MACHINE-LEARNING SYSTEMS AND TECHNIQUES TO OPTIMIZE TELEOPERATION AND/OR PLANNER DECISIONS
A system, an apparatus or a process may be configured to implement an application that applies artificial intelligence and/or machine-learning techniques to predict an optimal course of action (or a subset of courses of action) for an autonomous vehicle system (e.g., one or more of a planner of an autonomous vehicle, a simulator, or a teleoperator) to undertake based on suboptimal autonomous vehicle performance and/or changes in detected sensor data (e.g., new buildings, landmarks, potholes, etc.). The application may determine a subset of trajectories based on a number of decisions and interactions when resolving an anomaly due to an event or condition. The application may use aggregated sensor data from multiple autonomous vehicles to assist in identifying events or conditions that might affect travel (e.g., using semantic scene classification). An optimal subset of trajectories may be formed based on recommendations responsive to semantic changes (e.g., road construction).
This application is a continuation-in-part of the following U.S. non-provisional patent applications: U.S. patent application Ser. No. 14/932,959 filed Nov. 4, 2015 (Attorney Docket No. ZOO-001) entitled “Autonomous Vehicle Fleet Service And System,”; U.S. patent application Ser. No. 14/932,963 filed Nov. 4, 2015 (Attorney Docket No. ZOO-002) entitled “Adaptive Mapping To Navigate Autonomous Vehicles Responsive To Physical Environment Changes,”; U.S. patent application Ser. No. 14/932,966 filed Nov. 4, 2015 (Attorney Docket No. ZOO-004) entitled “Teleoperation System And Method For Trajectory Modification Of Autonomous Vehicles,”; and U.S. patent application Ser. No. 14/932,940 filed Nov. 4, 2015 (Attorney Docket No. ZOO-005) entitled “Automated Extraction Of Semantic Information To Enhance Incremental Mapping Modifications For Robotic Vehicles,” all of which are hereby incorporated by reference in their entirety for all purposes.FIELD
Various embodiments relate generally to autonomous vehicles and associated mechanical, electrical and electronic hardware, computer software and systems, and wired and wireless network communications to provide automated decision selection and execution in autonomous vehicles.BACKGROUND
Environments in which an autonomous vehicle may be tasked to navigate in may change over time due to changes in patterns of use (e.g., by pedestrians), road infrastructure (e.g., traffic signs, traffic lights, road markings, etc.) road conditions (e.g., road construction, a lane closure, potholes, an obstruction on a road surface, etc.). Changes in road conditions, such as lane closures, potholes and the like may require the autonomous vehicle to take appropriate actions to alter an initial guided path to a revised guided path. However, detecting events that may give rise to the need to implement a change may not be effective if the autonomous vehicle is not configured to predict a course of action to take based on a newly detected event. Moreover, information associated with an event and responses taken by the autonomous vehicle in response to the event may not benefit other autonomous vehicles that may also encounter a similar event unless the information and/or response is disseminated to the other autonomous vehicles.
Thus, what is needed is a solution to implement event detection that predicts optimal courses of action responsive to the event, without the limitations of conventional techniques.
Various embodiments or examples (“examples”) of the invention are disclosed in the following detailed description and the accompanying drawings:
Although the above-described drawings depict various examples of the invention, the invention is not limited by the depicted examples. It is to be understood that, in the drawings, like reference numerals designate like structural elements. Also, it is understood that the drawings are not necessarily to scale.DETAILED DESCRIPTION
Various embodiments or examples may be implemented in numerous ways, including as a system, a process, an apparatus, a user interface, or a series of program instructions on a computer readable medium such as a computer readable storage medium or a computer network where the program instructions are sent over optical, electronic, or wireless communication links. In general, operations of disclosed processes may be performed in an arbitrary order, unless otherwise provided in the claims.
A detailed description of one or more examples is provided below along with accompanying figures. The detailed description is provided in connection with such examples, but is not limited to any particular example. The scope is limited only by the claims, and numerous alternatives, modifications, and equivalents thereof. Numerous specific details are set forth in the following description in order to provide a thorough understanding. These details are provided for the purpose of example and the described techniques may be practiced according to the claims without some or all of these specific details. For clarity, technical material that is known in the technical fields related to the examples has not been described in detail to avoid unnecessarily obscuring the description.
According to some examples, at least some of autonomous vehicles 109a to 109e are configured as bidirectional autonomous vehicles, such as bidirectional autonomous vehicle (“AV”) 130. Bidirectional autonomous vehicle 130 may be configured to travel in either direction principally along, but not limited to, a longitudinal axis 131. Accordingly, bidirectional autonomous vehicle 130 may be configured to implement active lighting external to the vehicle to alert others (e.g., other drivers, pedestrians, cyclists, etc.) in the adjacent vicinity, and a direction in which bidirectional autonomous vehicle 130 is traveling. For example, active sources of light 136 may be implemented as active lights 138a when traveling in a first direction, or may be implemented as active lights 138b when traveling in a second direction. Active lights 138a may be implemented using a first subset of one or more colors, with optional animation (e.g., light patterns of variable intensities of light or color that may change over time). Similarly, active lights 138b may be implemented using a second subset of one or more colors and light patterns that may be different than those of active lights 138a. For example, active lights 138a may be implemented using white-colored lights as “headlights,” whereas active lights 138b may be implemented using red-colored lights as “taillights.” Active lights 138a and 138b, or portions thereof, may be configured to provide other light-related functionalities, such as provide “turn signal indication” functions (e.g., using yellow light). According to various examples, logic in autonomous vehicle 130 may be configured to adapt active lights 138a and 138b to comply with various safety requirements and traffic regulations or laws for any number of jurisdictions.
In some embodiments, bidirectional autonomous vehicle 130 may be configured to have similar structural elements and components in each quad portion, such as quad portion 194. The quad portions are depicted, at least in this example, as portions of bidirectional autonomous vehicle 130 defined by the intersection of a plane 132 and a plane 134, both of which pass through the vehicle to form two similar halves on each side of planes 132 and 134. Further, bidirectional autonomous vehicle 130 may include an autonomous vehicle controller 147 that includes logic (e.g., hardware or software, or as combination thereof) that is configured to control a predominate number of vehicle functions, including driving control (e.g., propulsion, steering, etc.) and active sources 136 of light, among other functions. Bidirectional autonomous vehicle 130 also includes a number of sensors 139 disposed at various locations on the vehicle (other sensors are not shown).
Autonomous vehicle controller 147 may be further configured to determine a local pose (e.g., local position) of an autonomous vehicle 109 and to detect external objects relative to the vehicle. For example, consider that bidirectional autonomous vehicle 130 is traveling in the direction 119 in road network 110. A localizer (not shown) of autonomous vehicle controller 147 can determine a local pose at the geographic location 111. As such, the localizer may use acquired sensor data, such as sensor data associated with surfaces of buildings 115 and 117, which can be compared against reference data, such as map data (e.g., 3D map data, including reflectance data) to determine a local pose. Further, a perception engine (not shown) of autonomous vehicle controller 147 may be configured to detect, classify, and predict the behavior of external objects, such as external object 112 (a “tree”) and external object 114 (a “pedestrian”). Classification of such external objects may broadly classify objects as static objects, such as external object 112, and dynamic objects, such as external object 114. The localizer and the perception engine, as well as other components of the AV controller 147, collaborate to cause autonomous vehicles 109 to drive autonomously.
According to some examples, autonomous vehicle service platform 101 is configured to provide teleoperator services should an autonomous vehicle 109 request teleoperation. For example, consider that an autonomous vehicle controller 147 in autonomous vehicle 109d detects an object 126 obscuring a path 124 on roadway 122 at point 191, as depicted in inset 120. If autonomous vehicle controller 147 cannot ascertain a path or trajectory over which vehicle 109d may safely transit with a relatively high degree of certainty, then autonomous vehicle controller 147 may transmit request message 105 for teleoperation services. In response, a teleoperator computing device 104 may receive instructions from a teleoperator 108 to perform a course of action to successfully (and safely) negotiate obstacles 126. Response data 107 then can be transmitted back to autonomous vehicle 109d to cause the vehicle to, for example, safely cross a set of double lines as it transits along the alternate path 121. In some examples, teleoperator computing device 104 may generate a response identifying geographic areas to exclude from planning a path. In particular, rather than provide a path to follow, a teleoperator 108 may define areas or locations that the autonomous vehicle must avoid.
In view of the foregoing, the structures and/or functionalities of autonomous vehicle 130 and/or autonomous vehicle controller 147, as well as their components, can perform real-time (or near real-time) trajectory calculations through autonomous-related operations, such as localization and perception, to enable autonomous vehicles 109 to self-drive.
In some cases, the bidirectional nature of bidirectional autonomous vehicle 130 provides for a vehicle that has quad portions 194 (or any other number of symmetric portions) that are similar or are substantially similar to each other. Such symmetry reduces complexity of design and decreases relatively the number of unique components or structures, thereby reducing inventory and manufacturing complexities. For example, a drivetrain and wheel system may be disposed in any of the quad portions 194. Further, autonomous vehicle controller 147 is configured to invoke teleoperation services to reduce the likelihood that an autonomous vehicle 109 is delayed in transit while resolving an event or issue that may otherwise affect the safety of the occupants. In some cases, the visible portion of road network 110 depicts a geo-fenced region that may limit or otherwise control the movement of autonomous vehicles 109 to the road network shown in
At 206, data representing a subset of candidate trajectories may be received from an autonomous vehicle responsive to the detection of the event. For example, a planner of an autonomous vehicle controller may calculate and evaluate large numbers of trajectories (e.g., thousands or greater) per unit time, such as a second. In some embodiments, candidate trajectories are a subset of the trajectories that provide for relatively higher confidence levels that an autonomous vehicle may move forward safely in view of the event (e.g., using an alternate path provided by a teleoperator). Note that some candidate trajectories may be ranked or associated with higher degrees of confidence than other candidate trajectories. According to some examples, subsets of candidate trajectories may originate from any number of sources, such as a planner, a teleoperator computing device (e.g., teleoperators can determine and provide approximate paths), etc., and may be combined as a superset of candidate trajectories. At 208, path guidance data may be identified at one or more processors. The path guidance data may be configured to assist a teleoperator in selecting a guided trajectory from one or more of the candidate trajectories. In some instances, the path guidance data specifies a value indicative of a confidence level or probability that indicates the degree of certainty that a particular candidate trajectory may reduce or negate the probability that the event may impact operation of an autonomous vehicle. A guided trajectory, as a selected candidate trajectory, may be received at 210, responsive to input from a teleoperator (e.g., a teleoperator may select at least one candidate trajectory as a guided trajectory from a group of differently-ranked candidate trajectories). The selection may be made via an operator interface that lists a number of candidate trajectories, for example, in order from highest confidence levels to lowest confidence levels. At 212, the selection of a candidate trajectory as a guided trajectory may be transmitted to the vehicle, which, in turn, implements the guided trajectory for resolving the condition by causing the vehicle to perform a teleoperator-specified maneuver. As such, the autonomous vehicle may transition from a non-normative operational state.
According to some embodiments, portions of the autonomous vehicle (“AV”) control logic 347 may be implemented using clusters of graphics processing units (“GPUs”) implementing a framework and programming model suitable for programming the clusters of GPUs. For example, a compute unified device architecture (“CUDA”)-compatible programming language and application programming interface (“API”) model may be used to program the GPUs. CUDA™ is produced and maintained by NVIDIA of Santa Clara, Calif. Note that other programming languages may be implemented, such as OpenCL, or any other parallel programming language.
According to some embodiments, autonomous vehicle control logic 347 may be implemented in hardware and/or software as autonomous vehicle controller 347a, which is shown to include a motion controller 362, a planner 364, a perception engine 366, and a localizer 368. As shown, autonomous vehicle controller 347a is configured to receive camera data 340a, Lidar data 346a, and radar data 348a, or any other range-sensing or localization data, including sonar data 341a or the like. Autonomous vehicle controller 347a is also configured to receive positioning data, such as GPS data 352, IMU data 354, and other position-sensing data (e.g., wheel-related data, such as steering angles, angular velocity, etc.). Further, autonomous vehicle controller 347a may receive any other sensor data 356, as well as reference data 339. In some cases, reference data 339 includes map data (e.g., 3D map data, 2D map data, 4D map data (e.g., including Epoch Determination)) and route data (e.g., road network data, including, but not limited to, RNDF data (or similar data), MDF data (or similar data), etc.
Localizer 368 is configured to receive sensor data from one or more sources, such as GPS data 352, wheel data, IMU data 354, Lidar data 346a, camera data 340a, radar data 348a, and the like, as well as reference data 339 (e.g., 3D map data and route data). Localizer 368 integrates (e.g., fuses the sensor data) and analyzes the data by comparing sensor data to map data to determine a local pose (or position) of bidirectional autonomous vehicle 330. According to some examples, localizer 368 may generate or update the pose or position of any autonomous vehicle in real-time or near real-time. Note that localizer 368 and its functionality need not be limited to “bi-directional” vehicles and can be implemented in any vehicle of any type. Therefore, localizer 368 (as well as other components of AV controller 347a) may be implemented in a “uni-directional” vehicle or any non-autonomous vehicle. According to some embodiments, data describing a local pose may include one or more of an x-coordinate, a y-coordinate, a z-coordinate (or any coordinate of any coordinate system, including polar or cylindrical coordinate systems, or the like), a yaw value, a roll value, a pitch value (e.g., an angle value), a rate (e.g., velocity), altitude, and the like.
Perception engine 366 is configured to receive sensor data from one or more sources, such as Lidar data 346a, camera data 340a, radar data 348a, and the like, as well as local pose data. Perception engine 366 may be configured to determine locations of external objects based on sensor data and other data. External objects, for instance, may be objects that are not part of a drivable surface. For example, perception engine 366 may be able to detect and classify external objects as pedestrians, bicyclists, dogs, other vehicles, etc. (e.g., perception engine 366 is configured to classify the objects in accordance with a type of classification, which may be associated with semantic information, including a label). Based on the classification of these external objects, the external objects may be labeled as dynamic objects or static objects. For example, an external object classified as a tree may be labeled as a static object, while an external object classified as a pedestrian may be labeled as a static object. External objects labeled as static may or may not be described in map data. Examples of external objects likely to be labeled as static include traffic cones, cement barriers arranged across a roadway, lane closure signs, newly-placed mailboxes or trash cans adjacent a roadway, etc. Examples of external objects likely to be labeled as dynamic include bicyclists, pedestrians, animals, other vehicles, etc. If the external object is labeled as dynamic, and further data about the external object may indicate a typical level of activity and velocity, as well as behavior patterns associated with the classification type. Further data about the external object may be generated by tracking the external object. As such, the classification type can be used to predict or otherwise determine the likelihood that an external object may, for example, interfere with an autonomous vehicle traveling along a planned path. For example, an external object that is classified as a pedestrian may be associated with some maximum speed, as well as an average speed (e.g., based on tracking data). The velocity of the pedestrian relative to the velocity of an autonomous vehicle can be used to determine if a collision is likely. Further, perception engine 364 may determine a level of uncertainty associated with a current and future state of objects. In some examples, the level of uncertainty may be expressed as an estimated value (or probability).
Planner 364 is configured to receive perception data from perception engine 366, and may also include localizer data from localizer 368. According to some examples, the perception data may include an obstacle map specifying static and dynamic objects located in the vicinity of an autonomous vehicle, whereas the localizer data may include a local pose or position. In operation, planner 364 generates numerous trajectories, and evaluates the trajectories, based on at least the location of the autonomous vehicle against relative locations of external dynamic and static objects. Planner 364 selects an optimal trajectory based on a variety of criteria over which to direct the autonomous vehicle in way that provides for collision-free travel. In some examples, planner 364 may be configured to calculate the trajectories as probabilistically-determined trajectories. Further, planner 364 may transmit steering and propulsion commands (as well as decelerating or braking commands) to motion controller 362. Motion controller 362 subsequently may convert any of the commands, such as a steering command, a throttle or propulsion command, and a braking command, into control signals (e.g., for application to actuators or other mechanical interfaces) to implement changes in steering or wheel angles 351 and/or velocity 353.
Localizer 468 is configured to localize autonomous vehicle (i.e., determine a local pose) relative to reference data, which may include map data, route data (e.g., road network data, such as RNDF-like data), and the like. In some cases, localizer 468 is configured to identify, for example, a point in space that may represent a location of autonomous vehicle 430 relative to features of a representation of an environment. Localizer 468 is shown to include a sensor data integrator 469, which may be configured to integrate multiple subsets of sensor data (e.g., of different sensor modalities) to reduce uncertainties related to each individual type of sensor. According to some examples, sensor data integrator 469 is configured to fuse sensor data (e.g., Lidar data, camera data, radar data, etc.) to form integrated sensor data values for determining a local pose. According to some examples, localizer 468 retrieves reference data originating from a reference data repository 405, which includes a map data repository 405a for storing 2D map data, 3D map data, 4D map data, and the like. Localizer 468 may be configured to identify at least a subset of features in the environment to match against map data to identify, or otherwise confirm, a pose of autonomous vehicle 430. According to some examples, localizer 468 may be configured to identify any amount of features in an environment, such that a set of features can one or more features, or all features. In a specific example, any amount of Lidar data (e.g., most or substantially all Lidar data) may be compared against data representing a map for purposes of localization. Generally, non-matched objects resulting from the comparison of the environment features and map data may be a dynamic object, such as a vehicle, bicyclist, pedestrian, etc. Note that detection of dynamic objects, including obstacles, may be performed with or without map data. In particular, dynamic objects may be detected and tracked independently of map data (i.e., in the absence of map data). In some instances, 2D map data and 3D map data may be viewed as “global map data” or map data that has been validated at a point in time by autonomous vehicle service platform 401. As map data in map data repository 405a may be updated and/or validated periodically, a deviation may exist between the map data and an actual environment in which the autonomous vehicle is positioned. Therefore, localizer 468 may retrieve locally-derived map data generated by local map generator 440 to enhance localization. Local map generator 440 is configured to generate local map data in real-time or near real-time. Optionally, local map generator 440 may receive static and dynamic object map data to enhance the accuracy of locally generated maps by, for example, disregarding dynamic objects in localization. According to at least some embodiments, local map generator 440 may be integrated with, or formed as part of, localizer 468. In at least one case, local map generator 440, either individually or in collaboration with localizer 468, may be configured to generate map and/or reference data based on simultaneous localization and mapping (“SLAM”) or the like. Note that localizer 468 may implement a “hybrid” approach to using map data, whereby logic in localizer 468 may be configured to select various amounts of map data from either map data repository 405a or local map data from local map generator 440, depending on the degrees of reliability of each source of map data. Therefore, localizer 468 may still use out-of-date map data in view of locally-generated map data.
Perception engine 466 is configured to, for example, assist planner 464 in planning routes and generating trajectories by identifying objects of interest in a surrounding environment in which autonomous vehicle 430 is transiting. Further, probabilities may be associated with each of the object of interest, whereby a probability may represent a likelihood that an object of interest may be a threat to safe travel (e.g., a fast-moving motorcycle may require enhanced tracking rather than a person sitting at a bus stop bench while reading a newspaper). As shown, perception engine 466 includes an object detector 442 and an object classifier 444. Object detector 442 is configured to distinguish objects relative to other features in the environment, and object classifier 444 may be configured to classify objects as either dynamic or static objects and track the locations of the dynamic and the static objects relative to autonomous vehicle 430 for planning purposes. Further, perception engine 466 may be configured to assign an identifier to a static or dynamic object that specifies whether the object is (or has the potential to become) an obstacle that may impact path planning at planner 464. Although not shown in
Planner 464 is configured to generate a number of candidate trajectories for accomplishing a goal to reaching a destination via a number of paths or routes that are available. Trajectory evaluator 465 is configured to evaluate candidate trajectories and identify which subsets of candidate trajectories are associated with higher degrees of confidence levels of providing collision-free paths to the destination. As such, trajectory evaluator 465 can select an optimal trajectory based on relevant criteria for causing commands to generate control signals for vehicle components 450 (e.g., actuators or other mechanisms). Note that the relevant criteria may include any number of factors that define optimal trajectories, the selection of which need not be limited to reducing collisions. For example, the selection of trajectories may be made to optimize user experience (e.g., user comfort) as well as collision-free trajectories that comply with traffic regulations and laws. User experience may be optimized by moderating accelerations in various linear and angular directions (e.g., to reduce jerking-like travel or other unpleasant motion). In some cases, at least a portion of the relevant criteria can specify which of the other criteria to override or supersede, while maintain optimized, collision-free travel. For example, legal restrictions may be temporarily lifted or deemphasized when generating trajectories in limited situations (e.g., crossing double yellow lines to go around a cyclist or travelling at higher speeds than the posted speed limit to match traffic flows). As such, the control signals are configured to cause propulsion and directional changes at the drivetrain and/or wheels. In this example, motion controller 462 is configured to transform commands into control signals (e.g., velocity, wheel angles, etc.) for controlling the mobility of autonomous vehicle 430. In the event that trajectory evaluator 465 has insufficient information to ensure a confidence level high enough to provide collision-free, optimized travel, planner 464 can generate a request to teleoperator 404 for teleoperator support.
Autonomous vehicle service platform 401 includes teleoperator 404 (e.g., a teleoperator computing device), reference data repository 405, a map updater 406, a vehicle data controller 408, a calibrator 409, and an off-line object classifier 410. Note that each element of autonomous vehicle service platform 401 may be independently located or distributed and in communication with other elements in autonomous vehicle service platform 401. Further, element of autonomous vehicle service platform 401 may independently communicate with the autonomous vehicle 430 via the communication layer 402. Map updater 406 is configured to receive map data (e.g., from local map generator 440, sensors 460, or any other component of autonomous vehicle controller 447), and is further configured to detect deviations, for example, of map data in map data repository 405a from a locally-generated map. Vehicle data controller 408 can cause map updater 406 to update reference data within repository 405 and facilitate updates to 2D, 3D, and/or 4D map data. In some cases, vehicle data controller 408 can control the rate at which local map data is received into autonomous vehicle service platform 408 as well as the frequency at which map updater 406 performs updating of the map data.
Calibrator 409 is configured to perform calibration of various sensors of the same or different types. Calibrator 409 may be configured to determine the relative poses of the sensors (e.g., in Cartesian space (x, y, z)) and orientations of the sensors (e.g., roll, pitch and yaw). The pose and orientation of a sensor, such a camera, Lidar sensor, radar sensor, etc., may be calibrated relative to other sensors, as well as globally relative to the vehicle's reference frame. Off-line self-calibration can also calibrate or estimate other parameters, such as vehicle inertial tensor, wheel base, wheel radius or surface road friction. Calibration can also be done online to detect parameter change, according to some examples. Note, too, that calibration by calibrator 409 may include intrinsic parameters of the sensors (e.g., optical distortion, beam angles, etc.) and extrinsic parameters. In some cases, calibrator 409 may be performed by maximizing a correlation between depth discontinuities in 3D laser data and edges of image data, as an example. Off-line object classification 410 is configured to receive data, such as sensor data, from sensors 470 or any other component of autonomous vehicle controller 447. According to some embodiments, an off-line classification pipeline of off-line object classification 410 may be configured to pre-collect and annotate objects (e.g., manually by a human and/or automatically using an offline labeling algorithm), and may further be configured to train an online classifier (e.g., object classifier 444), which can provide real-time classification of object types during online autonomous operation.
Teleoperator manager 707 is configured to manage a number of teleoperator computing devices 704 with which teleoperators 708 provide input. Simulator 740 is configured to simulate operation of one or more autonomous vehicles 730, as well as the interactions between teleoperator manager 707 and an autonomous vehicle 730. Simulator 740 may also simulate operation of a number of sensors (including the introduction of simulated noise) disposed in autonomous vehicle 730. Further, an environment, such as a city, may be simulated such that a simulated autonomous vehicle can be introduced to the synthetic environment, whereby simulated sensors may receive simulated sensor data, such as simulated laser returns. Simulator 740 may provide other functions as well, including validating software updates and/or map data. Policy manager 742 is configured to maintain data representing policies or rules by which an autonomous vehicle ought to behave in view of a variety of conditions or events that an autonomous vehicle encounters while traveling in a network of roadways. In some cases, updated policies and/or rules may be simulated in simulator 740 to confirm safe operation of a fleet of autonomous vehicles in view of changes to a policy. Some of the above-described elements of autonomous vehicle service platform 701 are further described hereinafter.
Communication channels 770 are configured to provide networked communication links among a fleet of autonomous vehicles 730 and autonomous vehicle service platform 701. For example, communication channel 770 includes a number of different types of networks 771, 772, 773, and 774, with corresponding subnetworks (e.g., 771a to 771n), to ensure a certain level of redundancy for operating an autonomous vehicle service reliably. For example, the different types of networks in communication channels 770 may include different cellular network providers, different types of data networks, etc., to ensure sufficient bandwidth in the event of reduced or lost communications due to outages in one or more networks 771, 772, 773, and 774.
An example of a data exchange for facilitating teleoperations via the communications protocol is described as follows. Consider that obstacle data 920 is generated by a perception system of an autonomous vehicle controller. Further, planner options data 924 is generated by a planner to notify a teleoperator of a subset of candidate trajectories, and position data 926 is generated by the localizer. Obstacle data 920, planner options data 924, and position data 926 are transmitted to a messaging service bridge 932, which, in accordance with message service configuration data 934, generates telemetry data 940 and query data 942, both of which are transmitted via data-centric messaging bus 972 into teleoperator application 901 as telemetry data 950 and query data 952. Teleoperator API 962 receives telemetry data 950 and inquiry data 952, which, in turn are processed in view of Route data 960 and message service configuration data 964. The resultant data is subsequently presented to a teleoperator 908 via teleoperator computing device 904 and/or a collaborative display (e.g., a dashboard display visible to a group of collaborating teleoperators 908). Teleoperator 908 reviews the candidate trajectory options that are presented on the display of teleoperator computing device 904, and selects a guided trajectory, which generates command data 982 and query response data 980, both of which are passed through teleoperator API 962 as query response data 954 and command data 956. In turn, query response data 954 and command data 956 are transmitted via data-centric messaging bus 972 into autonomous vehicle application 930 as query response data 944 and command data 946. Messaging service bridge 932 receives query response data 944 and command data 946 and generates teleoperator command data 928, which is configured to generate a teleoperator-selected trajectory for implementation by a planner. Note that the above-described messaging processes are not intended to be limiting, and other messaging protocols may be implemented as well.
Trajectory evaluator 1120 includes a state and event manager 1122, which, in turn, may include a confidence level generator 1123. Trajectory evaluator 1120 further includes a guided trajectory generator 1126 and a trajectory generator 1124. Further, planner 1164 is configured to receive policy data 1130, perception engine data 1132, and localizer data 1134.
Policy data 1130 may include criteria with which planner 1164 uses to determine a path that has a sufficient confidence level with which to generate trajectories, according to some examples. Examples of policy data 1130 include policies that specify that trajectory generation is bounded by stand-off distances to external objects (e.g., maintaining a safety buffer of 3 feet from a cyclist, as possible), or policies that require that trajectories must not cross a center double yellow line, or policies that require trajectories to be limited to a single lane in a 4-lane roadway (e.g., based on past events, such as typically congregating at a lane closest to a bus stop), and any other similar criteria specified by policies. Perception engine data 1132 includes maps of locations of static objects and dynamic objects of interest, and localizer data 1134 includes at least a local pose or position.
State and event manager 1122 may be configured to probabilistically determine a state of operation for an autonomous vehicle. For example, a first state of operation (i.e., “normative operation”) may describe a situation in which trajectories are collision-free, whereas a second state of operation (i.e., “non-normative operation”) may describe another situation in which the confidence level associated with possible trajectories are insufficient to guarantee collision-free travel. According to some examples, state and event manager 1122 is configured to use perception data 1132 to determine a state of autonomous vehicle that is either normative or non-normative. Confidence level generator 1123 may be configured to analyze perception data 1132 to determine a state for the autonomous vehicle. For example, confidence level generator 1123 may use semantic information associated with static and dynamic objects, as well as associated probabilistic estimations, to enhance a degree of certainty that planner 1164 is determining safe course of action. For example, planner 1164 may use perception engine data 1132 that specifies a probability that an object is either a person or not a person to determine whether planner 1164 is operating safely (e.g., planner 1164 may receive a degree of certainty that an object has a 98% probability of being a person, and a probability of 2% that the object is not a person).
Upon determining a confidence level (e.g., based on statistics and probabilistic determinations) is below a threshold required for predicted safe operation, a relatively low confidence level (e.g., single probability score) may trigger planner 1164 to transmit a request 1135 for teleoperation support to autonomous vehicle service platform 1101. In some cases, telemetry data and a set of candidate trajectories may accompany the request. Examples of telemetry data include sensor data, localization data, perception data, and the like. A teleoperator 1108 may transmit via teleoperator computing device 1104 a selected trajectory 1137 to guided trajectory generator 1126. As such, selected trajectory 1137 is a trajectory formed with guidance from a teleoperator. Upon confirming there is no change in the state (e.g., a non-normative state is pending), guided trajectory generator 1126 passes data to trajectory generator 1124, which, in turn, causes trajectory tracker 1128, as a trajectory tracking controller, to use the teleop-specified trajectory for generating control signals 1170 (e.g., steering angles, velocity, etc.). Note that planner 1164 may trigger transmission of a request 1135 for teleoperation support prior to a state transitioning to a non-normative state. In particular, an autonomous vehicle controller and/or its components can predict that a distant obstacle may be problematic and preemptively cause planner 1164 to invoke teleoperations prior to the autonomous vehicle reaching the obstacle. Otherwise, the autonomous vehicle may cause a delay by transitioning to a safe state upon encountering the obstacle or scenario (e.g., pulling over and off the roadway). In another example, teleoperations may be automatically invoked prior to an autonomous vehicle approaching a particular location that is known to be difficult to navigate. This determination may optionally take into consideration other factors, including the time of day, the position of the sun, if such situation is likely to cause a disturbance to the reliability of sensor readings, and traffic or accident data derived from a variety of sources.
In another state of operation (e.g., a normative state), static map data 1301, current and predicted object state data 1303, local pose data 1305, and plan data 1307 (e.g., global plan data) are received into trajectory calculator 1325, which is configured to calculate (e.g., iteratively) trajectories to determine an optimal one or more paths. Next, at least one path is selected and is transmitted as selected path data 1311. According to some embodiments, trajectory calculator 1325 is configured to implement re-planning of trajectories as an example. Nominal driving trajectory generator 1327 is configured to generate trajectories in a refined approach, such as by generating trajectories based on receding horizon control techniques. Nominal driving trajectory generator 1327 subsequently may transmit nominal driving trajectory path data 1372 to, for example, a trajectory tracker or a vehicle controller to implement physical changes in steering, acceleration, and other components.
Teleoperator action recommendation controller 1412 includes logic configured to receive and/or control a teleoperation service request via autonomous vehicle (“AV”) planner data 1472, which can include requests for teleoperator assistance as well as telemetry data and other data. As such, planner data 1472 may include recommended candidate trajectories or paths from which a teleoperator 1408 via teleoperator computing device 1404 may select. According to some examples, teleoperator action recommendation controller 1412 may be configured to access other sources of recommended candidate trajectories from which to select an optimum trajectory. For example, candidate trajectories contained in autonomous vehicle planner data 1472 may, in parallel, be introduced into simulator 1440, which is configured to simulate an event or condition being experienced by an autonomous vehicle requesting teleoperator assistance. Simulator 1440 can access map data and other data necessary for performing a simulation on the set of candidate trajectories, whereby simulator 1440 need not exhaustively reiterate simulations to confirm sufficiency. Rather, simulator 1440 may provide either confirm the appropriateness of the candidate trajectories, or may otherwise alert a teleoperator to be cautious in their selection.
Teleoperator interaction capture analyzer 1416 may be configured to capture numerous amounts of teleoperator transactions or interactions for storage in repository 1441, which, for example, may accumulate data relating to a number of teleoperator transactions for analysis and generation of policies, at least in some cases. According to some embodiments, repository 1441 may also be configured to store policy data for access by policy manager 1442. Further, teleoperator interaction capture analyzer 1416 may apply machine learning techniques to empirically determine how best to respond to events or conditions causing requests for teleoperation assistance. In some cases, policy manager 1442 may be configured to update a particular policy or generate a new policy responsive to analyzing the large set of teleoperator interactions (e.g., subsequent to applying machine learning techniques). Policy manager 1442 manages policies that may be viewed as rules or guidelines with which an autonomous vehicle controller and its components operate under to comply with autonomous operations of a vehicle. In some cases, a modified or updated policy may be applied to simulator 1440 to confirm the efficacy of permanently releasing or implementing such policy changes.
Simulator interface controller 1414 is configured to provide an interface between simulator 1440 and teleoperator computing devices 1404. For example, consider that sensor data from a fleet of autonomous vehicles is applied to reference data updater 1438 via autonomous (“AV”) fleet data 1470, whereby reference data updater 1438 is configured to generate updated map and route data 1439. In some implementations, updated map and route data 1439 may be preliminarily released as an update to data in map data repositories 1420 and 1422, or as an update to data in route data repository 1424. In this case, such data may be tagged as being a “beta version” in which a lower threshold for requesting teleoperator service may be implemented when, for example, a map tile including preliminarily updated information is used by an autonomous vehicle. Further, updated map and route data 1439 may be introduced to simulator 1440 for validating the updated map data. Upon full release (e.g., at the close of beta testing), the previously lowered threshold for requesting a teleoperator service related to map tiles is canceled. User interface graphics controller 1410 provides rich graphics to teleoperators 1408, whereby a fleet of autonomous vehicles may be simulated within simulator 1440 and may be accessed via teleoperator computing device 1404 as if the simulated fleet of autonomous vehicles were real.
Fleet optimization manager 1620 is shown to include a hybrid autonomous vehicle/non-autonomous vehicle processor 1640, which, in turn, includes an AV/non-AV optimization calculator 1642 and a non-AV selector 1644. According to some examples, hybrid autonomous vehicle/non-autonomous vehicle processor 1640 is configured to manage a hybrid fleet of autonomous vehicles and human-driven vehicles (e.g., as independent contractors). As such, autonomous vehicle service may employ non-autonomous vehicles to meet excess demand, or in areas, such as non-AV service region 1690, that may be beyond a geo-fence or in areas of poor communication coverage. AV/non-AV optimization calculator 1642 is configured to optimize usage of the fleet of autonomous and to invite non-AV drivers into the transportation service (e.g., with minimal or no detriment to the autonomous vehicle service). Non-AV selector 1644 includes logic for selecting a number of non-AV drivers to assist based on calculations derived by AV/non-AV optimization calculator 1642.
Communication event detector 1840 includes a policy download manager 1842 and communications-configured (“COMM-configured”) AV dispatcher 1844. Policy download manager 1842 is configured to provide autonomous vehicles 1830 an updated policy in view of reduced communications region 1880, whereby the updated policy may specify routes to quickly exit region 1880 if an autonomous vehicle enters that region. For example, autonomous vehicle 1864 may receive an updated policy moments before driving into region 1880. Upon loss of communications, autonomous vehicle 1864 implements the updated policy and selects route 1866 to drive out of region 1880 quickly. COMM-configured AV dispatcher 1844 may be configured to identify points 1865 at which to park autonomous vehicles that are configured as relays to establishing a peer-to-peer network over region 1880. As such, COMM-configured AV dispatcher 1844 is configured to dispatch autonomous vehicles 1862 (without passengers) to park at locations 1865 for the purposes of operating as communication towers in a peer-to-peer ad hoc network.
Further to diagram 2200, localization-based data 2250 and relative localization-based data 2251 may be fed into data integrator 2266a and localization data integrator 2266, respectively. Data integrator 2266a and localization data integrator 2266 may be configured to fuse corresponding data, whereby localization-based data 2250 may be fused at data integrator 2266a prior to being fused with relative localization-based data 2251 at localization data integrator 2266. According to some embodiments, data integrator 2266a is formed as part of localization data integrator 2266, or is absent. Regardless, a localization-based data 2250 and relative localization-based data 2251 can be both fed into localization data integrator 2266 for purposes of fusing data to generate local position data 2252. Localization-based data 2250 may include unary-constrained data (and uncertainty values) from projection processor 2254a, as well as binary-constrained data (and uncertainty values) from odometry processor 2254b and integrator processor 2254c. Relative localization-based data 2251 may include unary-constrained data (and uncertainty values) from localization processor 2254d and visual registration processor 2254e, and optionally from radar return processor 2254f. According to some embodiments, localization data integrator 2266 may implement non-linear smoothing functionality, such as a Kalman filter (e.g., a gated Kalman filter), a relative bundle adjuster, pose-graph relaxation, particle filter, histogram filter, or the like.
Referring back to
Simulator 2840 may be configured to generate a simulated autonomous vehicle controller 2847, which includes synthetic adaptations of a perception engine 2866, a localizer 2868, a motion controller 2862, and a planner 2864, each of which may have functionalities described herein within simulated environment 2803. Simulator 2840 may also generate simulated interfaces (“I/F”) 2849 to simulate the data exchanges with different sensors modalities and different sensor data formats. As such, simulated interface 2849 may simulate a software interface for packetized data from, for example, a simulated Lidar sensor 2872. Further, simulator 2840 may also be configured to generate a simulated autonomous vehicle 2830 that implements simulated AV controller 2847. Simulated autonomous vehicle 2830 includes simulated Lidar sensors 2872, simulated camera or image sensors 2874, and simulated radar sensors 2876. In the example shown, simulated Lidar sensor 2872 may be configured to generate a simulated laser consistent with ray trace 2892, which causes generation of simulated sensor return 2891. Note that simulator 2840 may simulate the addition of noise or other environmental effects on sensor data (e.g., added diffusion or reflections that affect simulated sensor return 2891, etc.). Further yet, simulator 2840 may be configured to simulate a variety of sensor defects, including sensor failure, sensor miscalibration, intermittent data outages, and the like.
Simulator 2840 includes a physics processor 2850 for simulating the mechanical, static, dynamic, and kinematic aspects of an autonomous vehicle for use in simulating behavior of simulated autonomous vehicle 2830. For example, physics processor 2850 includes a content mechanics module 2851 for simulating contact mechanics, a collision detection module 2852 for simulating the interaction between simulated bodies, and a multibody dynamics module 2854 to simulate the interaction between simulated mechanical interactions.
Simulator 2840 also includes a simulator controller 2856 configured to control the simulation to adapt the functionalities of any synthetically-generated element of simulated environment 2803 to determine cause-effect relationship, among other things. Simulator 2840 includes a simulator evaluator 2858 to evaluate the performance synthetically-generated element of simulated environment 2803. For example, simulator evaluator 2858 may analyze simulated vehicle commands 2880 (e.g., simulated steering angles and simulated velocities) to determine whether such commands are an appropriate response to the simulated activities within simulated environment 2803. Further, simulator evaluator 2858 may evaluate interactions of a teleoperator 2808 with the simulated autonomous vehicle 2830 via teleoperator computing device 2804. Simulator evaluator 2858 may evaluate the effects of updated reference data 2827, including updated map tiles and route data, which may be added to guide the responses of simulated autonomous vehicle 2830. Simulator evaluator 2858 may also evaluate the responses of simulator AV controller 2847 when policy data 2829 is updated, deleted, or added. The above-description of simulator 2840 is not intended to be limiting. As such, simulator 2840 is configured to perform a variety of different simulations of an autonomous vehicle relative to a simulated environment, which include both static and dynamic features. For example, simulator 2840 may be used to validate changes in software versions to ensure reliability. Simulator 2840 may also be used to determine vehicle dynamics properties and for calibration purposes. Further, simulator 2840 may be used to explore the space of applicable controls and resulting trajectories so as to effect learning by self-simulation.
Further to the example shown, autonomous vehicle application 3240 may also include a user identification controller 3246 that may be configured to detect that user 3202 is in a geographic region, or vicinity, near autonomous vehicle 3230, as the vehicle approaches. In some situations, user 3202 may not readily perceive or identify autonomous vehicle 3230 as it approaches for use by user 3203 (e.g., due to various other vehicles, including trucks, cars, taxis, and other obstructions that are typical in city environments). In one example, autonomous vehicle 3230 may establish a wireless communication link 3262 (e.g., via a radio frequency (“RF”) signal, such as WiFi or Bluetooth®, including BLE, or the like) for communicating and/or determining a spatial location of user 3202 relative to autonomous vehicle 3230 (e.g., using relative direction of RF signal and signal strength). In some cases, autonomous vehicle 3230 may detect an approximate geographic location of user 3202 using, for example, GPS data or the like. A GPS receiver (not shown) of mobile computing device 3203 may be configured to provide GPS data to autonomous vehicle service application 3240. Thus, user identification controller 3246 may provide GPS data via link 3260 to autonomous vehicle service platform 3201, which, in turn, may provide that location to autonomous vehicle 3230 via link 3261. Subsequently, autonomous vehicle 3230 may determine a relative distance and/or direction of user 3202 by comparing the user's GPS data to the vehicle's GPS-derived location.
Autonomous vehicle 3230 may also include additional logic to identify the presence of user 3202, such that logic configured to perform face detection algorithms to detect either user 3202 generally, or to specifically identify the identity (e.g., name, phone number, etc.) of user 3202 based on the user's unique facial characteristics. Further, autonomous vehicle 3230 may include logic to detect codes for identifying user 3202. Examples of such codes include specialized visual codes, such as QR codes, color codes, etc., specialized audio codes, such as voice activated or recognized codes, etc., and the like. In some cases, a code may be an encoded security key that may be transmitted digitally via link 3262 to autonomous vehicle 3230 to ensure secure ingress and/or egress. Further, one or more of the above-identified techniques for identifying user 3202 may be used as a secured means to grant ingress and egress privileges to user 3202 so as to prevent others from entering autonomous vehicle 3230 (e.g., to ensure third party persons do not enter an unoccupied autonomous vehicle prior to arriving at user 3202). According to various examples, any other means for identifying user 3202 and providing secured ingress and egress may also be implemented in one or more of autonomous vehicle service application 3240, autonomous vehicle service platform 3201, and autonomous vehicle 3230.
To assist user 3302 in identifying the arrival of its requested transportation, autonomous vehicle 3230 may be configured to notify or otherwise alert user 3202 to the presence of autonomous vehicle 3230 as it approaches user 3202. For example, autonomous vehicle 3230 may activate one or more light-emitting devices 3280 (e.g., LEDs) in accordance with specific light patterns. In particular, certain light patterns are created so that user 3202 may readily perceive that autonomous vehicle 3230 is reserved to service the transportation needs of user 3202. As an example, autonomous vehicle 3230 may generate light patterns 3290 that may be perceived by user 3202 as a “wink,” or other animation of its exterior and interior lights in such a visual and temporal way. The patterns of light 3290 may be generated with or without patterns of sound to identify to user 3202 that this vehicle is the one that they booked.
According to some embodiments, autonomous vehicle user controller 3244 may implement a software application that is configured to control various functions of an autonomous vehicle. Further, an application may be configured to redirect or reroute the autonomous vehicle during transit to its initial destination. Further, autonomous vehicle user controller 3244 may be configured to cause on-board logic to modify interior lighting of autonomous vehicle 3230 to effect, for example, mood lighting. Controller 3244 may also control a source of audio (e.g., an external source such as Spotify, or audio stored locally on the mobile computing device 3203), select a type of ride (e.g., modify desired acceleration and braking aggressiveness, modify active suspension parameters to select a set of “road-handling” characteristics to implement aggressive driving characteristics, including vibrations, or to select “soft-ride” qualities with vibrations dampened for comfort), and the like. For example, mobile computing device 3203 may be configured to control HVAC functions as well, like ventilation and temperature.
Note that various structures and/or functionalities of
In some cases, computing platform 3300 can be disposed in any device, such as a computing device 3390a, which may be disposed in an autonomous vehicle 3391, and/or mobile computing device 3390b.
Computing platform 3300 includes a bus 3302 or other communication mechanism for communicating information, which interconnects subsystems and devices, such as processor 3304, system memory 3306 (e.g., RAM, etc.), storage device 3308 (e.g., ROM, etc.), an in-memory cache (which may be implemented in RAM 3306 or other portions of computing platform 3300), a communication interface 3313 (e.g., an Ethernet or wireless controller, a Bluetooth controller, NFC logic, etc.) to facilitate communications via a port on communication link 3321 to communicate, for example, with a computing device, including mobile computing and/or communication devices with processors. Processor 3304 can be implemented with one or more graphics processing units (“GPUs”), with one or more central processing units (“CPUs”), such as those manufactured by Intel® Corporation, or one or more virtual processors, as well as any combination of CPUs and virtual processors. Computing platform 3300 exchanges data representing inputs and outputs via input-and-output devices 3301, including, but not limited to, keyboards, mice, audio inputs (e.g., speech-to-text devices), user interfaces, displays, monitors, cursors, touch-sensitive displays, LCD or LED displays, and other I/O-related devices.
According to some examples, computing platform 3300 performs specific operations by processor 3304 executing one or more sequences of one or more instructions stored in system memory 3306, and computing platform 3300 can be implemented in a client-server arrangement, peer-to-peer arrangement, or as any mobile computing device, including smart phones and the like. Such instructions or data may be read into system memory 3306 from another computer readable medium, such as storage device 3308. In some examples, hard-wired circuitry may be used in place of or in combination with software instructions for implementation. Instructions may be embedded in software or firmware. The term “computer readable medium” refers to any tangible medium that participates in providing instructions to processor 3304 for execution. Such a medium may take many forms, including but not limited to, non-volatile media and volatile media. Non-volatile media includes, for example, optical or magnetic disks and the like. Volatile media includes dynamic memory, such as system memory 3306.
Common forms of computer readable media includes, for example, floppy disk, flexible disk, hard disk, magnetic tape, any other magnetic medium, CD-ROM, any other optical medium, punch cards, paper tape, any other physical medium with patterns of holes, RAM, PROM, EPROM, FLASH-EPROM, any other memory chip or cartridge, or any other medium from which a computer can read. Instructions may further be transmitted or received using a transmission medium. The term “transmission medium” may include any tangible or intangible medium that is capable of storing, encoding or carrying instructions for execution by the machine, and includes digital or analog communications signals or other intangible medium to facilitate communication of such instructions. Transmission media includes coaxial cables, copper wire, and fiber optics, including wires that comprise bus 3302 for transmitting a computer data signal.
In some examples, execution of the sequences of instructions may be performed by computing platform 3300. According to some examples, computing platform 3300 can be coupled by communication link 3321 (e.g., a wired network, such as LAN, PSTN, or any wireless network, including WiFi of various standards and protocols, Bluetooth®, NFC, Zig-Bee, etc.) to any other processor to perform the sequence of instructions in coordination with (or asynchronous to) one another. Computing platform 3300 may transmit and receive messages, data, and instructions, including program code (e.g., application code) through communication link 3321 and communication interface 3313. Received program code may be executed by processor 3304 as it is received, and/or stored in memory 3306 or other non-volatile storage for later execution.
In the example shown, system memory 3306 can include various modules that include executable instructions to implement functionalities described herein. System memory 3306 may include an operating system (“O/S”) 3332, as well as an application 3336 and/or logic module(s) 3359. In the example shown in
Referring to the example shown in
Referring to the example shown in
Referring back to
In some embodiments, module 3350 of
In some cases, a mobile device, or any networked computing device (not shown) in communication with one or more modules 3359 (module 3350 of
For example, module 3350 of
As hardware and/or firmware, the above-described structures and techniques can be implemented using various types of programming or integrated circuit design languages, including hardware description languages, such as any register transfer language (“RTL”) configured to design field-programmable gate arrays (“FPGAs”), application-specific integrated circuits (“ASICs”), multi-chip modules, or any other type of integrated circuit.
For example, module 3350 of
Road network 3650 may represent a geo-fenced region or a portion of a geo-fenced region that may limit or otherwise control the movement of the autonomous vehicles 3630 and 3630a, for example. Autonomous vehicles 3630 and 3630a may be a subset of autonomous vehicles in a fleet of autonomous vehicles, for example. Other autonomous vehicles in the fleet of autonomous vehicles may be actively navigating other road networks in one or more other geo-fenced regions as denoted by 3698. Each autonomous vehicle 3630 and 3630a may implement sensing of an environment external to the autonomous vehicle using a sensor system included in each autonomous vehicle. For example, an autonomous vehicle denoted as 3630a may include sensor suites 3610a-3610d to sense the environment external to the autonomous vehicle 3630a as the vehicle autonomously navigates road surfaces within the road network 3650. Sensor suites 3610a-3610d may be configured to provide 360 degrees of overlapping sensor coverage regions (denoted as dashed circle 3610a) around the autonomous vehicle 3630a, for example. The other autonomous vehicles 3630 may also include sensor suites that provide 360 degrees of overlapping sensor coverage regions (denoted as dashed circle 3610) around the autonomous vehicles 3630, for example.
Further to diagram 3600, the autonomous vehicle 3630a may be autonomously navigating a road surface 3623 in a region 3624 of the road network 3650 in a direction of travel along a lane 3622 of the road surface 3623 generally indicated by arrow 3632, for example. A sensor system (not shown) of vehicle 3630a may detect objects on or proximate the road surface 3623 and sensor data indicative of those objects may be processed and classified by one or more systems (e.g., a perception system and/or a planner system) of the autonomous vehicle 3630a, for example. In the region 3624, the environment external to the autonomous vehicle 3630a may include but is not limited to the following detected objects (e.g., objects classified based on processed sensor data by a perception system): a building 3602; the road surface 3623, a lane marker 3625; lanes 3622 and 3624, curbs 3626 and 3627, pedestrians 3620, a fire hydrant 3631, and potholes 3634.
A planner 3670 of the autonomous vehicle 3630a may be configured to implement autonomous control of the autonomous vehicle 3630a in the road network 3650 based on data including but not limited to telemetry data 3671 and policy data 3672. Telemetry data 3671 may include but is not limited to sensor data 3673 generated by sensor types and modalities of sensors in the sensor system of the autonomous vehicle 3630a (e.g., one or more sensors depicted in
Policy data 3672 may include data 3674 being configured to implement a criteria with which planner 3670 uses to determine a path (e.g., a selected candidate trajectory) that has a sufficient confidence level with which to generate trajectories, according to some examples. Examples of data 3674 includes policies that specify that trajectory generation is bounded by stand-off distances to external objects (e.g., maintaining a safety buffer of 3 feet from a cyclist, as possible), or policies that require that trajectories must not cross a center double yellow line, etc. Data 3674 may include but is not limited to data representing confidence levels associated with candidate trajectories generated by the planner 3670 in response to an event (e.g., the potholes 3634) and an operational state of the autonomous vehicle (e.g., normative or non-normative), etc. In some examples, the data 3674 may be generated by a teleoperator. In other examples, the data 3674 may be generated by a simulator. In yet other examples, the data 3674 may be generated by the planner of autonomous vehicle 3630a or by another autonomous vehicle (e.g., by a planner in one of the vehicle's 3630).
Planner 3670 may communicate some or all of the data associated with the telemetry data 3671 and/or policy data 3672 as data 3696. Data 3639 may be communicated to a policy explorer 3699 being configured to receive the data 3639 and process the data 3639 using one or more computer resources 3692. The one or more computer resources 3692 may include but are not limited to circuitry, logic, field programmable gate array (FPGA), application specific integrated circuits (ASIC), programmable logic, a digital signal processor (DSP), a graphics processing unit (GPU), a microprocessor, a microcontroller, a big fat computer (BFC) or others, or clusters thereof. In some examples, the policy explorer (e.g., 3699) may be implemented in a planner of an autonomous vehicle.
Further to diagram 3600, the planner 3670 of autonomous vehicle 3630a may have calculated candidate trajectories 3651, each having a high confidence level associated with safely navigating autonomous vehicle 3630a around the potholes 3634 in road surface 3623. The planner 3670 may select one of the candidate trajectories 3651 for execution by the autonomous vehicle 3630a and the selected trajectory is denoted as 3652 (depicted in heavy line).
Due to the position of the potholes 3634 on the road surface 3623, the autonomous vehicle 3630a may not be able to avoid hitting one or both of the potholes; therefore, there is a possibility that hitting one or both of the potholes 3634 may cause damage to the autonomous vehicle 3630a and/or affect ride comfort of passengers being transported by the autonomous vehicle 3630a. In some examples, the selected trajectory 3652 may be calculated to allow the autonomous vehicle 3630a to continue its transit through region 3624 by traveling pass the potholes 3634. In other examples, the selected trajectory 3652 may be calculated to allow the autonomous vehicle 3630a to execute a safe stop trajectory to prevent the autonomous vehicle 3630a from potentially running over one or both of the potholes 3634 and being damaged and/or disturbing passengers. In yet other examples, the planner 3670 may implement one of an alternative set of candidate trajectories 3653 being configured to cause the autonomous vehicle 3630a to avoid the potholes 3634 (e.g., by changing lanes from lane 3266 to lane 3624). For example, the planner 3670 may select alternative candidate trajectory 3654 (denoted in heavy line) from a set of alternative candidate trajectories 3653 (e.g., based on a confidence level associated with each alternative candidate trajectory 3653). In other examples, an external system (e.g., a simulator or teleoperator) in communication with the autonomous vehicle 3630a may compute the alternative candidate trajectories 3653, rank the alternative candidate trajectories 3653 (e.g., ranked in order of confidence level from highest confidence level to lowest confidence level) and select one of the alternative candidate trajectories 3653 to be executed (e.g., alternative candidate trajectory 3654) by the autonomous vehicle 3630a under control of the planner 3670.
The data 3696 representing the telemetry data 3671 and/or policy data 3672 may be received by the policy explorer 3699. The policy explorer 3699 may be configured to implement generation of policy updated policy data 3694 (e.g., new policy data, revised policy data) based on the data 3696 (e.g., the telemetry data 3671, the policy data 3672 or both). The one or more computer resources 3692 may access one or more data repositories 3693, 3695, 3697, 3691) and may compare, compute, correlate or otherwise process data associated with data 3696 and one or more of the data repositories to generate updated policy data 3694. In some examples, the
Data repositories 3693, 3695, 3697, 3691 need not be internal to the policy explorer 3699 and some or all of the data repositories may be accessed from a resource external to the policy explorer 3699. Data repository 3693 may include data representative of sensor data, map data, local pose data, telemetry data, policy data and other data associated with autonomous vehicles in the of autonomous fleet. Data repository 3693 may include processed data and/or raw (e.g., unprocessed data). Data repository 3695 may include data representative of simulation data associated with one or more vehicles in the fleet of autonomous vehicles. Data repository 3697 may include data representative of teleoperations associated with one or more vehicles in the fleet of autonomous vehicles. Data repository 3691 may include other data associated with operation of a fleet of autonomous vehicles. Policy explorer 3699 may include and/or have access to more or fewer data repositories than depicted in diagram 3600, for example. Policy explorer 3699 may include and/or have access to data repositories having other types of data than described above in reference to diagram 3600, for example.
The policy explorer 3699 may communicate the updated policy data 3694 to one or more autonomous vehicles (e.g., 3630 or other vehicles in the fleet) and/or to the autonomous vehicle 3630a, for example. In other examples, the updated policy data 3694 may be communicated to a teleoperator, a simulator or both.
Updated policy data 3694 may include one or more candidate trajectories and the one or more candidate trajectories may have the same or different confidence levels associated with each candidate trajectory. Candidate trajectories may be arranged in a predetermined order or rank, such as a ranking order based on confidence level (e.g. ranked from highest confidence level to lowest confidence level or vice-versa).
Updated policy data 3694 may include other data being configured to cause a system that receives the data (e.g., an autonomous vehicle, a simulator, a teleoperator) to take an action, such as implementing a selected one of the one or more candidate trajectories, for example. In other examples, the other data may be configured to allow the system that receives the data to select, based on a criteria (e.g., ranking based on highest confidence level), one of the one or more candidate trajectories for implementation (e.g., to determine which of the one or more candidate trajectories is selected as the selected candidate trajectory). As one example, updated policy data 3694 may include data being configured to cause candidate trajectory 3654 to be selected by the system receiving the updated policy data 3694. As another example, updated policy data 3694 may include data being configured to cause the autonomous vehicle 3630a to decide which of the candidate trajectories to be selected based on a criteria (e.g., applied by planner 3670) that may be applied to data included in updated policy data 3694, such as the above mentioned confidence level associated with each candidate trajectory.
In some examples, a selected candidate trajectory (e.g., candidate trajectory 3654) may be selected based on a policy requiring the selected candidate trajectory having a confidence level predicted to provide a safe trajectory for the autonomous vehicle to navigate based on an event or condition that prompted the updated policy data to be generated. In other examples, a selected candidate trajectory (e.g., candidate trajectory 3654) may be selected based on a policy underlining a preferred customer experience (e.g., a smooth ride absent the jolting effects of potholes) or other metric associated with operation of the autonomous vehicle (e.g., averting potential damage due to an event such as potholes, road debris etc.).
In other examples, the updated policy data 3694 may be received by a system (e.g., autonomous vehicle 3630a) after the event has already occurred and policy data (e.g., policy data 3672) in place at the time of the event may have been used to determine a selected candidate trajectory (e.g., 3652) from one or more computed (e.g., by planner 367) candidate trajectories (3653). The later received updated policy data 3694 may be accessed to implement candidate trajectory determinations at a future time when another event (e.g., detection by a sensor system and semantic classification by a perception and/or planner of potholes 3634) occasions machine learning to consider alternative policy data to avoid or otherwise mitigate the effects of the event on operation of the autonomous vehicle. For example, the updated policy data 3694 may be accessed by a planner to compute candidate trajectories 3653, to compare the candidate trajectories 3653 with candidate trajectories 3651, and select (e.g., based on confidence levels) which candidate trajectory to select for guiding the path of the autonomous vehicle. The updated policy data 3694 may include data representing rules associated with events or semantic classifications of pothole objects detected along a computed path of an autonomous vehicle and the rule may determine that candidate trajectories 3651 be rejected in favor of candidate trajectories 3653 because candidate trajectories 3653 avoid the possibility of contact with the potholes 3634, for example.
At a stage 3704, data representing a confidence level associated with the event may be extracted from the policy data. For example, the confidence level may be ranked confidence level values associated with candidate trajectories computed by a planner of the autonomous vehicle.
At a stage 3706, a state of operation of the autonomous vehicle may be determined based on the confidence level. For example, the confidence level extracted at the stage 3704 may exceed a range of acceptable confidence levels associated with normative operation of the autonomous vehicle (e.g., an optimized path having a computed probability of facilitating collision-free travel of the autonomous vehicle, complying with traffic laws, providing a comfortable user experience or a comfortable user ride). As another example, the confidence level extracted at the stage 3704 may be indicative of confidence levels associated with non-normative operation of the autonomous vehicle due to possible trajectories that are insufficient to guarantee collision-free travel of the autonomous vehicle.
At a stage 3708, data representing candidate trajectories responsive to the event (e.g., trajectories that avoid contact with potholes 3634) may be calculated based on the telemetry data and the state of operation of the autonomous vehicle (e.g., normative or non-normative operation). Each candidate trajectory calculated at the stage 3708 may have an associated confidence level.
At a stage 3710, updated policy data associated with the event may be generated. The updated policy data may include the candidate trajectories.
At a stage 3712, the updated policy data may be communicated (e.g., transmitted). As one example, the updated policy data may be communicated to a planner of at least one autonomous vehicle (e.g., to one or more autonomous vehicles in a fleet of autonomous vehicles). In other examples, the updated policy data may be communicated to a teleoperator (e.g., being configured to influence path planning and/or navigational control in one or more autonomous vehicles). In yet other examples, the updated policy data may be communicated to a simulator (e.g., configured to simulate an autonomous vehicle in a synthetic environment). In some examples, the updated policy data may be communicated to a data store, memory (e.g., non-volatile memory), data repository or other data storage system.
One or more of the stages of flow chart 3700 may be implemented by a policy explorer (e.g., policy explorer 3699 of
Policy explorer 3899 may access functionality of physics processor 3850 and its associated sub-blocks 3851, 3852 and 3854, controller 3856 and evaluator 3858, for example. Policy explorer 3899 may access functionality of correlator 3861 to determine correlation between items of data, comparator 3862 to compare data with other data, statistical change point detector 3864 and/or probabilistic inference calculator 3863 to determine confidence intervals. Policy explorer 3899 may generate simulated policy data to be simulated in simulator 3840 and generate updated policy data 3894 based on analysis of simulation results, such as confidence levels associated with simulated candidate trajectories. For example, updated policy data 3894 may be determined, at least in part, by simulated candidate trajectories having high confidence levels that meet and/or exceed a predetermined value (e.g., 95% or higher confidence level).
Updated policy data 3894 may be communicated (e.g., via a communications network) to one or more destinations including but not limited to a teleoperator 3894, one or more autonomous vehicles 3830, and a data repository 3890, for example. Teleoperator 3894 may generate policy data 3897 based on the updated policy data 3894. Teleoperator 3894 may generate policy data 3897 based on other data, such as data 3896, updated policy data 3894 or both, for example.
Policy explorer 3899 of
According to some embodiments, the term “circuit” can refer, for example, to any system including a number of components through which current flows to perform one or more functions, the components including discrete and complex components. Examples of discrete components include transistors, resistors, capacitors, inductors, diodes, and the like, and examples of complex components include memory, processors, analog circuits, digital circuits, and the like, including field-programmable gate arrays (“FPGAs”), application-specific integrated circuits (“ASICs”). Therefore, a circuit can include a system of electronic components and logic components (e.g., logic configured to execute instructions, such that a group of executable instructions of an algorithm, for example, and, thus, is a component of a circuit). According to some embodiments, the term “module” can refer, for example, to an algorithm or a portion thereof, and/or logic implemented in either hardware circuitry or software, or a combination thereof (i.e., a module can be implemented as a circuit). In some embodiments, algorithms and/or the memory in which the algorithms are stored are “components” of a circuit. Thus, the term “circuit” can also refer, for example, to a system of components, including algorithms. These can be varied and are not limited to the examples or descriptions provided.
Although the foregoing examples have been described in some detail for purposes of clarity of understanding, the above-described inventive techniques are not limited to the details provided. There are many alternative ways of implementing the above-described invention techniques. The disclosed examples are illustrative and not restrictive.
1. A method, comprising:
- receiving, from one or more of a sensor or a communication interface, telemetry data associated with an event, the event being a situation or condition associated with operation of an autonomous vehicle;
- obtaining policy data, the policy data including instructions for operating an autonomous vehicle according to a trajectory and the policy data including a confidence level associated with the received trajectory, the confidence level indicating a degree of certainty that the autonomous vehicle, in operating according to the trajectory, will operate safely in view of the event;
- obtaining candidate trajectories responsive to the event, based on the telemetry data, each candidate trajectory having an associated confidence level;
- generating, by machine-learning by a processor and based at least in part on the candidate trajectories and the telemetry data, updated policy data that includes instructions for operating the autonomous vehicle responsive to the event differently than according to the policy data; and
- communicating the updated policy data to at least one autonomous vehicle via a communications interface.
3. The method of claim 1 further comprising:
- determining that the confidence level associated with the received trajectory is below a confidence level threshold, indicating a non-normative operational state of the autonomous vehicle.
4. The method of claim 1, wherein communicating the updated policy data to at least one autonomous vehicle comprises communicating the updated policy data to an autonomous vehicle other than the autonomous vehicle with which the event is associated.
5. The method of claim 4, wherein the telemetry data and policy data are obtained from the autonomous vehicle with which the event is associated.
6. The method of claim 1, wherein the telemetry data is received and the policy data is obtained at a teleoperator computing device and wherein the teleoperator computing device is configured to communicate the updated policy data to at least one autonomous vehicle.
8. The method of claim 1, wherein the telemetry data is received and the policy data is obtained at a simulator, the method further comprising, by the simulator:
- obtaining the candidate trajectories responsive to the event by simulating the candidate trajectories based at least in part on one or more of the telemetry data or the policy data;
- determining confidence levels associated with the simulated candidate trajectories;
- receiving selection of one of the candidate trajectories for implementation by the autonomous vehicle;
- providing the selected trajectory to the at least one autonomous vehicle as part of the updated policy data.
10. The method of claim 8 the method further comprising, by the simulator:
- conducting the generation of the updated policy data based at least in part on one or more of simulating the candidate trajectories or the selected trajectory; and
- communicating one or more of the policy data or the updated policy data to one or more of: a teleoperator computing device, or the at least one autonomous vehicle.
13. The method of claim 1, wherein the telemetry data is received and the policy data is obtained at a simulator configured to simulate an autonomous vehicle in a synthetic environment.
14. The method of claim 1, wherein the telemetry data is received, the policy data is obtained, and the updated policy data is generated at an autonomous vehicle.
15. The method of claim 1, wherein the policy data is obtained from one or more of an autonomous vehicle that captured the telemetry data or a data repository associated with a service for a fleet of autonomous vehicles.
16. The method of claim 1, further comprising:
- determining that the at least one autonomous vehicle has operated according to the teleoperation trajectory;
- obtaining further telemetry data from the at least one autonomous vehicle;
- determining from at least the further telemetry data that operation of the at least one autonomous vehicle according to the teleoperation trajectory and responsive to the event resulted in normative operation of the at least one autonomous vehicle;
- generating, by machine-learning, the updated policy responsive to the determining that operating the at least one autonomous vehicle according to the teleoperation trajectory resulted in normative operation; and
- communicating the updated policy data to at least one autonomous vehicle.
17. The method of claim 1, wherein the policy data includes one or more of:
- criteria for selecting a candidate trajectory for operating the autonomous vehicle;
- rules defining constraints for generating trajectories;
- threshold confidence levels for determining whether operations of the autonomous vehicle are normative;
- a classifier for detecting, from the telemetry data, objects in an environment in which the autonomous vehicle is operating;
- a semantic classification of an object detected from the telemetry data; or
- a semantic classification associated with the event; and
- wherein the updated policy data includes an update to one or more of the criteria, the rules, the threshold confidence levels, the semantic classification of the object, or the semantic classification associated with the event.
18. The method of claim 3, wherein at least one of the associated confidence levels is greater than the confidence level associated with the received trajectory.
19. The method of claim 3, wherein the candidate trajectories are obtained at least in part responsive to the confidence level associated with the received trajectory indicating the non-normative operational state.
20. A system comprising:
- one or more processors;
- a memory having stored thereon first criteria for controlling motion of an autonomous vehicle and a planner module executable by the one or more processors and that, when executed by the one or more processors, configure the system to perform operations including: obtaining sensor data related to operation of an autonomous vehicle and an event; determining, based at least on the sensor data and the first criteria, a first trajectory for operation of the autonomous vehicle responsive to the event; receiving instructions for implementing a second trajectory for operation of the autonomous vehicle responsive to the event; and learning, from at least the instructions and the sensor data, second criteria to update the first criteria, the second criteria configuring the one or more processors to determine the second trajectory based at least on the sensor data and the instructions.
21. A system as claim 20 recites, the first criteria including one or more of:
- criteria for selecting a candidate trajectory for operating the autonomous vehicle;
- rules defining constraints for generating trajectories;
- threshold confidence levels for determining whether operations of the autonomous vehicle are normative;
- a classifier for detecting, from the sensor data, objects in an environment in which the autonomous vehicle is operating;
- a semantic classification of an object detected by the perception engine from the sensor data;
- a semantic classification associated with the event;
- one or more candidate trajectories; or
- confidence levels associated with the candidate trajectories or rules for determining confidence levels associated with the candidate trajectories;
- and wherein the second criteria includes a modification of or addition to one or more of:
- the criteria for selecting a candidate trajectory,
- the rules defining constraints for generating trajectories,
- the threshold confidence levels,
- the perception engine configuration,
- the semantic classifications,
- the candidate trajectories,
- the confidence levels, or
- rules for determining confidence levels.
22. The system as claim 20 recites, the system comprising an autonomous vehicle and further including one or more of:
- one or more sensors for obtaining the sensor data or
- a communication interface for receiving the instructions.
23. A system as claim 20 recites, the operations further including simulating in a synthetic environment candidate trajectories based at least in part on one or more of the sensor data and the first criteria, and wherein receiving a second trajectory for operation of the autonomous vehicle includes receiving selection of one of the simulated candidate trajectories.
24. The system as claim 20 recites, the system comprising a teleoperator computing device, the teleoperator computing device being configured to control operations of the autonomous vehicle.
25. One or more computer-readable media storing processor-executable instructions that, when executed, cause one or more processors to perform operations, the operations comprising:
- receiving telemetry data associated with operation of an autonomous vehicle;
- receiving instructions for operating the autonomous vehicle responsive to the telemetry data, the instructions including instructions for the autonomous vehicle to operate according to a defined trajectory;
- learning updated criteria for operating the autonomous vehicle from the instructions and the telemetry data; and
- communicating the updated criteria to at least one autonomous vehicle to update criteria of the at least one autonomous vehicle, the updated criteria and the criteria of the at least one autonomous vehicle including instructions for the at least one autonomous vehicle to determine a trajectory of the at least one autonomous vehicle, and the instructions of the updated criteria configuring the at least one autonomous vehicle to conduct the determination differently than the instructions of the criteria of the at least one autonomous vehicle configure the at least one autonomous vehicle to conduct the determination.
26. One or more non-transitory computer-readable media as claim 25 recites, the operations further comprising:
- generating a plurality of candidate trajectories for the autonomous vehicle responsive to the event;
- determining confidence levels corresponding to respective ones of the plurality of candidate trajectories, a confidence level indicating a degree of certainty that the autonomous vehicle, in operating according to one of the plurality of candidate trajectories, will operate safely in view of the event; and
- at least one of: rank the plurality of candidate trajectories according to confidence levels of the plurality of candidate trajectories, and select, based at least in part on the ranking, the confidence levels, and an acceptable confidence level specified by the criteria of the at least one autonomous vehicle or the updated criteria, respectively, a trajectory from the plurality of candidate trajectories for implementation by the autonomous vehicle; or transmit a request for teleoperation of the autonomous vehicle if none of the plurality of candidate trajectories is associated with a confidence level that meets the acceptable confidence level.
27. One or more non-transitory computer-readable media as claim 26 recites, determining the confidence levels being based at least in part on the criteria of the at least one autonomous vehicle or the updated criteria, respectively, and determining a confidence level for a respective candidate trajectory in view of the event and the determining including one or more of:
- determining a probability that by operating according to the candidate trajectory, the autonomous vehicle will operate safely in view of the event;
- determining whether the candidate trajectory comports with traffic laws;
- determining whether the candidate trajectory violates the traffic laws;
- determining consequences of violating the traffic laws;
- determining consequences of not violating the traffic laws; or
- determining whether the candidate trajectory violates a passenger comfort policy,
- wherein the acceptable confidence level is a minimum required confidence level to permit the autonomous vehicle to implement the candidate trajectory.
28. One or more non-transitory computer-readable media as claim 25 recites, the one or more computer-readable media being a memory of an autonomous vehicle, a teleoperator computing device, a computing device storing a simulator, or a fleet management service device.