Abstract: Automated training dataset generators that generate feature training datasets for use in real-world autonomous driving applications based on virtual environments are disclosed herein. The feature training datasets may be associated with training a machine learning model to control real-world autonomous vehicles. In some embodiments, an occupancy grid generator is used to generate an occupancy grid indicative of an environment of an autonomous vehicle from an imaging scene that depicts the environment. The occupancy grid is used to control the vehicle as the vehicle moves through the environment. In further embodiments, a sensor parameter optimizer may determine parameter settings for use by real-world sensors in autonomous driving applications.

Abstract: A computer-implemented method of determining a relative velocity between a vehicle and an object. The method includes receiving sensor data generated by one or more sensors of the vehicle configured to sense an environment following a scan pattern. The method also includes obtaining, based on the sensor data, a point cloud frame. The point cloud frame comprises a plurality of points of depth data and a time at which the depth data was captured. Additionally, the method includes selecting two or more points of the scan pattern that overlap the object. The selected points are located on or near a two-dimensional surface corresponding to the object, and the depth data for two or more of the selected points are captured at different times. The method includes calculating the relative velocity between the vehicle and the object based on the depth data and capture times associated with the selected points.

Abstract: A computer-implemented method of determining relative velocity between a vehicle and an object. The method includes receiving sensor data generated by one or more sensors of the vehicle configured to sense an environment by following a scan pattern comprising component scan lines. The method includes obtaining, based on the sensor data, a point cloud frame. Additionally, the method includes identifying a first pixel and a second pixel that are co-located within a field of regard and overlap a point cloud object within the point cloud frame and calculating a difference between a depth associated with the first pixel and a depth associated with the second pixel. The method includes determining a relative velocity of the point cloud object by dividing the difference in depth data by a time difference between when the depth associated with the first pixel was sensed and the depth associated with the second pixel was sensed.

Abstract: A computer-implemented method of determining relative velocity between a vehicle and an object. The method includes receiving sensor data generated by one or more sensors of the vehicle. The one or more sensors are configured to sense an environment through which the vehicle is moving by following a scan pattern comprising component scan lines. The method includes obtaining, by one or more processors, a point cloud frame based on the sensor data and representative of the environment and identifying, by the one or more processors, a point cloud object within the point cloud frame. The method further includes determining, by the one or more processors, that the point cloud object is skewed relative to an expected configuration of the point cloud object, and determining, by the one or more processors, a relative velocity of the point cloud object by analyzing the skew of the object.

Abstract: A computer-implemented method of determining a relative velocity between a vehicle and an object. The method includes receiving sensor data generated by one or more sensors of the vehicle configured to sense an environment following a scan pattern. The method also includes obtaining, based on the sensor data, a point cloud frame. The point cloud frame comprises a plurality of points of depth data and a time at which the depth data was captured. Additionally, the method includes selecting two or more points of the scan pattern that overlap the object. The selected points are located on or near a two-dimensional surface corresponding to the object, and the depth data for two or more of the selected points are captured at different times. The method includes calculating the relative velocity between the vehicle and the object based on the depth data and capture times associated with the selected points.

Abstract: A method for configuring a perception component of a vehicle having one or more sensors includes generating a first set of training data that includes first sensor data corresponding to a first setting of one or more sensor parameters, and an indicator of the first setting. The method also includes generating a second set of training data that includes second sensor data corresponding to a second setting of the sensor parameter(s), and an indicator of the second setting. The method further includes training the perception component, at least by training a machine learning based model using the first and second training data sets. The trained perception component is configured to generate signals descriptive of a current state of the vehicle environment by processing sensor data generated by the sensor(s), and one or more indicators indicating which setting of the sensor parameter(s) corresponds to which portions of the generated sensor data.

Abstract: A method for controlling a first sensor configured to sense an environment through which a vehicle is moving includes receiving sensor data generated by one or more sensors of the vehicle as the vehicle moves through the environment, identifying, by one or more processors and based on at least a portion of the received sensor data, one or more road portions along which the vehicle is expected to travel, and determining, by one or more processors, a configuration of the identified road portions, at least in part by determining a slope of at least one of the identified road portions. The method also includes determining, by one or more processors analyzing at least the determined configuration, an elevation of a field of regard of the first sensor that satisfies one or more visibility criteria, and causing the first sensor to be adjusted in accordance with the determined elevation.

Abstract: A computer-readable medium stores instructions executable by one or more processors to implement an aggregate self-driving control architecture (SDCA) for controlling an autonomous vehicle. The aggregate SDCA includes a plurality of SDCAs each including a different motion planner. Each motion planner is configured to receive signals descriptive of a current state of an environment through which the autonomous vehicle is moving, and each SDCA is configured to generate candidate decisions for controlling the autonomous vehicle by using the respective motion planner to process the received signals. The aggregate SDCA also includes a decision arbiter configured to receive the candidate decisions generated by the SDCAs, generate decisions for controlling the autonomous vehicle by processing the candidate decisions, and provide signals indicative of the generated decisions to one or more operational subsystems of the vehicle to effectuate maneuvering of the vehicle.

Abstract: Various software techniques for managing operation of autonomous vehicles based on sensor data are disclosed herein. A computing system may generate, based on a set of signals descriptive of a current state of an environment in which the autonomous vehicle is operating, a normal path plan separate from a safe path plan, or a hybrid path plan including a normal path plan and a safe path plan. In generating the safe path plan, the computing system may generate and concatenate a set of motion primitives. When a fault condition occurs, the computing device may transition from executing the normal path plan to executing the safe path plan to safely stop the autonomous vehicle.

Abstract: A method for controlling at least a first sensor of a vehicle, which senses an environment through which the vehicle is moving by producing a plurality of scan lines arranged according to a spatial distribution, includes receiving sensor data generated by one or more sensors. The one or more sensors are configured to sense the environment through which the vehicle is moving. The method also includes identifying, by one or more processors and based on the received sensor data, one or more areas of interest in the environment, and causing, by one or more processors and based on the areas of interest, the spatial distribution of the plurality of scan lines produced by the first sensor to be adjusted.

Abstract: Various software techniques for managing operation of autonomous vehicles based on sensor data are disclosed herein. A computing system may generate, based on a set of signals descriptive of a current state of an environment in which the autonomous vehicle is operating, a normal path plan separate from a safe path plan, or a hybrid path plan including a normal path plan and a safe path plan. In generating the safe path plan, the computing system may generate and concatenate a set of motion primitives. When a fault condition occurs, the computing device may transition from executing the normal path plan to executing the safe path plan to safely stop the autonomous vehicle.

Abstract: A method for processing point clouds having variable spatial distributions of scan lines includes receiving a point cloud frame generated by a sensor configured to sense a vehicle environment. Each of the points in the frame has associated two-dimensional coordinates and an associated parameter value. The method also includes generating a normalized point cloud frame by adding interpolated points not present in the received frame, at least by, for each interpolated point, identifying one or more neighboring points having associated two-dimensional coordinates that are within a threshold distance of two-dimensional coordinates for the interpolated point, and calculating an estimated parameter value of the interpolated point using, for each of the identified neighboring points, a distance between the two-dimensional coordinates and the parameter value associated with the identified neighboring point.

Abstract: A method for controlling at least a first vehicle sensor includes receiving sensor data generated by one or more vehicle sensors that are configured to sense an environment through which the vehicle is moving, and identifying, based on the received sensor data, one or more current and/or predicted positions of one or more dynamic objects that are currently moving, or are capable of movement, within the environment. The method also includes causing, based on the current and/or predicted positions of the dynamic objects, an area of focus of the first sensor to be adjusted, at least by causing (i) a field of regard of the first sensor, and/or (ii) a spatial distribution of scan lines produced by the first sensor, to be adjusted.

Abstract: A non-transitory computer-readable medium stores instructions executable by one or more processors to implement a sensor control architecture for controlling at least a first sensor of a vehicle. The sensor control architecture is configured to receive sensor data generated by one or more sensors of the vehicle. The one or more sensors are configured to sense an environment through which the vehicle is moving. The sensor control architecture is also configured to determine, based on the received sensor data and using an attention model that is trained using a machine learning technique, one or more sensor settings, and to cause one or more sensor parameters of the first sensor to be adjusted in accordance with the determined sensor settings. The one or more sensor parameters include at least one sensor parameter that defines an area of focus for the first sensor.

Abstract: A method for configuring a perception component of a vehicle having one or more sensors includes generating a first set of training data that includes first sensor data corresponding to a first setting of one or more sensor parameters, and an indicator of the first setting. The method also includes generating a second set of training data that includes second sensor data corresponding to a second setting of the sensor parameter(s), and an indicator of the second setting. The method further includes training the perception component, at least by training a machine learning based model using the first and second training data sets. The trained perception component is configured to generate signals descriptive of a current state of the vehicle environment by processing sensor data generated by the sensor(s), and one or more indicators indicating which setting of the sensor parameter(s) corresponds to which portions of the generated sensor data.

Abstract: A method for processing point clouds having variable spatial distributions of scan lines includes receiving a point cloud frame generated by a sensor configured to sense an environment through which a vehicle is moving. The point cloud frame includes scan lines arranged according to a particular spatial distribution. The method also includes either generating an enhanced point cloud frame with a larger number of points than the received point cloud frame, or constructing, by one or more processors and based on points of the received point cloud frame, a three-dimensional mesh. The method also includes generating, by performing an interpolation function on the enhanced point cloud frame or a virtual surface provided by the three-dimensional mesh, a normalized point cloud frame, and generating, using the normalized point cloud frame, signals descriptive of a current state of the environment through which the vehicle is moving.

Abstract: A method for controlling a vehicle based on sensor data having variable sensor parameter settings includes receiving sensor data generated by a vehicle sensor while the sensor is configured with a first sensor parameter setting. The method also includes receiving an indicator specifying the first sensor parameter setting, and selecting, based on the received indicator, one of a plurality of neural networks of a perception component, each neural network having been trained using training data corresponding to a different sensor parameter setting. The method also includes generating signals descriptive of a current state of the environment using the selected neural network and based on the received sensor data. The method further includes generating driving decisions based on the signals descriptive of the current state of the environment, and causing one or more operational subsystems of the vehicle to maneuver the vehicle in accordance with the generated driving decisions.

Abstract: A method for processing point clouds having variable spatial distributions of scan lines includes receiving a point cloud frame generated by a sensor configured to sense a vehicle environment. Each of the points in the frame has associated two-dimensional coordinates and an associated parameter value. The method also includes generating a normalized point cloud frame by adding interpolated points not present in the received frame, at least by, for each interpolated point, identifying one or more neighboring points having associated two-dimensional coordinates that are within a threshold distance of two-dimensional coordinates for the interpolated point, and calculating an estimated parameter value of the interpolated point using, for each of the identified neighboring points, a distance between the two-dimensional coordinates and the parameter value associated with the identified neighboring point.

Abstract: A method for controlling at least a first sensor of a vehicle, which senses an environment through which the vehicle is moving by producing a plurality of scan lines arranged according to a spatial distribution, includes receiving sensor data generated by one or more sensors. The one or more sensors are configured to sense the environment through which the vehicle is moving. The method also includes identifying, by one or more processors and based on the received sensor data, one or more areas of interest in the environment, and causing, by one or more processors and based on the areas of interest, the spatial distribution of the plurality of scan lines produced by the first sensor to be adjusted.

Abstract: A method for processing point clouds having variable spatial distributions of scan lines includes receiving a point cloud portion corresponding to an object in a vehicle environment, the point cloud portion including scan lines arranged according to a particular spatial distribution. The method also includes constructing a voxel grid corresponding to the received point cloud portion. The voxel grid includes a plurality of volumes in a stacked, three-dimensional arrangement, and constructing the voxel grid includes (i) determining an initial classification of the object, (ii) setting one or more parameters of the voxel grid based on the initial classification, and (iii) associating each volume of the plurality of volumes with an attribute specifying how many points, from the point cloud portion, fall within that volume. The method also includes generating, using the constructed voxel grid, signals descriptive of a current state of the environment through which the vehicle is moving.