Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a student machine learning model using a teacher machine learning model that has a pre-trained feature extractor.

Abstract: Aspects and implementations of the present disclosure relate, generally, to optimization of autonomous vehicle (AV) technology and, more specifically, to a transfer hub for autonomous trucking operations. In one example, the disclosed techniques include obtaining, at a first time, one or more first images of an outside environment of an AV and shutting down a data processing system of the AV. The techniques further include initiating a starting-up of the data processing system of the AV and obtaining, at a second time, one or more second images of the outside environment of the AV. The techniques further include determining, based on a comparison of the one or more first images to the one or more second images, that the AV has not moved between the first time and the second time and completing the starting-up of the data processing system of the AV.

Abstract: The technology relates to developing a highly accurate understanding of a vehicle's sensor fields of view in relation to the vehicle itself. A training phase is employed to gather sensor data in various situations and scenarios, and a modeling phase takes such information and identifies self-returns and other signals that should either be excluded from analysis during real-time driving or accounted for to avoid false positives. The result is a sensor field of view model for a particular vehicle, which can be extended to other similar makes and models of that vehicle. This approach enables a vehicle to determine when sensor data is of the vehicle or something else. As a result, the detailed modeling allowing the on-board computing system to make driving decisions and take other actions based on accurate sensor information.

Abstract: The described aspects and implementations enable efficient object detection and tracking. In one implementation, disclosed is a method and a system to perform the method, the system including the sensing system configured to obtain sensing data characterizing an environment of the vehicle. The system further includes a data processing system operatively coupled to the sensing system and configured to process the sensing data using a first (second) set of neural network (NN) layers to obtain a first (second) set of features for a first (second) region of the environment, the first (second) set of features is associated with a first (second) spatial resolution. The data processing system is further to process the two sets of features using a second set of NN layers to detect a location of object(s) in the environment of the vehicle and a state of motion of the object(s).

Abstract: Methods and systems for the use of detected objects for image processing are described. A computing device autonomously controlling a vehicle may receive images of the environment surrounding the vehicle from an image-capture device coupled to the vehicle. In order to process the images, the computing device may receive information indicating characteristics of objects in the images from one or more sources coupled to the vehicle. Examples of sources may include RADAR, LIDAR, a map, sensors, a global positioning system (GPS), or other cameras. The computing device may use the information indicating characteristics of the objects to process received images, including determining the approximate locations of objects within the images. Further, while processing the image, the computing device may use information from sources to determine portions of the image to focus upon that may allow the computing device to determine a control strategy based on portions of the image.

Abstract: The technology involves identifying suitable pickup and drop-off locations based on detected pedestrian walking paths. Mapped areas have specific physical configurations, which may suggest places to pick up or drop off a rider (or a delivery). A walking path heatmap can be generated based on obtained historical and/or real-time pedestrian-related information, which can be obtained by autonomous vehicles driving in areas of interest. Incorporating heatmap information into the evaluation, the system identifies locations for optimized pickup or drop-off in accordance with where pedestrians would likely go. One aspect involves classifying different objects, for instance identifying one or more objects as people who may be walking versus riding a bicycle. Once classified, information about the paths is used to obtain a the heatmap associated with the walking paths.

Abstract: A system and method include scanning a light detection and ranging (LIDAR) device through a range of orientations corresponding to a scanning zone while emitting light pulses from the LIDAR device. The method also includes receiving returning light pulses corresponding to the light pulses emitted from the LIDAR device and determining initial point cloud data based on time delays between emitting the light pulses and receiving the corresponding returning light pulses and the orientations of the LIDAR device. The initial point cloud data has an initial angular resolution. The method includes identifying, based on the initial point cloud data, a reflective feature in the scanning zone and determining an enhancement region and an enhanced angular resolution for a subsequent scan to provide a higher spatial resolution in at least a portion of subsequent point cloud data from the subsequent scan corresponding to the reflective feature.

Abstract: The described aspects and implementations enable efficient utilization of multiple machine learning (ML) models with autonomous vehicles (AVs) to quickly and efficiently determine trajectory information of objects in a driving environment. During a time interval, a perception system of an AV generates, using a first ML model, an embedding based on data that characterizes an object's trajectory in the driving environment. At each sub-interval of multiple sub-intervals during the time interval, a second ML model uses the embedding to generate a probability distribution for the trajectory of the object, and a planning system of the AV generates an update to the AV's trajectory based on the probability distribution.

Abstract: Example embodiments relate techniques and systems for generating multi-axis radar velocity images using stereo radar. A vehicle radar system receives radar first radar data from a first radar and second radar data from a second radar, which are coupled at different locations on a vehicle traveling in an environment. The system determines a first radial speed and a second radial speed for an object based on the first and second radar data, respectively, and then estimates a velocity vector for the object relative to the vehicle based on the first and second radial speeds. The system can provide the estimated velocity vector as an input into a neural network and enable vehicle systems to control the vehicle based on the output from the neural network.

Abstract: Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for intervention behavior prediction. A method includes receiving data characterizing a scene that includes a first agent and a second agent and receiving intervention data specifying a planned intervention to be performed by the second agent. A conditional behavior prediction output that assigns, to each of a plurality of possible future behaviors, (i) a respective conditional likelihood that the first agent performs the possible future behavior given that the second agent performs the planned intervention and (ii) a predicted value of a confounder variable for the possible future behavior is-generated using a conditional behavior prediction model.

Abstract: A method includes obtaining, by a processing device, input data derived from a set of sensors of an autonomous vehicle (AV), generating, by the processing device using a set of lane detection classifier heads, at least one heatmap based on a fused bird's eye view (BEV) feature generated from the input data, obtaining, by the processing device, a set of polylines using the at least one heatmap, wherein each polyline of the set of polylines corresponds to a respective track of a first set of tracks for a first frame, and generating, by the processing device, a second set of tracks for a second frame after the first frame by using a statistical filter based on a set of extrapolated tracks for the second frame and a set of track measurements for the second frame, wherein each track measurement of the set of track measurements corresponds to a respective updated polyline obtained for the second frame.

Abstract: The technology relates to maneuvering a vehicle prior to making a turn from a current lane of the vehicle. As an example, a route that includes making the turn from the lane is identified. An area of the lane prior to the turn having a lane width of at least a predetermined size is also identified. Sensor data identifying an object within the lane is received from a perception system of a vehicle. Characteristics of the object, including a location of the object relative to the vehicle, are identified from the sensor data. The vehicle is then maneuvered through the area prior to making the turn using the identified characteristics.

Abstract: Aspects of the disclosure relate to methods for controlling a vehicle having an autonomous driving mode. For instance, sensor data may be received from one or more sensors of the perception system of the vehicle, the sensor data identifying characteristics of an object perceived by the perception system. When it is determined that the object is no longer being perceived by the one or more sensors of the perception system, predicted characteristics for the object may be generated based on one or more of the identified characteristics. The predicted characteristics of the object may be used to control the vehicle in the autonomous driving mode such that the vehicle is able to respond to the object when it is determined that the object is no longer being perceived by the one or more sensors of the perception system.

Abstract: An example method involves detecting a sensor-testing trigger. Detecting the sensor-testing trigger may comprise determining that a vehicle is within a threshold distance to a target in an environment of the vehicle. The method also involves obtaining sensor data collected by a sensor of the vehicle after the detection of the sensor-testing trigger. The sensor data is indicative of a scan of a region of the environment that includes the target. The method also involves comparing the sensor data with previously-collected sensor data indicating detection of the target by one or more sensors during one or more previous scans of the environment. The method also involves generating performance metrics related to the sensor of the vehicle based on the comparison.

Abstract: Described examples relate to an apparatus comprising a first sensor configured to scan an area of interest during a first time period and a second sensor configured to capture a plurality of images of a field of view. The apparatus may include at least one controller configured to receive the plurality of images captured by the second sensor, compare the timestamp information associated with at least one image of the plurality of images to at least one time period of the first time period, and select a base image from the plurality of images based on the comparison.

Abstract: Example implementations may relate to determining a strategy for a drop process associated with a light detection and ranging (LIDAR) device. In particular, the LIDAR device could emit light pulses and detect return light pulses, and could generate a set of data points representative of the detected return light pulses. The drop process could involve a computing system discarding data point(s) of the set and/or preventing emission of light pulse(s) by the LIDAR device. Accordingly, the computing system could detect a trigger to engage in the drop process, and may responsively (i) use information associated with the environment around the vehicle, operation of the vehicle, and/or operation of the LIDAR device as a basis to determine the strategy for the drop process, and (ii) engage in the drop process in accordance with the determined strategy.