Abstract: The present disclosure provides methods and techniques for evaluating and improving algorithms for autonomous driving planning and control (PNC), using one or more metrics (e.g., similarity scores) computed based on expert demonstrations. For example, the one or more metrics allow for improving PNC based on human, as opposed to or in addition to optimizing certain oversimplified properties, such as the least distance or time, as an objective. When driving in certain scenarios, such as taking a turn, people may drive in a distributed probability pattern instead of in a uniform line (e.g., different speeds and different curvatures at the same corner). As such, there can be more than one “correct” control trajectory for an autonomous vehicle to perform in the same turn. Safety, comfort, speeds, and other criteria may lead to different preferences and judgment as to how well the controlled trajectory has been computed.

Abstract: According to some embodiments, systems, methods and media for dynamically generating scenario parameters for an autonomous driving vehicles (ADV) are described. In one embodiment, when an ADV enters a driving scenario, the ADV can invoke a map-based scenario checker to determine the type of scenario, and invokes a corresponding neural network model to generate a set of parameters for the scenario based on real-time environmental conditions (e.g., traffics) and vehicle status information (e.g., speed). The set of scenario parameters can be a set of extra constraints for configuring the ADV to drive in a driving mode corresponding to the scenario.

Abstract: Embodiments of a methodology for generalized evolutionary training of a neural network may comprise: (i) obtaining a set of model snapshots by training a set of input models until at least one snapshot condition is satisfied for each input model from the set of input models, wherein each model snapshot comprises values of model components of its respective input model when the at least one snapshot condition was satisfied; (ii) generating model snapshot evaluation results by evaluating performance of each model snapshot; (iii) based upon the model snapshot evaluation results, selecting one or more parent models from the set of model snapshots; (iv) generating one or more child models by perturbing at least one or more model components of a parent model from the one or more parent models; and (v) setting the one or more child models as the set of input models for a subsequent training iteration.

Abstract: In one embodiment, an exemplary method includes the operations of receiving, at a profiling application, a record file recorded by the ADV for a driving scenario in an area, and a high definition map matching the area; extracting planning messages and perception messages from the record file; and aligning the planning message and the perception messages based on their timestamps. The method further includes calculating an individual performance score for each planning cycle of the ADV for the driving scenario based on the planning messages; calculating a weight for each planning cycle based on the perception messages and the high definition map; and then calculating a weighted score for the driving scenario based on individual performance scores and their corresponding weights.

Abstract: Systems, methods, and media for factoring localization uncertainty of an ADV into its planning and control process to increase the safety of the ADV. The uncertainty of the localization can be caused by sensor inaccuracy, map matching algorithm inaccuracy, and/or speed uncertainty. The localization uncertainty can have negative impact on trajectory planning and vehicle control. Embodiments described herein are intended to increase the safety of the ADV by considering localization uncertainty in trajectory planning and vehicle control. An exemplary method includes determining a confidence region for an ADV that is automatically driving on a road segment based on localization uncertainty and speed uncertainty; determining that an object is within the confidence region, and a probability of collision with the ADV based on a distance of the object to the ADV; and planning a trajectory based on the probability of collision, and controlling the ADV based on the probability of collision.

Abstract: Disclosed are performance metrics for evaluating the accuracy of a dynamic model in predicting the trajectory of ADV when simulating the behavior of the ADV under the control commands. The performance metrics may indicate the degree of similarity between the predicted trajectory of the dynamic model and the actual trajectory of the vehicle when applied with identical control commands. The performance metrics measure deviations of the predicted trajectory of the dynamic model from the actual trajectory based on the ground truths. The performance metrics may include cumulative or mean absolute trajectory error, end-pose difference (ED), two-sigma defect rate (?2?), the Hausdirff Distance (HAU), the longest common sub-sequence error (LCSS), or dynamic time warping (DTW). The two-sigma defect rate represents the ratio of the number of points with true location error falling out of the 2? range of the predicted location error over the total number of points in the trajectory.

Abstract: An obstacle is detected based on sensor data obtained from a plurality of sensors of the ADV. A distribution of a plurality of positions of the obstacle at a point of time may be predicted. A range of positions of the plurality of positions of the obstacle may be determined based on a confidence level of the range. A modified shape with a modified length of the obstacle may be determined based on the range of positions of the obstacle. A trajectory of the ADV based on the modified shape with the modified length of the obstacle may be planned. The ADV may be controlled to drive according to the planned trajectory to drive safely to avoid a collision with the obstacle.

Abstract: According to some embodiments, described herein is a method and a system for guaranteeing safety at a control level of an ADV when at least a portion of a planned path generated by a planning module of the ADV is uncertain due to traffics and/or road condition changes. The planning module, when generating a path, also generate a confidence level of each segment of the path based on one or more of perception data, map information, or traffic rules. The confidence levels are decreasing further away from the ADV. When the control module of the ADV obtains the path and the associated confidence levels, the control module issue control commands to track only one or two segments whose confidence levels exceeds a threshold hold, and issue default control commands for the rest of the path.

Abstract: According to various embodiments, systems, methods, and media for evaluating an open space planner in an autonomous vehicle are disclosed. In one embodiment, an exemplary method includes receiving, at a profiling application, a record file recorded by the ADV while driving in an open space using the open space planner, and a configuration file specifying parameters of the ADV; extracting planning messages and prediction messages from the record file, each extracted message being associated with the open space planner. The method further includes generating features from the planning message and the prediction messages in view of the specified parameters of the ADV; and calculating statistical metrics from the features. The statistical metrics are then provided to an automatic tuning framework for tuning the open space planner.

Abstract: According to some embodiments, systems, methods, and media for operating an autonomous driving vehicle (ADV) encountered with small objects are described. According to a method, the ADV, when detecting an object in a lane in which the ADV is travelling, can determine whether the ADV can safely drive over the object based on attributes of the object and attributes of the ADV, and if so, can generate one or more planned trajectories that each enable the ADV to drive over the object. From the one or more planned trajectories, the ADV can select a planned trajectory that enables the ADV to drive over the object along the centerline of the ADV without causing the ADV to drive out of the lane. If no such planned trajectory exists, the ADV can bypass the object, or stop within a predetermined distance of the object.

Abstract: Embodiments of the invention are provided to automatically generate corner simulation scenarios. In an embodiment, an exemplary method includes performing the following operations for a predetermined number of iterations for each set of predefined parameters. The operations include generating a set of parameter values for the set of predefined parameters; determining whether the set of parameter values is valid or invalid based on a set of predefined metrics; and if the set of parameter values is valid, performing a simulation task to simulate a trajectory planner of the ADV in a simulation scenario configured by the set of parameter values. The method further includes calculating a performance score for the simulation task; and if the performance score of the simulation task is below a predetermined threshold, saving the set of parameter values in a storage, wherein the set of parameter values is used for re-tuning the trajectory planner.

Abstract: According to some embodiments, systems, methods and media for operating an autonomous driving vehicles (ADV) in an unforeseen scenario are disclosed. In one embodiment, an exemplary method includes determining that the ADV has entered an unforeseen scenario; and identifying one or more surrounding vehicles that are navigating the unforeseen scenario. The method further includes generating a trajectory by mimicking driving behaviors of one or more of the one or more surrounding vehicles; and operating the ADV to follow the trajectory to navigate the unforeseen scenario.

Abstract: A system perceives an environment of an autonomous driving vehicle (ADV) based on a plurality of sensors and map data. The system determines an obstacle in the perceived environment to be a moving vehicle and the moving vehicle is to a left lane, to a right lane, or in front of the ADV. The system performs an inference on the obstacle using a neural network model to determine whether a behavior of the obstacle is anomalous. The system determines the obstacle is anomalous based on the performed inference.

Abstract: According to various embodiments, the disclosure discloses systems, methods and media for formulating training datasets for learning-based components in an autonomous driving vehicle (ADV). In an embodiment, an exemplary method includes allocating training datasets for training a learning-based model in the ADV, each training dataset being allocated to one of multiple predefined driving scenarios; determining a weight of each training dataset out of the training datasets; and optimizing the weight of each training dataset in one or more iterations according to a predetermined algorithm until a performance of the learning-based model reaches a predetermined threshold. The predetermined algorithm is one of a random search algorithm, a grid search algorithm, or a Bayesian algorithm.

Abstract: A trajectory of an obstacle is predicted by a prediction module of the ADV. A trajectory of the ADV is determined based on the trajectory of the obstacle by a planning module of the ADV. A loss function of an analysis model of the prediction module is decomposed to multiple components with multiple weightings to generate a weighted loss function based on the trajectory of the ADV. A performance of the prediction module is evaluated based on the weighted loss function to improve the performance of the prediction module to increase a safety and comfort of the ADV.

Abstract: A plurality of trajectories of a plurality of obstacles are predicted, at an autonomous driving simulation platform, by a prediction module of an autonomous driving vehicle (ADV). A trajectory of the ADV is planned, at the autonomous driving simulation platform, by a planning module of the ADV based on the plurality of trajectories of the plurality of obstacles. A performance of the planning module is determined based on one or more evaluation metrics regarding the trajectory of the ADV. A performance of the prediction module is evaluated based on the performance of the planning module to improve the performance of the prediction module to deploy the prediction module to the ADV to drive autonomously.

Abstract: The present disclosure provides methods and techniques for evaluating and improving algorithms for autonomous driving planning and control (PNC), using one or more metrics (e.g., similarity scores) computed based on expert demonstrations. For example, the one or more metrics allow for improving PNC based on human, as opposed to or in addition to optimizing certain oversimplified properties, such as the least distance or time, as an objective. When driving in certain scenarios, such as taking a turn, people may drive in a distributed probability pattern instead of in a uniform line (e.g., different speeds and different curvatures at the same corner). As such, there can be more than one “correct” control trajectory for an autonomous vehicle to perform in the same turn. Safety, comfort, speeds, and other criteria may lead to different preferences and judgment as to how well the controlled trajectory has been computed.

Abstract: According to some embodiments, systems, methods and media for dynamically generating scenario parameters for an autonomous driving vehicles (ADV) are described. In one embodiment, when an ADV enters a driving scenario, the ADV can invoke a map-based scenario checker to determine the type of scenario, and invokes a corresponding neural network model to generate a set of parameters for the scenario based on real-time environmental conditions (e.g., traffics) and vehicle status information (e.g., speed). The set of scenario parameters can be a set of extra constraints for configuring the ADV to drive in a driving mode corresponding to the scenario.

Abstract: Systems and methods for generating labelled audio data and onboard validation of the labelled audio data utilizing an autonomous driving vehicle (ADV) while the ADV is operating within a driving environment are disclosed. The method includes recording a sound emitted by an object within the driving environment of the ADV, and converting the recorded sound into audio samples. The method further includes labelling the audio samples, and refining the labelled audio samples to produce refined labelled audio data. The refined labelled audio data is utilized to subsequently train a machine learning algorithm to recognize a sound source during autonomous driving of the ADV. The method further includes generating a performance profile of the refined labelled audio data based at least on the audio samples, a position of the object, and a relative direction of the object. The position of the object and the relative direction of the object are determined by a perception system of the ADV.

Abstract: Systems and methods are disclosed for optimizing values of a set of tunable parameters of an autonomous driving vehicle (ADV). The controllers can be a linear quadratic regular, a “bicycle model,” a model-reference adaptive controller (MRAC) that reduces actuation latency in control subsystems such as steering, braking, and throttle, or other controller (“controllers”). An optimizer selects a set tunable parameters for the controllers. A task distribution system pairs each set of parameters with each of a plurality of simulated driving scenarios, and dispatches a task to the simulator to perform the simulation with the set of parameters. Each simulation is scored. A weighted score is generated from the simulation. The optimizer uses the weighted score as a target objective for a next iteration of the optimizer, for a fixed number of iterations. A physical real-world ADV is navigated using the optimized set of parameters for the controllers in the ADV.