Baidu USA LLC

Baidu.com Times Technology (beijing) Co. Ltd.

Runxin He

Jiangtao Hu

Qi Luo

Jinghao Miao

Yu Wang

Shu Jiang

Jinyun Zhou

Jiaxuan Xu

Shiyu Song

Li Yu

Pengfei Yuan

Yong Xiao

Jingao Wang

Wendong Ding

Weiman Lin

Xiangquan Xiao

A method of navigating an autonomous driving vehicle (ADV) includes determining a target function for an open space model based on one or more obstacles and map information within a proximity of the ADV, then iteratively performing first and second quadratic programming (QP) optimizations on the target function. Then, generating a second trajectory based on results of the first and second QP optimizations to control the ADV autonomously using the second trajectory. The first QP optimization is based on fixing a first set of variables of the target function. The second QP optimization is based on maximizing a sum of the distances from the ADV to each of the obstacles over a plurality of points of the first trajectory, and minimizing a difference between a target end-state of the ADV and a determined final state of the ADV using the first trajectory.

End dynamics and constraints relaxation algorithm on optimizing an open space trajectory

In one embodiment, a computer-implemented method of operating an autonomous driving vehicle (ADV) includes perceiving a driving environment surrounding the ADV based on sensor data obtained from one or more sensors mounted on the ADV, determining a driving scenario, in response to a driving decision based on the driving environment, applying a predetermined machine-learning model to data representing the driving environment and the driving scenario to generate a set of one or more driving parameters, and planning a trajectory to navigate the ADV using the set of the driving parameters according to the driving scenario through the driving environment.

Neural network approach for parameter learning to speed up planning for complex driving scenarios

In one embodiment, a method of training dynamic models for autonomous driving vehicles includes the operations of receiving a first set of training data from a training data source, the first set of training data representing driving statistics for a first set of features; training a dynamic model based on the first set of training data for the first set of features; determining a second set of features as a subset of the first set of features based on evaluating the dynamic model, each of the second set of features representing a feature whose performance score is below a predetermined threshold. The method further includes the following operations for each of the second set of features: retrieving a second set of training data associated with the corresponding feature of the second set of features, and retraining the dynamic model using the second set of training data.

Automated training data extraction method for dynamic models for autonomous driving vehicles

In one embodiment, a system and method for partitioning a region for point cloud registration of LIDAR poses of an autonomous driving vehicle (ADV) using a regional iterative closest point (ICP) algorithm is disclosed. The method determines the frame pair size of one or more pairs of related LIDAR poses of a region of an HD map to be constructed. If the frame pair size is greater than a threshold, the region is further divided into multiple clusters. The method may perform the ICP algorithm for each cluster. Inside a cluster, the ICP algorithm focuses on a partial subset of the decision variables and assumes the rest of the decision variables are fixed. To construct the HD map, the method may determine if the results of the ICP algorithms from the clusters converge. If the solutions converge, a solution to the point cloud registration for the region is found.

Coordinate gradient method for point cloud registration for autonomous vehicles

In one embodiment, a system is disclosed for registration of point clouds for autonomous driving vehicles (ADV). The system receives a number of point clouds and corresponding poses from ADVs equipped with LIDAR sensors capturing point clouds of a navigable area to be mapped, where the point clouds correspond to a first coordinate system. The system partitions the point clouds and the corresponding poses into one or more loop partitions based on navigable loop information captured by the point clouds. For each of the loop partitions, the system applies an optimization model to point clouds corresponding to the loop partition to register the point clouds. They system merges the one or more loop partitions together using a pose graph algorithm, where the merged partitions of point clouds are utilized to perceive a driving environment surrounding the ADV.

Point clouds registration system for autonomous vehicles

In one embodiment, a system and method for point cloud registration of LIDAR poses of an autonomous driving vehicle (ADV) is disclosed. The method selects poses of the point clouds that possess higher confidence level during the data capture phase as fixed anchor poses. The fixed anchor points are used to estimate and optimize the poses of non-anchor poses during point cloud registration. The method may partition the points clouds into blocks to perform the ICP algorithm for each block in parallel by minimizing the cost function of the bundle adjustment equation updated with a regularity term. The regularity term may measure the difference between current estimates of the poses and previous or the initial estimates. The method may also minimize the bundle adjustment equation updated with a regularity term when solving the pose graph problem to merge the optimized poses from the blocks to make connections between the blocks.

Updated point cloud registration pipeline based on ADMM algorithm for autonomous vehicles

In one embodiment, a system generates a plurality of driving scenarios to train a reinforcement learning (RL) agent and replays each of the driving scenarios to train the RL agent by: applying a RL algorithm to an initial state of a driving scenario to determine a number of control actions from a number of discretized control/action options for the ADV to advance to a number of trajectory states which are based on a number of discretized trajectory state options, determining a reward prediction by the RL algorithm for each of the controls/actions, determining a judgment score for the trajectory states, and updating the RL agent based on the judgment score.

Offline agent using reinforcement learning to speedup trajectory planning for autonomous vehicles

In one embodiment, a computer-implemented method of autonomously parking an autonomous driving vehicle, includes generating environment descriptor data describing a driving environment surrounding the autonomous driving vehicle (ADV), including identifying a parking space and one or more obstacles within a predetermined proximity of the ADV, generating a parking trajectory of the ADV based on the environment descriptor data to autonomously park the ADV into the parking space, including optimizing the parking trajectory in view of the one or more obstacles, segmenting the parking trajectory into one or more trajectory segments based on a vehicle state of the ADV, and controlling the ADV according to the one or more trajectory segments of the parking trajectory to autonomously park the ADV into the parking space without collision with the one or more obstacles.

Segmenting a parking trajectory to control an autonomous driving vehicle to park

In one embodiment, a system identifies a road to be navigated by an ADV, the road being captured by one or more point clouds from one or more LIDAR sensors. The system extracts road marking information of the identified road from the point clouds, the road marking information describing one or more road markings of the identified road. The system partitions the road into one or more road partitions based on the road markings. The system generates a point cloud map based on the road partitions, where the point cloud map is utilized to perceive a driving environment surrounding the ADV.

Map partition system for autonomous vehicles

In one embodiment, a system uses an actor-critic reinforcement learning model to generate a trajectory for an autonomous driving vehicle (ADV) in an open space. The system perceives an environment surrounding an ADV. The system applies a RL algorithm to an initial state of a planning trajectory based on the perceived environment to determine a plurality of controls for the ADV to advance to a plurality of trajectory states based on map and vehicle control information for the ADV. The system determines a reward prediction by the RL algorithm for each of the plurality of controls in view of a target destination state. The system generates a first trajectory from the trajectory states by maximizing the reward predictions to control the ADV autonomously according to the first trajectory.

Online agent using reinforcement learning to plan an open space trajectory for autonomous vehicles

In one embodiment, a system receives a stream of frames of point clouds from one or more LIDAR sensors of an ADV and corresponding poses in real-time. The system extracts segment information for each frame of the stream based on geometric or spatial attributes of points in the frame, where the segment information includes one or more segments of at least a first frame corresponding to a first pose. The system registers the stream of frames based on the segment information. The system generates a first point cloud map for the stream of frames based on the frame registration.

Real-time map generation system for autonomous vehicles

An autonomous driving vehicle (ADV) receives instructions for a human test driver to drive the ADV in manual mode and to collect a specified amount of driving data for one or more specified driving categories. As the user drivers the ADV in manual mode, driving data corresponding to the one or more driving categories is logged. A user interface of the ADV displays the one or more driving categories that the human driver is instructed collect data upon, and a progress indicator for each of these categories as the human driving progresses. The driving data is uploaded to a server for machine learning. If the server machine learning achieves a threshold grading amount of the uploaded data to variables of a dynamic self-driving model, then the server generates an ADV self-driving model, and distributes the model to one or more ADVs that are navigated in the self-driving mode.

Data collection automation system

In one embodiment, an open space model is generated for a system to plan trajectories for an ADV in an open space. The system perceives an environment surrounding an ADV including one or more obstacles. The system determines a target function for the open space model based on constraints for the one or more obstacles and map information. The system iteratively, performs a first quadratic programming (QP) optimization on the target function based on a first trajectory while fixing a first set of variables, and performs a second QP optimization on the target function based on a result of the first QP optimization while fixing a second set of variables. The system generates a second trajectory based on results of the first and the second QP optimizations to control the ADV autonomously according to the second trajectory.

Relaxation optimization model to plan an open space trajectory for autonomous vehicles

Generating control effort to control an autonomous driving vehicle (ADV) includes determining a direction (forward or reverse) in which the ADV is driving and selecting a driving model and a predictive model based upon the direction. In a forward direction, the driving model is a dynamic model, such as a 'bicycle model,' and the predictive model is a look-ahead model. In a reverse direction, the driving model is a hybrid dynamic and kinematic model and the predictive model is a look-back model. Current and predicted lateral error and heading error are determined using the driving model and predictive model, respectively. A linear quadratic regulator (LQR) uses the current and predicted lateral error and heading errors to determine a first control effort An augmented control logic determines a second, additional, control effort, to determine a final control effort that is output to a control module to drive the ADV.

Low-speed, backward driving vehicle controller design

In one embodiment, an autonomous driving system of an autonomous driving vehicle perceives a driving environment surrounding the autonomous driving vehicle traveling along a path, including perceiving an obstacle in the driving environment. The system detects a vertical acceleration of the autonomous driving vehicle based on sensor data obtained from a sensor on the autonomous driving vehicle. The system further calibrates the perceived obstacle based on the vertical acceleration of the autonomous driving vehicle. The system then controls the autonomous driving vehicle to navigate through the driving environment in view of the calibrated perceived obstacle.

Growing community of inventors

Sunnyvale, CA, United States of America

Runxin He