Gm Global Technolgoy Operations, Inc.

Carnegie Mellon University

John M Dolan

Upali Priyantha Mudalige

Praveen Palanisamy

Katharina Muelling

Bakhtiar Brian Litkouhi

Yilun Chen

Jin-Woo Lee

Michael A Garcia

John W Lambert

Tianyu Gu

Chiyu Dong

Junqing Wei

Zhiqian Qiao

Robert Mueller

Systems and methods are provided for end-to-end learning of commands for controlling an autonomous vehicle. A pre-processor pre-processes image data acquired by sensors at a current time step (CTS) to generate pre-processed image data that is concatenated with additional input(s) (e.g., a segmentation map and/or optical flow map) to generate a dynamic scene output. A convolutional neural network (CNN) processes the dynamic scene output to generate a feature map that includes extracted spatial features that are concatenated with vehicle kinematics to generate a spatial context feature vector. An LSTM network processes, during the (CTS), the spatial context feature vector at the (CTS) and one or more previous LSTM outputs at corresponding previous time steps to generate an encoded temporal context vector at the (CTS). The fully connected layer processes the encoded temporal context vector to learn control command(s) (e.g., steering angle, acceleration rate and/or a brake rate control commands).

Method and system for end-to-end learning of control commands for autonomous vehicle

Systems and methods are provided that employ spatial and temporal attention-based deep reinforcement learning of hierarchical lane-change policies for controlling an autonomous vehicle. An actor-critic network architecture includes an actor network that process image data received from an environment to learn the lane-change policies as a set of hierarchical actions, and a critic network that evaluates the lane-change policies to calculate loss and gradients to predict an action-value function (Q) that is used to drive learning and update parameters of the lane-change policies. The actor-critic network architecture implements a spatial attention module to select relevant regions in the image data that are of importance, and a temporal attention module to learn temporal attention weights to be applied to past frames of image data to indicate relative importance in deciding which lane-change policy to select.

Spatial and temporal attention-based deep reinforcement learning of hierarchical lane-change policies for controlling an autonomous vehicle

The present application generally relates to a method and apparatus for generating an action policy for controlling an autonomous vehicle. In particular, the system performs a deep learning algorithm in order to determine the action policy and an automatically generated curriculum system to determine a number of increasingly difficult tasks in order to refine the action policy.

Method and apparatus for automatically generated curriculum sequence based reinforcement learning for autonomous vehicles

Systems and method are provided for controlling a first vehicle. In one embodiment, a method includes: determining, by a processor, that a second vehicle is going to merge into a current lane of the first vehicle at a merge point; and in response to the determining: obtaining, by the processor, a velocity history of the second vehicle, determining, by the processor, a first time value associated with the first vehicle and the merge point; determining, by the processor, a second time value associated with the second vehicle and the merge point; determining, by the processor, an intention of the second vehicle based on the velocity history, the first time value, and the second time value; and selectively generating, by the processor a command to at least one of control the first vehicle and provide notification based on the intention.

Systems and methods for autonomous driving merging management

A method for automated lane centering and/or lane changing purposes for a vehicle traveling on a roadway that employs roadway points from a map database to determine a reference vehicle path and sensors on the vehicle for detecting static and moving objects to adjust the reference path. The method includes reducing the curvature of the reference path to generate a reduced curvature reference path that reduces the turning requirements of the vehicle and setting the speed of the vehicle from posted roadway speeds from the map database. The method also includes providing multiple candidate vehicle paths and vehicle speeds to avoid the static and moving objects in front of the vehicle.

Unified motion planning algorithm for autonomous driving vehicle in obstacle avoidance maneuver

An autonomous driving merge management system includes an autonomous driving control device and an intention decision management system. The management system includes a candidate strategy subsystem generating a plurality of candidate driving strategies, a merging vehicle behavior recognition subsystem predicting a merging intention of a merging vehicle; an intention-based interactive prediction subsystem predicting future merging scenarios between the host vehicle and merging vehicle as a function of inputs by the merging vehicle behavior recognition subsystem and inputs by the candidate strategy subsystem, and a cost function-based evaluation subsystem determining a cost for each future merging scenario generated by the intention-based interactive prediction subsystem. A processor selects a merge strategy of the host vehicle based on intention-based prediction results and cost function-based evaluation results. The autonomous driving control device applies the merge strategy to the host vehicle for allowing the merging vehicle to cooperatively merge with the host vehicle.

Growing community of inventors

Pittsburgh, PA, United States of America

John M Dolan