Apple Inc.

Martin Levihn

Nicolas F Meuleau

Luke Burke Johnson

Joshua D Redding

Sebastian Brechtel

Pekka Tapani Raiko

Sandip Aine

Jamie R Lesser

Jürgen Wiest

Anayo Kenechukwu Akametalu

Ujjwal Das Gupta

Trevor P Barron

Luke B Johnson

Implementations described and claimed herein provide systems and methods for controlling an autonomous vehicle. In one implementation, the autonomous vehicle is navigated towards a flow of traffic with a first gap between first and second vehicles and a second gap following the second vehicle. A motion plan for directing the autonomous vehicle into the flow of traffic at an interaction zone is generated based on whether an ability of the autonomous vehicle to enter the interaction zone at the second gap exceeds a confidence threshold. The autonomous vehicle is autonomously navigated into the flow of traffic at the first gap when the confidence threshold is exceeded. The motion plan forgoes navigation of the autonomous vehicle into the flow of traffic at the first and second gaps when the ability of the autonomous vehicle to enter the interaction zone at the second gap does not exceed the confidence threshold.

Systems and methods for hedging for different gaps in an interaction zone

A graph representation of a tactical map representing a plurality of static components of an environment of a vehicle is generated. Nodes of the graph represent static components, and edges represent relationships between multiple static components. Different edge types are used to indicate respective relationship semantics among the static components. Individual nodes are represented as having the same number and types of edges in the graph. Using the graph as input to a neural network based model, a set of results is obtained. A motion control directive based at least in part on the results is transmitted to a motion-control subsystem of the vehicle.

Processing graph representations of tactical maps using neural networks

A behavior planner for a vehicle generates a plurality of conditional action sequences of the vehicle using a tree search algorithm and heuristics obtained from one or more machine learning models. Each sequence corresponds to a sequence of anticipated states of the vehicle. At least some of the action sequences are provided to a motion selector of the vehicle. The motion selector generates motion-control directives based on the received conditional action sequences and on data received from one or more sensors of the vehicle, and transmits the directives to control subsystems of the vehicle.

Decision making for autonomous vehicle motion control

A set of actions corresponding to a particular state of the environment of a vehicle is identified. A respective encoding is generated for different actions of the set, using elements such as distinct colors to distinguish attributes such as target lane segments. Using the encodings as inputs to respective instances of a machine learning model, respective value metrics are estimated for each of the actions. One or more motion-control directives to implement a particular action selected using the value metrics are transmitted to motion-control subsystems of the vehicle.

Evaluating varying-sized action spaces using reinforcement learning

Growing community of inventors

San Jose, CA, United States of America

Martin Levihn