Google Inc.

X Development LLC

Gdm Holding LLC

Matthew Bennice

Paul Bechard

Yunfei Bai

Brent Austin

Joséphine Simon

Sergey Levine

Seyed Mohammad Khansari Zadeh

Eric Jang

Daniel Alex Lam

Chelsea Finn

Daniel Kappler

Alexander Irpan

Chuyuan Fu

Wenlong Lu

Tigran Gasparian

Nicolas Sievers

Andreas Christiansen

Jiayi Lin

Active utilization of a robotic simulator in control of one or more real world robots. A simulated environment of the robotic simulator can be configured to reflect a real world environment in which a real robot is currently disposed, or will be disposed. The robotic simulator can then be used to determine a sequence of robotic actions for use by the real world robot(s) in performing at least part of a robotic task. The sequence of robotic actions can be applied, to a simulated robot of the robotic simulator, to generate a sequence of anticipated simulated state data instances. The real robot can be controlled to implement the sequence of robotic actions. The implementation of one or more of the robotic actions can be contingent on a real state data instance having at least a threshold degree of similarity to a corresponding one of the anticipated simulated state data instances.

Simulation driven robotic control of real robot(s)

Implementations are provided for training an imitation robot control stack model to approximate the behavior of a robot control stack. In various implementations one or more training examples for training an imitation robot control stack model may be received. The training examples may comprise: input data to a robot control stack, wherein the input data comprises (i) a high-level command for controlling a robot, and (ii) current state data of the robot and an environment, and output data generated based on processing, by the robot control stack, the input data to the robot control stack, wherein the output data comprises a low-level command for controlling the robot according to the high-level command. The imitation robot control stack model may be trained based on the one or more training examples. Operation of the robot may be simulated based in part on controlling the operation of the robot using the trained imitation robot control stack model.

Imitation robot control stack models

Implementations described herein relate to training and refining robotic control policies using imitation learning techniques. A robotic control policy can be initially trained based on human demonstrations of various robotic tasks. Further, the robotic control policy can be refined based on human interventions while a robot is performing a robotic task. In some implementations, the robotic control policy may determine whether the robot will fail in performance of the robotic task, and prompt a human to intervene in performance of the robotic task. In additional or alternative implementations, a representation of the sequence of actions can be visually rendered for presentation to the human can proactively intervene in performance of the robotic task.

System(s) and method(s) of using imitation learning in training and refining robotic control policies

Implementations are provided for increasing realism of robot simulation by injecting noise into various aspects of the robot simulation. In various implementations, a three-dimensional (3D) environment may be simulated and may include a simulated robot controlled by an external robot controller. Joint command(s) issued by the robot controller and/or simulated sensor data passed to the robot controller may be intercepted. Noise may be injected into the joint command(s) to generate noisy commands. Additionally or alternatively, noise may be injected into the simulated sensor data to generate noisy sensor data. Joint(s) of the simulated robot may be operated in the simulated 3D environment based on the one or more noisy commands. Additionally or alternatively, the noisy sensor data may be provided to the robot controller to cause the robot controller to generate joint commands to control the simulated robot in the simulated 3D environment.

Injecting noise into robot simulation

Implementations are provided for operably coupling multiple robot controllers to a single virtual environment, e.g., to generate training examples for training machine learning model(s). In various implementations, a virtual environment may be simulated that includes an interactive object and a plurality of robot avatars that are controlled independently and contemporaneously by a corresponding plurality of robot controllers that are external from the virtual environment. Sensor data generated from a perspective of each robot avatar of the plurality of robot avatars may be provided to a corresponding robot controller. Joint commands that cause actuation of one or more joints of each robot avatar may be received from the corresponding robot controller. Joint(s) of each robot avatar may be actuated pursuant to corresponding joint commands. The actuating may cause two or more of the robot avatars to act upon the interactive object in the virtual environment.

Simulating multiple robots in virtual environments

Implementations are provided for generating a plurality of simulated training instances based on a recorded user-directed robot control episode, and training one or more robot control policies based on such training instances. In various implementations, a three-dimensional environment may be simulated and may include a robot controlled by an external robot controller. A user may operate the robot controller to control the robot in the simulated 3D environment to perform one or more robotic tasks. The user-directed robot control episode, including responses of the external robot controller and the simulated robot to user commands and/or the virtual environment, can be captured. Features of the captured user-directed robot control episode can be altered in order to generate a plurality of training instances. One or more robot control policies can then be trained based on the plurality of training instances.

Re-simulation of recorded episodes

Implementations are described herein for single iteration, multiple permutation robot simulation. In various implementations, one or more poses of a simulated object may be determined across one or more virtual environments. A plurality of simulated robots may be operated across the one or more virtual environments. For each simulated robot of the plurality of simulated robots, a camera transformation may be determined based on respective poses of the simulated robot and simulated object in the particular virtual environment. The camera transformation may be applied to the simulated object in the particular virtual environment of the one or more virtual environments in which the simulated robot operates. Based on the camera transformation, simulated vision data may be rendered that depicts the simulated object from a perspective of the simulated robot. Each of the plurality of simulated robots may be operated based on corresponding simulated vision data.

Single iteration, multiple permutation robot simulation

Growing community of inventors

San Jose, CA, United States of America

Matthew Bennice