Google Inc.

X Development LLC

Alexander Herzog

Mrinal Kalakrishnan

Yunfei Bai

Benjamin Holson

Adrian Li

Sergey Levine

Seyed Mohammad Khansari Zadeh

Eric Jang

Paul Wohlhart

Julian Ibarz

Chelsea Finn

Daniel Kappler

Chuyuan Fu

Wenlong Lu

Daniel Ho

Zhuo Xu

C Karen Liu

Allan Zhou

Dmitry Kalashnikov

Wenhao Yu

Implementations disclosed herein relate to utilizing at least one existing manually engineered policy, for a robotic task, in training an RL policy model that can be used to at least selectively replace a portion of the engineered policy. The RL policy model can be trained for replacing a portion of a robotic task and can be trained based on data from episodes of attempting performance of the robotic task, including episodes in which the portion is performed based on the engineered policy and/or other portion(s) are performed based on the engineered policy. Once trained, the RL policy model can be used, at least selectively and in lieu of utilization of the engineered policy, to perform the portion of robotic task, while other portion(s) of the robotic task are performed utilizing the engineered policy and/or other similarly trained (but distinct) RL policy model(s).

Training a policy model for a robotic task, using reinforcement learning and utilizing data that is based on episodes, of the robotic task, guided by an engineered policy

Techniques are disclosed that enable training a meta-learning model, for use in causing a robot to perform a task, using imitation learning as well as reinforcement learning. Some implementations relate to training the meta-learning model using imitation learning based on one or more human guided demonstrations of the task. Additional or alternative implementations relate to training the meta-learning model using reinforcement learning based on trials of the robot attempting to perform the task. Further implementations relate to using the trained meta-learning model to few shot (or one shot) learn a new task based on a human guided demonstration of the new task.

Efficient adaption of robot control policy for new task using meta-learning based on meta-imitation learning and meta-reinforcement learning

Utilization of past dynamics sample(s), that reflect past contact physics information, in training and/or utilizing a neural network model. The neural network model represents a learned value function (e.g., a Q-value function) and that, when trained, can be used in selecting a sequence of robotic actions to implement in robotic manipulation (e.g., pushing) of an object by a robot. In various implementations, a past dynamics sample for an episode of robotic manipulation can include at least two past images from the episode, as well as one or more past force sensor readings that temporally correspond to the past images from the episode.

Utilizing past contact physics in robotic manipulation (e.g., pushing) of an object

Asynchronous robotic control utilizing a trained critic network. During performance of a robotic task based on a sequence of robotic actions determined utilizing the critic network, a corresponding next robotic action of the sequence is determined while a corresponding previous robotic action of the sequence is still being implemented. Optionally, the next robotic action can be fully determined and/or can begin to be implemented before implementation of the previous robotic action is completed. In determining the next robotic action, most recently selected robotic action data is processed using the critic network, where such data conveys information about the previous robotic action that is still being implemented. Some implementations additionally or alternatively relate to determining when to implement a robotic action that is determined in an asynchronous manner.

Asynchronous robotic control using most recently selected robotic action data

Utilizing at least one existing policy (e.g. a manually engineered policy) for a robotic task, in generating reinforcement learning (RL) data that can be used in training an RL policy for an instance of RL of the robotic task. The existing policy can be one that, standing alone, will not generate data that is compatible with the instance of RL for the robotic task. In contrast, the generated RL data is compatible with RL for the robotic task at least by virtue of it including state data that is in a state space of the RL for the robotic task, and including actions that are in the action space of the RL for the robotic task. The generated RL data can be used in at least some of the initial training for the RL policy using reinforcement learning.

Generating reinforcement learning data that is compatible with reinforcement learning for a robotic task

Growing community of inventors

San Jose, CA, United States of America

Alexander Herzog