Google Inc.

University of California

X Development LLC

Toyota Research Institute, Inc.

Sergey Levine

Eric Jang

Peter Pastor Sampedro

Shixiang Gu

Chelsea Finn

Alexander Irpan

Mrinal Kalakrishnan

Yunfei Bai

Sudheendra Vijayanarasimhan

Timothy Paul Lillicrap

Seyed Mohammad Khansari Zadeh

Honglak Lee

Julian Ibarz

Daniel Kappler

Alex Krizhevsky

Ethan Holly

Daniel Alex Lam

Matthew Bennice

Paul Wohlhart

Konstantinos Bousmalis

Yevgen Chebotar

Adrian Ling Hin Li

Brent Austin

Ali Hamid Yahya Valdovinos

Nicolas Sievers

Peng Xu

Adrien David Gaidon

Quoc V Le

Vincent Olivier Vanhoucke

Kurt Konolige

Ilya Sutskever

Alexander Toshev

Christopher Kenneth Harris

Rahul Sukthankar

Kanury Kanishka Rao

Ian Goodfellow

Aleksandra Faust

Blake Warren Wulfe

Jie Tan

Saurabh Gupta

Alexander Herzog

James Christopher Davidson

Sehoon Ha

Logan Michael Ellis

Jitendra Malik

Daiyi Peng

Chris Harris

Coline Manon Devin

Kanishka Rao

Allan Zhou

Zhenyu Tan

Yingjie Miao

John Dalton Co-Reyes

Jean Mercat

Rowan Thomas Mcallister

Frederik David Ebert

Akhil Amar Padmanabha

Stephen Tian

Roberto Calandra

Fereshteh Sadeghi

Matthew Laurance Kelcey

Implementations disclosed herein relate to mitigating the reality gap through training a simulation-to-real machine learning model ('Sim2Real' model) using a vision-based robot task machine learning model. The vision-based robot task machine learning model can be, for example, a reinforcement learning (“RL”) neural network model (RL-network), such as an RL-network that represents a Q-function.

Mitigating reality gap through training a simulation-to-real model using a vision-based robot task model

Training and/or utilizing a hierarchical reinforcement learning (HRL) model for robotic control. The HRL model can include at least a higher-level policy model and a lower-level policy model. Some implementations relate to technique(s) that enable more efficient off-policy training to be utilized in training of the higher-level policy model and/or the lower-level policy model. Some of those implementations utilize off-policy correction, which re-labels higher-level actions of experience data, generated in the past utilizing a previously trained version of the HRL model, with modified higher-level actions. The modified higher-level actions are then utilized to off-policy train the higher-level policy model. This can enable effective off-policy training despite the lower-level policy model being a different version at training time (relative to the version when the experience data was collected).

Data-efficient hierarchical reinforcement learning

Methods, computer systems, and apparatus, including computer programs encoded on computer storage media, for generating and searching reinforcement learning algorithms. In some implementations, a computer-implemented system generates a sequence of candidate reinforcement learning algorithms. Each candidate reinforcement learning algorithm in the sequence is configured to receive an input environment state characterizing a state of an environment and to generate an output that specifies an action to be performed by an agent interacting with the environment. For each candidate reinforcement learning algorithm in the sequence, the system performs a performance evaluation for a set of a plurality of training environments. For each training environment, the system adjusts a set of environment-specific parameters of the candidate reinforcement learning algorithm by performing training of the candidate reinforcement learning algorithm to control a corresponding agent in the training environment. The system generates an environment-specific performance metric for the candidate reinforcement learning algorithm that measures a performance of the candidate reinforcement learning algorithm in controlling the corresponding agent in the training environment as a result of the training. After performing training in the set of training environments, the system generates a summary performance metric for the candidate reinforcement learning algorithm by combining the environment-specific performance metrics generated for the set of training environments. After evaluating each of the candidate reinforcement learning algorithms in the sequence, the system selects one or more output reinforcement learning algorithms from the sequence based on the summary performance metrics of the candidate reinforcement learning algorithms.

Reinforcement learning algorithm search

Implementations utilize deep reinforcement learning to train a policy neural network that parameterizes a policy for determining a robotic action based on a current state. Some of those implementations collect experience data from multiple robots that operate simultaneously. Each robot generates instances of experience data during iterative performance of episodes that are each explorations of performing a task, and that are each guided based on the policy network and the current policy parameters for the policy network during the episode. The collected experience data is generated during the episodes and is used to train the policy network by iteratively updating policy parameters of the policy network based on a batch of collected experience data. Further, prior to performance of each of a plurality of episodes performed by the robots, the current updated policy parameters can be provided (or retrieved) for utilization in performance of the episode.

Deep reinforcement learning for robotic manipulation

Optical tactile sensors are provided that include a scaffolding structure, a transparent elastomer material covering at least an end portion of the scaffolding structure, and one or multiple cameras situated on the end portion of the scaffolding structure and embedded within the transparent elastomer material, wherein the one or multiple cameras are situated so as to provide an extended, e.g., up to 360°, field of view about the end portion of the scaffolding structure.

Multi-directional high-resolution optical tactile sensors

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

A method of generating an output trajectory of an ego vehicle includes recording trajectory data of the ego vehicle and pedestrian agents from a scene of a training environment of the ego vehicle. The method includes identifying at least one pedestrian agent from the pedestrian agents within the scene of the training environment of the ego vehicle causing a prediction-discrepancy by the ego vehicle greater than the pedestrian agents within the scene. The method includes updating parameters of a motion prediction model of the ego vehicle based on a magnitude of the prediction-discrepancy caused by the at least one pedestrian agent on the ego vehicle to form a trained, control-aware prediction objective model. The method includes selecting a vehicle control action of the ego vehicle in response to a predicted motion from the trained, control-aware prediction objective model regarding detected pedestrian agents within a traffic environment of the ego vehicle.

Planning-aware prediction for control-aware autonomous driving modules

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

Techniques are disclosed that enable training a plurality of policy networks, each policy network corresponding to a disparate robotic training task, using a mobile robot in a real world workspace. Various implementations include selecting a training task based on comparing a pose of the mobile robot to at least one parameter of a real world training workspace. For example, the training task can be selected based on the position of a landmark, within the workspace, relative to the pose. For instance, the training task can be selected such that the selected training task moves the mobile robot towards the landmark.

System and methods for training robot policies in the real world

A method includes: receiving data identifying, for each of one or more objects, a respective target location to which a robotic agent interacting with a real-world environment should move the object; causing the robotic agent to move the one or more objects to the one or more target locations by repeatedly performing the following: receiving a current image of a current state of the real-world environment; determining, from the current image, a next sequence of actions to be performed by the robotic agent using a next image prediction neural network that predicts future images based on a current action and an action to be performed by the robotic agent; and directing the robotic agent to perform the next sequence of actions.

Control policies for robotic agents

Deep machine learning methods and apparatus related to semantic robotic grasping are provided. Some implementations relate to training a training a grasp neural network, a semantic neural network, and a joint neural network of a semantic grasping model. In some of those implementations, the joint network is a deep neural network and can be trained based on both: grasp losses generated based on grasp predictions generated over a grasp neural network, and semantic losses generated based on semantic predictions generated over the semantic neural network. Some implementations are directed to utilization of the trained semantic grasping model to servo, or control, a grasping end effector of a robot to achieve a successful grasp of an object having desired semantic feature(s).

Machine learning methods and apparatus for semantic robotic grasping

Training and/or using a recurrent neural network model for visual servoing of an end effector of a robot. In visual servoing, the model can be utilized to generate, at each of a plurality of time steps, an action prediction that represents a prediction of how the end effector should be moved to cause the end effector to move toward a target object. The model can be viewpoint invariant in that it can be utilized across a variety of robots having vision components at a variety of viewpoints and/or can be utilized for a single robot even when a viewpoint, of a vision component of the robot, is drastically altered. Moreover, the model can be trained based on a large quantity of simulated data that is based on simulator(s) performing simulated episode(s) in view of the model. One or more portions of the model can be further trained based on a relatively smaller quantity of real training data.

Viewpoint invariant visual servoing of robot end effector using recurrent neural network

Growing community of inventors

Berkeley, CA, United States of America

Sergey Levine