X Development LLC

Google Inc.

Mrinal Kalakrishnan

Adrian Li

Kurt Konolige

Peter Pastor Sampedro

Paul Wohlhart

Vincent Louis Dureau

Adrian Ling Hin Li

Sergey Levine

Yunfei Bai

Seyed Mohammad Khansari Zadeh

Ali Yahya Valdovinos

Ethan Rublee

Nicolas Hudson

Konstantinos Bousmalis

Alexander Herzog

Benjamin Holson

Yevgen Chebotar

Kuan Fang

Daniel Kappler

Ali Hamid Yahya Valdovinos

Stephen James

Vincent Olivier Vanhoucke

Stefan Hinterstoisser

Jeffrey Bingham

Vikas Sindhwani

Eric Jang

Julian Ibarz

Chelsea Finn

Alexander Irpan

Jianlan Luo

Torsten Kroeger

Karol Hausman

Matthew Laurance Kelcey

Allan Zhou

Mengyuan Yan

Ali Yahya Valdovinoa

Cristian Bodnar

Sergey Vladimir Levine

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium for predicting object pose. In one aspect, a method includes receiving an image of an object having one or more feature points; providing the image as an input to a neural network subsystem trained to receive images of objects and to generate an output including a heat map for each feature point; applying a differentiable transformation on each heat map to generate respective one or more feature coordinates for each feature point; providing the feature coordinates for each feature point as input to an object pose solver configured to compute a predicted object pose for the object, wherein the predicted object pose for the object specifies a position and an orientation of an object; and receiving, at the output of the object pose solver, a predicted object pose for the object in the image.

Object pose neural network system

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

Implementations are directed to training a machine learning model that, once trained, is used in performance of robotic grasping and/or other manipulation task(s) by a robot. The model can be trained using simulated training examples that are based on simulated data that is based on simulated robot(s) attempting simulated manipulations of various simulated objects. Portion(s) of the model can also be trained based on real training examples that are based on data from real-world physical robots attempting manipulations of various objects. The simulated training examples can be utilized to train the model to predict an output that can be utilized in a particular task—and the real training examples used to adapt at least a portion of the model to the real-world domain can be tailored to a distinct task. In some implementations, domain-adversarial similarity losses are determined during training, and utilized to regularize at least portion(s) of the model.

Machine learning methods and apparatus for robotic manipulation and that utilize multi-task domain adaptation

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

Methods and apparatus related to receiving a request that includes robot instructions and/or environmental parameters, operating each of a plurality of robots based on the robot instructions and/or in an environment configured based on the environmental parameters, and storing data generated by the robots during the operating. In some implementations, at least part of the stored data that is generated by the robots is provided in response to the request and/or additional data that is generated based on the stored data is provided in response to the request.

Operating multiple testing robots based on robot instructions and/or environmental parameters received in a request

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training a generator neural network to adapt input images.

Domain adaptation using simulation to simulation transfer

Methods, apparatus, and computer readable media related to combining and/or training one or more neural network modules based on version identifier(s) assigned to the neural network module(s). Some implementations are directed to using version identifiers of neural network modules in determining whether and/or how to combine multiple neural network modules to generate a combined neural network model for use by a robot and/or other apparatus. Some implementations are additionally or alternatively directed to assigning a version identifier to an endpoint of a neural network module based on one or more other neural network modules to which the neural network module is joined during training of the neural network module.

Neural network modules

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

Training and/or use of a machine learning model for placement of an object secured by an end effector of a robot. A trained machine learning model can be used to process: (1) a current image, captured by a vision component of a robot, that captures an end effector securing an object; (2) a candidate end effector action that defines a candidate motion of the end effector; and (3) a target placement input that indicates a target placement location for the object. Based on the processing, a prediction can be generated that indicates likelihood of successful placement of the object in the target placement location with application of the motion defined by the candidate end effector action. At many iterations, the candidate end effector action with the highest probability is selected and control commands provided to cause the end effector to move in conformance with the corresponding end effector action. When at least one release criteria is satisfied, control commands can be provided to cause the end effector to release the object, thereby leading to the object being placed in the target placement location.

Machine learning methods and apparatus for automated robotic placement of secured object in appropriate location

Generating and utilizing action image(s) that represent a candidate pose (e.g., a candidate end effector pose), in determining whether to utilize the candidate pose in performance of a robotic task. The action image(s) and corresponding current image(s) can be processed, using a trained critic network, to generate a value that indicates a probability of success of the robotic task if component(s) of the robot are traversed to the particular pose. When the value satisfies one or more conditions (e.g., satisfies a threshold), the robot can be controlled to cause the component(s) to traverse to the particular pose in performing the robotic task.

Robotic control using action image(s) and critic network

Techniques are described herein for robotic control using value distributions. In various implementations, as part of performing a robotic task, state data associated with the robot in an environment may be generated based at least in part on vision data captured by a vision component of the robot. A plurality of candidate actions may be sampled, e.g., from continuous action space. A trained critic neural network model that represents a learned value function may be used to process a plurality of state-action pairs to generate a corresponding plurality of value distributions. Each state-action pair may include the state data and one of the plurality of sampled candidate actions. The state-action pair corresponding to the value distribution that satisfies one or more criteria may be selected from the plurality of state-action pairs. The robot may then be controlled to implement the sampled candidate action of the selected state-action pair.

Robotic control using value distributions

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for training an action selection neural network that is used to control a robotic agent interacting with a real-world environment.

Using simulation and domain adaptation for robotic control

Deep machine learning methods and apparatus related to the manipulation of an object by an end effector of a robot are described herein. Some implementations relate to training an action prediction network to predict a probability density which can include candidate actions of successful grasps by the end effector given an input image. Some implementations are directed to utilization of an action prediction network to visually servo a grasping end effector of a robot to achieve a successful grasp of an object by the grasping end effector.

Action prediction networks for robotic grasping

Growing community of inventors

Palo Alto, CA, United States of America

Mrinal Kalakrishnan