Google Inc.

X Development LLC

Adobe, Inc.

Gdm Holding LLC

Yunfei Bai

Seyed Mohammad Khansari Zadeh

Mrinal Kalakrishnan

Sergey Levine

Matthew Bennice

Brent Austin

Eric Jang

Honglak Lee

Yuchen Wu

Chelsea Finn

Alexander Irpan

Daniel Kappler

Jasmine Hsu

Xinchen Yan

Abhinav Gupta

Daniel Max Kaufman

Daniel Alex Lam

Paul Wohlhart

Elmar Mair

Soeren Pirk

Alexander Herzog

Kuan Fang

James Christopher Davidson

Ian Wilkes

Tigran Gasparian

Andreas Christiansen

Yuanzheng Gong

Nicolas Sievers

Max Moroz

Weidan Wu

Arkanath Pathak

Vincent Olivier Vanhoucke

Kurt Konolige

Stefan Hinterstoisser

Paul Bechard

Konstantinos Bousmalis

Pierre Sermanet

Julian Ibarz

Jie Tan

Chuyuan Fu

Wenlong Lu

Harrison Lynch

Daniel Ho

Tingnan Zhang

Atil Iscen

Erwin Coumans

Allan Zhou

C Karen Liu

Zhuo Xu

Wenhao Yu

Matthew Laurance Kelcey

Sergey Vladimir Levine

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 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

Training and/or using a machine learning model for locomotion control of a robot, where the model is decoupled. In many implementations, the model is decoupled into an open loop component and a feedback component, where a user can provide a desired reference trajectory (e.g., a symmetric sine curve) as input for the open loop component. In additional and/or alternative implementations, the model is decoupled into a pattern generator component and a feedback component, where a user can provide controlled parameter(s) as input for the pattern generator component to generate pattern generator phase data (e.g., an asymmetric sine curve). The neural network model can be used to generate robot control parameters.

Systems and methods for learning agile locomotion for multiped robots

Generating edge-depth values for an object, utilizing the edge-depth values in generating a 3D point cloud for the object, and utilizing the generated 3D point cloud for generating a 3D bounding shape (e.g., 3D bounding box) for the object. Edge-depth values for an object are depth values that are determined from frame(s) of vision data (e.g., left/right images) that captures the object, and that are determined to correspond to an edge of the object (an edge from the perspective of frame(s) of vision data). Techniques that utilize edge-depth values for an object (exclusively, or in combination with other depth values for the object) in generating 3D bounding shapes can enable accurate 3D bounding shapes to be generated for partially or fully transparent objects. Such increased accuracy 3D bounding shapes directly improve performance of a robot that utilizes the 3D bounding shapes in performing various tasks.

Robotic control based on 3D bounding shape, for an object, generated using edge-depth values for the object

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

Implementations relate to training a point cloud prediction model that can be utilized to process a single-view two-and-a-half-dimensional (2.5D) observation of an object, to generate a domain-invariant three-dimensional (3D) representation of the object. Implementations additionally or alternatively relate to utilizing the domain-invariant 3D representation to train a robotic manipulation policy model using, as at least part of the input to the robotic manipulation policy model during training, the domain-invariant 3D representations of simulated objects to be manipulated. Implementations additionally or alternatively relate to utilizing the trained robotic manipulation policy model in control of a robot based on output generated by processing generated domain-invariant 3D representations utilizing the robotic manipulation policy model.

Robotic manipulation using domain-invariant 3D representations predicted from 2.5D vision data

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

Mitigating the reality gap through training and utilization of at least one difference model. The difference model can be utilized to generate, for each of a plurality of instances of simulated state data generated by a robotic simulator, a corresponding instance of modified simulated state data. The difference model is trained so that a generated modified instance of simulated state data is closer to 'real world data' than is a corresponding initial instance of simulated state data. Accordingly, the difference model can be utilized to mitigate the reality gap through modification of initially generated simulated state data, to make it more accurately reflect what would occur in a real environment. Moreover, the difference representation from the difference model can be used as input to the control policy to adapt the control learned from simulator to the real environment.

Mitigating reality gap through modification of simulated state data of robotic simulator

Training a machine learning model (e.g., a neural network model such as a convolutional neural network (CNN) model) so that, when trained, the model can be utilized in processing vision data (e.g., from a vision component of a robot), that captures an object, to generate a rich object-centric embedding for the vision data. The generated embedding can enable differentiation of even subtle variations of attributes of the object captured by the vision data.

Training a deep neural network model to generate rich object-centric embeddings of robotic vision data

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

Implementations are directed to generating simulated training examples for training of a machine learning model, training the machine learning model based at least in part on the simulated training examples, and/or using the trained machine learning model in control of at least one real-world physical robot. Implementations are additionally or alternatively directed to performing one or more iterations of quantifying a 'reality gap' for a robotic simulator and adapting parameter(s) for the robotic simulator based on the determined reality gap. The robotic simulator with the adapted parameter(s) can further be utilized to generate simulated training examples when the reality gap of one or more iterations satisfies one or more criteria.

Generating simulated training examples for training of machine learning model used for robot control

Mitigating the reality gap through optimization of one or more simulated hardware parameters for simulated hardware components of a simulated robot. Implementations generate and store real navigation data instances that are each based on a corresponding episode of locomotion of a real robot. A real navigation data instance can include a sequence of velocity control instances generated to control a real robot during a real episode of locomotion of the real robot, and one or more ground truth values, where each of the ground truth values is a measured value of a corresponding property of the real robot (e.g., pose). The velocity control instances can be applied to a simulated robot, and one or more losses can be generated based on comparing the ground truth value(s) to corresponding simulated value(s) generated from applying the velocity control instances to the simulated robot. The simulated hardware parameters and environmental parameters can be optimized based on the loss(es).

Mitigating reality gap through optimization of simulated hardware parameter(s) of simulated robot

Growing community of inventors

Fremont, CA, United States of America

Yunfei Bai