Nvidia Corporation

Thang Hong To

Jonathan Tremblay

Stanley Thomas Birchfield

Varun Jampani

David Jesus Acuna Marrero

Aayush Prakash

Cem Anil

Mark A Brophy

Jan Kautz

Stephen Walter Tyree

Stan Birchfield

Duncan J McKay

Shariq Iqbal

Erik Leitch

Jia Cheng

Artem Molchanov

Various embodiments enable a robot, or other autonomous or semi-autonomous device or system, to receive data involving the performance of a task in the physical world. The data can be provided as input to a perception network to infer a set of percepts about the task, which can correspond to relationships between objects observed during the performance. The percepts can be provided as input to a plan generation network, which can infer a set of actions as part of a plan. Each action can correspond to one of the observed relationships. The plan can be reviewed and any corrections made, either manually or through another demonstration of the task. Once the plan is verified as correct, the plan (and any related data) can be provided as input to an execution network that can infer instructions to cause the robot, and/or another robot, to perform the task.

Learning robotic tasks using one or more neural networks

In at least one embodiment, under the control of a robotic control system, a gripper on a robot is positioned to grasp a 3-dimensional object. In at least one embodiment, the relative position of the object and the gripper is determined, at least in part, by using a camera mounted on the gripper.

Robotic control system

Training deep neural networks requires a large amount of labeled training data. Conventionally, labeled training data is generated by gathering real images that are manually labelled which is very time-consuming. Instead of manually labelling a training dataset, domain randomization technique is used generate training data that is automatically labeled. The generated training data may be used to train neural networks for object detection and segmentation (labelling) tasks. In an embodiment, the generated training data includes synthetic input images generated by rendering three-dimensional (3D) objects of interest in a 3D scene. In an embodiment, the generated training data includes synthetic input images generated by rendering 3D objects of interest on a 2D background image. The 3D objects of interest are objects that a neural network is trained to detect and/or label.

Neural network model trained using generated synthetic images

Generation of synthetic images for training a neural network model

An object detection neural network receives an input image including an object and generates belief maps for vertices of a bounding volume that encloses the object. The belief maps are used, along with three-dimensional (3D) coordinates defining the bounding volume, to compute the pose of the object in 3D space during post-processing. When multiple objects are present in the image, the object detection neural network may also generate vector fields for the vertices. A vector field comprises vectors pointing from the vertex to a centroid of the object enclosed by the bounding volume defined by the vertex. The object detection neural network may be trained using images of computer-generated objects rendered in 3D scenes (e.g., photorealistic synthetic data). Automatically labelled training datasets may be easily constructed using the photorealistic synthetic data. The object detection neural network may be trained for object detection using only the photorealistic synthetic data.

Detecting and estimating the pose of an object using a neural network model

Growing community of inventors

Redmond, WA, United States of America

Thang Hong To