Scaling Manipulation Learning with Visual Kinematic Chain Prediction

Xinyu Zhang, Yuhan Liu, Haonan Chang, Abdeslam Boularias
Rutgers University

Abstract

Learning general-purpose models from diverse datasets has achieved great success in machine learning. In robotics, however, existing methods in multi-task learning are typically constrained to a single robot and workspace, while recent work such as RT-X requires a non-trivial action normalization procedure to manually bridge the gap between different action spaces in diverse environments. In this paper, we propose the visual kinematics chain as a precise and universal representation of quasi-static actions for robot learning over diverse environments, which requires no manual adjustment since the visual kinematic chains can be automatically obtained from the robot’s model and camera parameters. We propose the Visual Kinematics Transformer (VKT), a convolution-free architecture that supports an arbitrary number of camera viewpoints, and that is trained with a single objective of forecasting kinematic structures through optimal point-set matching. We demonstrate the superior performance of VKT over BC transformers as a general agent on Calvin, RLBench, Open-X, and real robot manipulation tasks.

MY ALT TEXT

We also find visual kinematic forecasting enables the use of 2D image augmentation for manipulation learning because the action space also resides in image plane, which is important to prevent overfitting given the small training set.
Existing manipulation learning augmentations are only applied in the 3D space, which requires depth and has fewer categories.