Scaling Manipulation Learning with Visual Kinematic Chain Prediction

TL;DR

This paper introduces the Visual Kinematics Transformer (VKT), a novel model that predicts robot kinematic structures from multiple camera views, enabling versatile multi-task robot learning without manual action normalization.

Contribution

The paper proposes the visual kinematics chain as a universal representation and introduces VKT, a convolution-free architecture supporting multiple viewpoints for diverse robot manipulation tasks.

Findings

VKT outperforms BC transformers on various benchmarks.

VKT supports arbitrary camera viewpoints without manual adjustments.

The approach enables general-purpose robot learning across diverse environments.

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. Video demonstrations can be found at https://mlzxy.github.io/visual-kinetic-chain.