Cross-Embodiment Robot Learning
Overview
The field studies how robots can reuse skills and data across different morphologies, from dexterous hands and humanoids to mobile, aerial, and legged platforms. Common approaches use canonical or latent action spaces, morphology-aware graph and transformer policies, motion retargeting, and embodiment-specific decoders; human videos, simulation, and robot demonstrations often provide complementary training data. Vision-language-action and world-action models also learn shared task intent or physical dynamics, then translate predictions into controls suited to a target robot, with evaluation spanning manipulation, locomotion, and navigation benchmarks and real hardware. The central tension is how to share enough structure for transfer without ignoring embodiment-specific kinematics, contact, sensing, and safety constraints; zero-shot results are promising, but reliable performance on unfamiliar robots and settings remains difficult.
Latest papers
350 papers87 papers in the last four weeks, up 98% on the four weeks before. 0.5% of all new papers.