AnyBody: A Benchmark Suite for Cross-Embodiment Manipulation

Meenal Parakh, Alexandre Kirchmeyer, Beining Han, Jia Deng
Princeton University
arXiv Code (coming soon)

AnyBody is a benchmark to evaluate cross-embodiment generalization ability by focusing on two foundational tasks—reach and push. It enables testing generalization across three axes of morphological variations: (1) interpolation, which evaluates performance within a robot category sharing the same link structure; (2) extrapolation, which tests performance on robots with different link structures; and (3) composition, which assesses performance on combinations of link structures.

Morphology-aware Baselines

With AnyBody, we evaluate morphology-aware RL agents, that learn from environments interactions of train morphologies, and are tested for their ability to generalize to unseen morphologies.

Experimental Results

Can ME training outperform the SE baseline on train morphs?

Multi-task Scores

Can ME training zero-shot surpass SE on test morph?

Zero-shot Scores

Can fine-tuning help in complex scenarios?

Fine-tuning Scores

For more details, please refer to the paper.

BibTeX

@article{parakh2025anybody,
  author    = {Parakh, Meenal and Kirchmeyer, Alexandre and Han, Beining and Deng, Jia},
  title     = {AnyBody: A Benchmark Suite for Cross-Embodiment Manipulation},
  url       = {http://arxiv.org/abs/2505.14986},
  journal   = {arxiv},
  year      = {2025}
}