AnyBody: A Benchmark Suite for Cross-Embodiment Manipulation

TL;DR

The paper introduces AnyBody, a benchmark to evaluate manipulation policies across diverse robot morphologies, addressing cross-embodiment generalization along interpolation, extrapolation, and composition. It evaluates morphology-conditioned PPO policies using both MLP and Transformer backbones, comparing single-embodiment and multi-embodiment training. Results indicate multi-embodiment training improves in-distribution performance and can enhance zero-shot interpolation generalization with transformers, but zero-shot generalization to unseen morphologies—especially in extrapolation and composition—remains challenging. The work provides an open-source IsaacSim extension and detailed guidance on architecture and training choices, highlighting directions for improving morphological reasoning and adaptation in robotic manipulation.

Abstract

Generalizing control policies to novel embodiments remains a fundamental challenge in enabling scalable and transferable learning in robotics. While prior works have explored this in locomotion, a systematic study in the context of manipulation tasks remains limited, partly due to the lack of standardized benchmarks. In this paper, we introduce a benchmark for learning cross-embodiment manipulation, focusing on two foundational tasks-reach and push-across a diverse range of morphologies. The benchmark is designed to test generalization along three axes: interpolation (testing performance within a robot category that shares the same link structure), extrapolation (testing on a robot with a different link structure), and composition (testing on combinations of link structures). On the benchmark, we evaluate the ability of different RL policies to learn from multiple morphologies and to generalize to novel ones. Our study aims to answer whether morphology-aware training can outperform single-embodiment baselines, whether zero-shot generalization to unseen morphologies is feasible, and how consistently these patterns hold across different generalization regimes. The results highlight the current limitations of multi-embodiment learning and provide insights into how architectural and training design choices influence policy generalization.

Figures (12)

• Figure 1: We introduce AnyBody, a benchmark suite for evaluating policy generalization across diverse robot morphologies. It consists of 18 robot variations: 8 procedurally generated robot categories and 10 based on real-world robots. The benchmark tasks comprise two manipulation tasks —reach and push — two scene variations (with and without obstacles), and two input types—state-based and point cloud-based.
• Figure 2: Three categories of benchmark tasks. We aim to test the zero-shot generalization ability of a multi-embodiment policy $\pi(s)$ on unseen morphologies.
• Figure 3: Cosine similarity of test morphologies with those in the train set.
• Figure 4: Robot morphology is represented by a sequence of links. We approximate the link geometries by the shape parameters of a fitted primitive.
• Figure 5: Large-scale multi-embodiment training in Isaac-Sim.