Towards Embodiment Scaling Laws in Robot Locomotion

TL;DR

This paper investigates how increasing the diversity of training robot embodiments improves the ability of policies to generalize to unseen robots, demonstrating positive scaling laws in locomotion tasks.

Contribution

It introduces embodiment scaling laws in robot locomotion, showing that more diverse training embodiments enhance zero-shot transfer to new robots in simulation and real-world.

Findings

Embodiment diversity improves generalization more than data scaling.

Policies trained on full embodiment datasets transfer zero-shot to unseen robots.

Positive scaling trends support embodiment diversity as a key factor in generalization.

Abstract

Cross-embodiment generalization underpins the vision of building generalist embodied agents for any robot, yet its enabling factors remain poorly understood. We investigate embodiment scaling laws, the hypothesis that increasing the number of training embodiments improves generalization to unseen ones, using robot locomotion as a test bed. We procedurally generate ~1,000 embodiments with topological, geometric, and joint-level kinematic variations, and train policies on random subsets. We observe positive scaling trends supporting the hypothesis, and find that embodiment scaling enables substantially broader generalization than data scaling on fixed embodiments. Our best policy, trained on the full dataset, transfers zero-shot to novel embodiments in simulation and the real world, including the Unitree Go2 and H1. These results represent a step toward general embodied intelligence, with relevance to adaptive control for configurable robots, morphology co-design, and beyond.