Towards Embodiment Scaling Laws in Robot Locomotion
This addresses the problem of building generalist embodied agents for any robot, with relevance to adaptive control and morphology co-design, though it is incremental as it tests a hypothesis in a specific domain.
The paper investigates whether training policies on more robot embodiments improves generalization to unseen ones, finding positive scaling trends where embodiment scaling enables broader generalization than data scaling on fixed embodiments, with their best policy achieving zero-shot transfer to novel real-world robots like Unitree Go2 and H1.
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.