Towards Embodiment Scaling Laws in Robot Locomotion
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2025
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866918132352811008 |
|---|---|
| author | Ai, Bo Dai, Liu Bohlinger, Nico Li, Dichen Mu, Tongzhou Wu, Zhanxin Fay, K. Christensen, Henrik I. Peters, Jan Su, Hao |
| author_facet | Ai, Bo Dai, Liu Bohlinger, Nico Li, Dichen Mu, Tongzhou Wu, Zhanxin Fay, K. Christensen, Henrik I. Peters, Jan Su, Hao |
| contents | 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. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2505_05753 |
| institution | arXiv |
| publishDate | 2025 |
| record_format | arxiv |
| spellingShingle | Towards Embodiment Scaling Laws in Robot Locomotion Ai, Bo Dai, Liu Bohlinger, Nico Li, Dichen Mu, Tongzhou Wu, Zhanxin Fay, K. Christensen, Henrik I. Peters, Jan Su, Hao Robotics Artificial Intelligence Machine Learning 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. |
| title | Towards Embodiment Scaling Laws in Robot Locomotion |
| topic | Robotics Artificial Intelligence Machine Learning |
| url | https://arxiv.org/abs/2505.05753 |