This is a page that curates AI-related papers published worldwide. All content here is summarized using Google Gemini and operated on a non-profit basis. Copyright for each paper belongs to the authors and their institutions; please make sure to credit the source when sharing.
Towards Embodiment Scaling Laws in Robot Locomotion
Created by
Haebom
Author
Bo Ai, Liu Dai, Nico Bohlinger, Dichen Li, Tongzhou Mu, Zhanxin Wu, K. Fay, Henrik I. Christensen, Jan Peters, Hao Su
Outline
This paper explores "embodiment scaling laws" that improve generalization performance to new robot morphologies by training on diverse robot morphologies. We generated approximately 1,000 different robot morphologies and trained a policy on a subset of them. We observed that generalization performance to new robot morphologies improved as the number of trained morphologies increased. This demonstrates that our approach enables more effective generalization than conventional data scaling methods. Notably, the top-performing policy trained on the full dataset successfully performed zero-shot transfer learning on new robot morphologies on both simulations and real robots (including Unitree Go2 and H1). This research suggests potential applications in diverse fields such as adaptive control and shape co-design, and represents a step toward general embodied intelligence.
Takeaways, Limitations
•
Takeaways:
◦
We demonstrate that training using a variety of robot shapes can significantly improve generalization performance to new robot shapes.
◦
Experimental confirmation of the existence of 'embodiment scaling laws'.
◦
A more efficient generalization learning method than conventional data expansion methods is presented.
◦
Zero-shot transfer learning success on both simulations and real robots.
◦
It presents potential applications in various fields such as adaptive control and shape co-design.
•
Limitations:
◦
Further research is needed on the diversity and complexity of the robot forms used in this study.
◦
Further verification is needed regarding the diversity of robot shapes and robustness to environmental changes in real-world applications.
◦
Further research is needed to evaluate training and generalization performance for large-scale robot configurations.
