Towards Embodiment Scaling Laws in Robot Locomotion

2025-05-09

AI Summary

This work investigates how scaling robot embodiment affects cross-embodiment generalization—a core challenge for generalist robots requiring adaptation to diverse physical forms. To address this, we introduce a large-scale simulation dataset programmatically generating 1,000 heterogeneous morphologies (bipedal, quadrupedal, hexapedal) and propose a unified policy representation with mechanisms for adapting to heterogeneous observation and action spaces. We provide the first empirical evidence that increasing the number of training morphologies significantly improves zero-shot transfer performance to unseen embodiments—outperforming mere data scaling within a single morphology. Critically, policies trained across all morphologies achieve zero-shot deployment on real-world novel platforms—including Unitree Go2 and H1—without fine-tuning, substantially surpassing single-morphology baselines. Our results establish a scalable paradigm and key empirical foundation for embodied intelligence toward cross-embodiment general-purpose control.

Technology Category

• Intelligent Robots: Embodied AI
• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Machine Learning: Imitation Learning & Inverse Reinforcement Learning

Application Category

• Search and Retrieval-Augmented AI: Multilingual and cross-lingual Web search
• Responsible Web: Machine-in-the-loop, human agency and autonomy
• Economics, Online Markets and Human Computation: Economic ramifications for generative AI infrastructure and applications

Abstract

Developing generalist agents that can operate across diverse tasks, environments, and physical embodiments is a grand challenge in robotics and artificial intelligence. In this work, we focus on the axis of embodiment and investigate embodiment scaling laws$unicode{x2013}$the hypothesis that increasing the number of training embodiments improves generalization to unseen ones. Using robot locomotion as a test bed, we procedurally generate a dataset of $sim$1,000 varied embodiments, spanning humanoids, quadrupeds, and hexapods, and train generalist policies capable of handling diverse observation and action spaces on random subsets. We find that increasing the number of training embodiments improves generalization to unseen ones, and scaling embodiments is more effective in enabling embodiment-level generalization than scaling data on small, fixed sets of embodiments. Notably, our best policy, trained on the full dataset, zero-shot transfers to novel embodiments in the real world, such as Unitree Go2 and H1. These results represent a step toward general embodied intelligence, with potential relevance to adaptive control for configurable robots, co-design of morphology and control, and beyond.

Problem

• Investigates how increasing training embodiments improves generalization to unseen ones
• Develops generalist policies for diverse robot locomotion tasks and embodiments
• Demonstrates zero-shot transfer to novel real-world robot embodiments

Innovation

• Procedurally generate 1,000 diverse robot embodiments
• Train generalist policies on varied observation and action spaces
• Zero-shot transfer to novel real-world embodiments