Active Embodiment Identification with Reinforcement Learning for Legged Robots

2026-05-08

AI Summary

This work addresses the challenge of accurately identifying intrinsic body parameters for legged robots operating under unknown morphologies. To this end, the authors propose an active self-identification framework that unifies information-seeking exploratory behaviors with explicit parameter prediction within a single modeling paradigm. The approach integrates a history-augmented URMA architecture with reinforcement learning, enabling the robot to actively infer both joint-level and global morphological parameters through environmental interaction in simulation. Experimental results demonstrate that the method achieves high-precision parameter identification across diverse morphologies, significantly enhancing cross-morphology adaptability and generalization performance.

Abstract

We present an active embodiment identification method for legged robots that jointly learns information-seeking behavior and explicit embodiment prediction. Using a history-augmented URMA architecture, the method infers joint-level and global embodiment parameters through interaction with the environment in simulation across different morphologies.

Problem

• embodiment identification
• legged robots
• active learning
• morphology
• reinforcement learning

Innovation

• active embodiment identification
• reinforcement learning
• legged robots
• history-augmented URMA
• embodiment prediction