Learning Humanoid Locomotion with Morphology Aware Transformers
Abstract: Current humanoid locomotion policies trained with deep reinforcement learning are tightly coupled to the specific robot morphology they were trained on, making them brittle to variations in joint friction, link mass, or limb proportions. Prior transformer-based approaches rely on one-hot encodings to distinguish joints, which carry no structural meaning and fail to transfer to unseen morphologies, leaving a critical gap in morphology-agnostic generalization. This work addresses that gap by representing the robot as a kinematic graph and introducing two structure-aware components: Random Walk Structural Encodings (RWSE) that capture each joint's relational position within the body, and a FiLM conditioning layer that injects global physical properties into the transformer at inference time without retraining. Compared to one-hot baselines, this approach produces semantically richer joint embeddings (validated via t-SNE) and improved locomotion performance on unseen morphology variants in IsaacGym experiments with the Unitree G1 robot. Future work will extend evaluation to robots with varying joint and link counts, characterize the limits of zero-shot transfer, and explore integration of formal behavioral specifications for reasoning about policy reliability under morphology change.
Advisor: Professor Mahdi Agheli
Committee: Professor Jing Xiao, Professor Constantinos Chamzas