LocoGPT: GPT-Based Multi-Humanoid-Task Policy for Humanoid Locomotion

Siddharth Padmanabhan, Kazuki Miyazawa, Takato Horii

Abstract: Recent studies have explored developing controllers that can generalize across several humanoid robots that differ in shape and size. However, limited studies have developed cross-embodiment policies for humanoids in a low data regime with varied observation and action spaces. In this study, we propose a GPT-based multi-humanoid locomotion policy called LocoGPT, (where GPT refers to Generative Pre-Trained Transformer), trained on multiple humanoids with varied observation and action spaces that can generate various locomotion behaviors for durations that touch close to one minute on an average. From our evaluations, we show that LocoGPT has high cross-embodiment learning capability and can generalize locomotion behaviors like walking, running, and carrying across different humanoids better than the baselines. LocoGPT outperforms the Multi-Layer Perceptron (MLP) baselines in generating actions for locomotion-based tasks, and can do so by leveraging body specific modules with a shared actor network in a low data regime. We further show that the motions generated by LocoGPT are stable and relatively close to expert data using metrics such as action rate and joint acceleration and jerk smoothness ratios. Additionally, we condense the model as much as possible and show that we can retain similar performance levels.

External IDs:dblp:journals/access/PadmanabhanMH26