Data-Efficient Approach to Humanoid Control by Fine-Tuning a Pre-Trained GPT on Action Data

Recent advances in imitation learning have enabled robots to learn multiple tasks from large-scale datasets. However, developing a model for multi-tasking humanoid control faces significant challenges. Human kinematic data is available in open-source datasets for humanoid motion learning, but learning policies from this data requires simulation due to the lack of actions. While dynamic data can accelerate learning via supervision, datasets typically lack substantial amounts of such action labels, that are also difficult to be directly used for training due to the unique structure of each human/humanoid systems. In this study, we pre-trained a Generative Pre-trained Transformer (GPT) based model on expert policy-rollout observations only (without actions) from a humanoid motion dataset. Upon fine-tuning on a smaller dataset with both observations and action labels, we demonstrate that our GPT-based model can predict actions to achieve human-like movements faster in training than training a GPT on the entire dataset from scratch directly. Furthermore, performance evaluation based on motion generation across various behaviors showed that our approach achieves efficient learning comparable to baselines.