NUS HINT-WAM Uses Future-Video Hindsight to Self-Distill World Action Models Without Reward Shaping
Summary
Researchers at the National University of Singapore released HINT-WAM, the first on-policy self-distillation framework for world action models in robot manipulation. The student policy generates actions along its own denoising trajectory, while the same model acts as a teacher conditioned on the demonstration's future video (hindsight) and supplies dense action-distribution targets — avoiding the sparse-reward problem of RL and the expert-trajectory bias of supervised fine-tuning. Paper posted Nov. 19 by Phuong Tuan Dat, Qi Li and Xinchao Wang.
Originally reported by huggingface.co
Read the original article →Original headline: NUS HINT-WAM Uses Future-Video Hindsight to Self-Distill World Action Models Without Reward Shaping