Robot action improves with better future video predictions using self learning
Self-Distilling Future Video Representations for World Action Models
RoboticsComputer Vision and Pattern Recognition
Summary
Robots need to predict future events to plan their actions well. The authors found a way to help robots learn better future video information from noisy data in just one step, using a technique called self-distillation. They also let the robot consider past observations to improve future predictions. This method helps robots perform tasks better, especially in new situations and when learning skills from humans, even if those tasks were not part of the robot’s original training.
What this means in practice
- •For robotics engineers: Build robots that better predict future scenes to improve manipulation and task success in varied environments without extra retraining.
- •For automation system designers: Create automation systems capable of transferring human demonstration skills to robots for tasks unseen during robot training.
Authors
Huiwon Jang, Suhyeok Jang, Hyeonmo Kang, Seungku Kim, Jinwoo Shin, Sihyun Yu
Abstract
World Action Models (WAMs) have emerged as a promising approach to learning generalist robot policies using rich spatiotemporal priors from pre-trained video diffusion models. A recent line of WAMs improves inference efficiency by conditioning a separate action decoder on intermediate future video representations extracted from a single denoising step, avoiding costly iterative denoising required for video generation. This design makes action prediction depend directly on the quality of these future representations. In this paper, we introduce Future REpresentation learning via self-Distillation with temporal context (FRED), a framework that trains WAMs to improve the quality of future video representations from noise in a single denoising step. Specifically, FRED uses self-distillation to predict future representations from noise, using representations extracted from ground-truth future observations as targets. We further condition the video model on historical observations, providing temporal context for predicting future representations better. Across diverse real-world manipulation scenarios, FRED consistently outperforms competing WAMs. It achieves particularly large gains on long-horizon and out-of-domain tasks under zero-shot environment generalization, and substantially improves human-to-robot skill transfer, reaching 50.0% task success on tasks absent from robot training data and demonstrated only in human videos, compared with 19.4% for the best competing WAM.