TimeRewarder: Learning Dense Reward from Passive Videos via Frame-wise Temporal Distance
Researchers have developed TimeRewarder, a novel method for learning dense reward signals from passive videos. This technique models temporal distances between frame pairs to estimate task progress, which can then guide reinforcement learning agents. Experiments on ten Meta-World tasks showed TimeRewarder significantly improved success rates and sample efficiency, outperforming manually designed rewards and previous methods. The approach also demonstrated potential in leveraging real-world human videos for scalable reward signal generation. AI
IMPACT Enables more efficient training of reinforcement learning agents by automating reward design from video data.
