Reproduce “Imitation Learning from a Single Temporally Misaligned Video” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
William Huey, Huaxiaoyue Wang, Anne Wu, Yoav Artzi, Sanjiban Choudhury
ORCA (ORdered Coverage Alignment) is presented as a reward function for imitation learning, enabling agents to learn complex sequential tasks from a single, temporally misaligned video demonstration. This method achieves a 4.5x improvement in average normalized return over state-of-the-art baselines on Meta-world tasks and a 6.6x improvement on Humanoid-v4, by robustly enforcing subgoal ordering and coverage.
Key insights
- For sequential tasks, effective imitation requires matching entire trajectories based on the *ordered coverage* of subgoals, rather than just individual frame-level states.
- A reward function must be inherently non-Markovian to accurately penalize violations of temporal ordering and ensure full task completion.
- Recursively calculating the probability of covering subgoals in sequence provides a principled way to construct a dense, task-aware reward from a single, potentially misaligned demonstration.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.