Reproduce “Efficient Self-Supervised Data Collection for Offline Robot Learning” 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
Shadi Endrawis, Gal Leibovich, Guy Jacob, Gal Novik, Aviv Tamar
A self-supervised method for efficient data collection in offline robot learning combines Random Network Distillation (RND) with dynamically sampled goal-conditioned policies to generate diverse, high-quality datasets. The resulting datasets consistently yield higher performance in downstream supervised learning and offline reinforcement learning tasks.
Key insights
- Active, goal-conditioned exploration guided by intrinsic novelty is more efficient than passive or random exploration for covering state spaces.
- Dynamically sampling goals from a replay buffer based on RND novelty effectively directs the agent to continuously discover novel states.
- Random Network Distillation (RND) can robustly estimate novelty in high-dimensional visual state spaces, enabling its use for complex robotic tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.