Reproduce “Behavioral Cloning Mystery” 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
Seohong Park, Sergey Levine
OCBench provides a simulated laboratory for studying behavioral cloning with controllable scripted demonstrations designed to have narrow, temporally correlated trajectories and varied strategies. It covers 28 tasks and supports fresh-data experiments, generating up to 200,000 simulation environment steps per second on one H100 GPU. In flow behavioral cloning, 25-action policies outperform one-action policies on two tested tasks.
Key insights
- Success can keep improving after validation loss worsens, and larger datasets can sometimes reduce success—even when the data are sampled from the same distribution.
- Executing 25-action chunks can outperform replanning at every step; tests suggest that both temporal dependencies in demonstrations and accumulated errors may contribute.
- Task performance depends on model scale and input representation, while offline loss and action MSE can fail to reflect generalization to states encountered during execution.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.