Reproduce “Keyframe-Focused Visual Imitation 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

Chuan Wen, Jierui Lin, Jianing Qian, Yang Gao, Dinesh Jayaraman

Keyframe-Focused Visual Imitation Learning addresses the copycat problem in behavioral cloning by reweighting training samples based on action prediction error. This approach enables visual imitation policies to robustly leverage observation histories, leading to improved task success rates and reduced distributional shift in environments like CARLA and MuJoCo-Image.

Key insights

  • The copycat problem stems from strong temporal correlations in expert actions combined with an underrepresentation of critical "changepoint" samples in the training data.
  • By increasing the effective training error of a simple copycat strategy through sample reweighting, the learning algorithm is guided away from this shortcut.
  • A lightweight, separately trained model can effectively identify crucial changepoint samples, enabling a scalable solution for visual imitation learning.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.