Reproduce “Copper-Policy: Focus on the Representation for Robust Robot Manipulation” 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

Zexin Feng, Yixu Feng, Lingyu Xiao, Shang Su, Kexin Zheng, Chang Xu, Mengkai Shi, Shuo Feng and 1 more

Copper-Policy trains a robot policy using a compact task-conditioned representation of scene changes and a separate pathway for current-frame visual detail. It jointly trains prediction and action decoding; at deployment, its Future Expert encodes context without generating future frames. On three real-robot tasks, it averages 96.3, versus 94.3 for π₀.₅, but this is an average score, not universal success.

Key insights

  • Future prediction can shape a policy’s internal representation without requiring future-image generation or test-time rollout planning.
  • Ablations indicate that predictive World tokens and direct access to current visual features provide complementary benefits, especially on the real-robot tasks.
  • Strong results on several evaluations do not imply uniform robustness: randomized RoboTwin performance is low, and language and layout generalization remain challenging.

Summary by alphaXiv.

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