Reproduce “ImplicitRDP: An End-to-End Visual-Force Diffusion Policy with Structural Slow-Fast 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

Wendi Chen, Han Xue, Yi Wang, Fangyuan Zhou, Jun Lv, Yang Jin, Shirun Tang, Chuan Wen and 1 more

The policy combines visual and proprioceptive context with action-rate force feedback in one causal model, letting the robot revise action chunks as force readings arrive. It also predicts a compliance-inspired virtual position target during training. Across three real-world tasks, it succeeds in 18/20 switch-toggling trials, versus 10/20 with RDP and 8/20 with vision-only DP. The evidence is limited to tested tasks and robot setup.

Key insights

  • Achieving closed-loop force control within action chunking in an end-to-end architecture is possible by enforcing a temporally causal structure and using a deterministic diffusion inference mechanism.
  • A physics-grounded virtual target formulation, coupled with adaptive stiffness weighting, provides a robust representation regularization that encourages the policy to effectively leverage force feedback and mitigates modality collapse.
  • Specific choices in prediction parameterization (velocity-prediction) and rotation representation (Euler angles) enhance the stability and success rates of diffusion policies in contact-rich manipulation scenarios.

Summary by alphaXiv.

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