Reproduce “Visual-Tactile Peg-in-Hole Assembly Learning from Peg-out-of-Hole Disassembly” 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

Yongqiang Zhao, Xuyang Zhang, Zhuo Chen, Matteo Leonetti, Emmanouil Spyrakos-Papastavridis, Shan Luo

Researchers from King's College London developed a visual-tactile skill learning framework for robot peg-in-hole assembly by leveraging expert data from the inverse peg-out-of-hole disassembly task. This method achieved an 80.6% average success rate across diverse object pairs in simulation and 72.1% in real-world experiments, outperforming direct reinforcement learning and pure supervised learning baselines.

Key insights

  • The inherent simplicity of the inverse PooH task makes it an effective and data-efficient source for generating expert demonstrations for the more complex PiH assembly.
  • Bridging the dynamic gap between reversed PooH and actual PiH contact patterns is critical, achieved by regenerating tactile data and introducing action randomization during demonstration generation.
  • The synergy of visual (for global context and coarse alignment) and tactile (for local contact and fine force control) sensing is crucial for robust and compliant PiH assembly, leading to lower contact forces.

Summary by alphaXiv.

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