Reproduce “TacCoRL: Integrating Tactile Feedback into VLA via Simulation” 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
Siyu Ma, Yuqi Liang, Chang Yu, Yunuo Chen, Hao Su, Yixin Zhu, Yin Yang, Chenfanfu Jiang
TacCoRL integrates tactile feedback into pretrained Vision-Language-Action (VLA) policies using a simulation-to-real pipeline to improve robot performance in contact-rich manipulation. The system learns contact-guided corrective behaviors from simulated "near-failure" states, increasing average real-world success rates from 50.0% (vision-only) to 72.5% across tasks.
Key insights
- Accurate sim-to-real alignment across visual, control, and tactile modalities is fundamental for effectively transferring learned contact-guided behaviors from simulation to physical robots.
- Combining sim-real co-training for initialization with real-data-anchored RL post-training effectively enables learning critical error recovery actions in simulation that generalize to real-world tasks.
- Tactile feedback proves essential for correcting localized contact errors, especially when visual information is ambiguous or occluded, leading to significantly enhanced robustness in manipulation tasks.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.