Reproduce “TacPAC: Tactile Prediction and Real-Time Action Correction in World-Action Models for Contact-Rich 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
Zipei Ma, Xiaofei Wei, Junzhe Jiang, Shunlin Lu, Li Zhang
TacPAC combines a world-action model’s predicted visual and tactile observations with a tactile expert that uses cached expected contact to adjust only unexecuted actions during execution. Across five real-robot tasks, it achieved a 64% average episode success rate versus 48% for T-Rex. This result covers one robot setup and 20 trials per method per task.
Key insights
- Tactile prediction alone is insufficient for optimal contact-rich manipulation because crucial tactile changes are temporally sparse and typically occur *during* action execution, after an initial plan is committed.
- Interpreting real-time tactile feedback not in isolation, but *against an anticipated contact expectation* (derived from the original plan's prediction), is critical for enabling effective and context-aware action correction.
- An efficient, single-pass tactile expert that computes delta actions from cached predictions allows for high-frequency, real-time corrections without the computational overhead of full replanning, preventing error propagation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.