Reproduce “Blind Dexterity: Whole-Body Humanoid Manipulation via Pure Proprioception” 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
Aditya Bhatt, Oleg Kaidanov, Puze Liu, Jan Peters
The paper shows that humanoid joint-encoder histories can support contact-driven object manipulation without cameras or dedicated touch sensors. Under compliant position control, policies use joint motion, previous commands, and stock IMU to search through sweeping and tapping. In simulated football trapping, blind-from-scratch policies reached 92.9 ± 1.2% success, but this does not establish real-world reliability or generalization to unseen geometries.
Key insights
- Compliant joint-space dynamics, specifically joint tracking residuals, can serve as an implicit, whole-body tactile sensing channel for contact-rich interaction.
- Robots can learn active exploratory behaviors through physical contact, systematically generating informative interactions to infer latent object states from proprioceptive data.
- Short histories of raw proprioceptive data often provide sufficient information for robust control, with explicit state estimator feedback not consistently yielding significant performance benefits.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.