Reproduce “Vegetable Peeling: A Case Study in Constrained Dexterous 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
Tao Chen, Eric Cousineau, Naveen Kuppuswamy, Pulkit Agrawal
A robotic system learns to reorient various vegetables using a dexterous hand while satisfying constraints like firm grasping, avoiding finger obstruction, and enabling precise control via external stop signals. This constrained manipulation capability enables the system to reliably integrate with a second arm performing vegetable peeling, demonstrating robust performance across diverse objects and achieving autonomous peeling through a vision-based pipeline.
Key insights
- A simplified reward design using a single 'keyframe demonstration' effectively guides policy learning for contact-rich dexterous manipulation, reducing hyperparameter tuning.
- Reset-based constraint enforcement, where episodes are immediately reset upon violation, implicitly shapes policy behavior to adhere to spatial and contact constraints without complex reward penalties.
- Transformer architectures demonstrate superior learning speed and final performance compared to RNNs for processing historical observations in sequential dexterous manipulation policies.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.