Reproduce “DORA: Object Affordance-Guided Reinforcement Learning for Dexterous Robotic 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

Lei Zhang, Soumya Mondal, Zhenshan Bing, Kaixin Bai, Diwen Zheng, Zhaopeng Chen, Alois Christian Knoll, Jianwei Zhang

DORA presents a framework that integrates object affordance knowledge into reinforcement learning for dexterous robotic hands, using semantic information to guide functional grasp pose generation and shape the reward function. This approach improved task success rates by an average of 15.4% and reduced training time in simulated multi-fingered manipulation scenarios involving objects like jugs and hammers.

Key insights

  • Integrating object affordances as both constraints for grasp selection and components of the reward function provides a powerful semantic prior, effectively bridging high-level understanding with low-level motor control.
  • Affordance guidance significantly enhances sample efficiency and accelerates policy convergence by constraining the exploration space to functionally relevant and physically feasible interactions.
  • The approach is particularly effective for tasks demanding the selection of semantically meaningful contact points, reducing ambiguity in the large contact space of dexterous hands.

Summary by alphaXiv.

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