Reproduce “Learning Dual-Arm Push and Grasp Synergy in Dense Clutter” 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

Yongliang Wang, Hamidreza Kasaei

A dual-arm hierarchical Deep Reinforcement Learning framework is introduced for grasping specific objects within densely cluttered environments. This approach learns adaptive dual-arm push and 6-DoF grasp actions directly from visual inputs, demonstrating improved completion rates and grasp success using fewer motions than single-arm systems, with successful direct transfer from simulation to real-world robots.

Key insights

  • Coordinated dual-arm push and 6-DoF grasp actions enable far greater efficiency and success in dense, occluded clutter than single-arm methods, mimicking human dexterity.
  • A unified CNN-based DRL model, guided by an adaptive fuzzy reward function, can effectively learn complex push-grasp synergy directly from visual inputs without needing separate networks for each action.
  • Robustness and generalization, allowing seamless sim-to-real transfer without fine-tuning, are achieved by leveraging pre-trained visual backbones and an efficient learning framework.

Summary by alphaXiv.

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