Multi-Agent Deep Reinforcement Learning for Variable-Finger Dexterous Grasping through Multi-Stream Embedding Fusion

Mahdi Bonyani, Maryam Soleymani, Chao Wang

Keywords: Variable-Finger Dexterous Grasping, Multi-Agent Deep Reinforcement Learning, Multi-Stream Embedding Fusion

Abstract:

Dexterous robotic hands offer unparalleled potential for high-precision, contact-rich manipulation, but their
control remains a formidable challenge due to high-dimensional action spaces and diverse object-hand interactions. In this paper, we propose a novel framework for dexterous grasping based on multi-agent deep reinforcement learning (MADRL) and multi-stream embedding fusion. Each component of the robotic hand, fingers, wrist, and arm, is modeled as an independent agent that learns cooperative control strategies guided
by multi-stream embedding fusion. By leveraging high-quality static grasp data from the MultiDex dataset as reference targets, our method eliminates the need for human demonstrations or generative sampling during training. Experimental results demonstrate that our method achieves stable, compliant, and generalizable grasps across diverse objects and hand configurations, outperforming traditional single-agent baselines.

Submission Number: 17