Multi-Agent Deep Reinforcement Learning for Collision-Free Posture Control of Multi-Manipulators in Shared Workspaces

In multi-manipulator systems operating within shared workspaces, achieving collision-free posture control is challenging due to high degrees of freedom and complex inter-manipulator interactions. Traditional motion planning methods often struggle with scalability and computational efficiency in such settings, motivating the need for learning-based approaches. This paper presents a multi-agent deep reinforcement learning (MADRL) framework for real-time collision-free posture control of multiple manipulators. The proposed method employs a line-segment representation of manipulator links to enable efficient interlink distance computation to guide cooperative collision avoidance. Employing a centralized training and decentralized execution (CTDE) framework, the approach leverages global state information during training, while enabling each manipulator to rely on local observations for real-time collision-free trajectory planning. By integrating efficient state representation with a scalable training paradigm, the proposed framework provides a principled foundation for addressing coordination challenges in dense industrial workspaces. The approach is implemented and validated in NVIDIA Isaac Sim across various overlapping workspace scenarios. Compared to conventional state representations, the proposed method achieves faster learning convergence and superior computational efficiency. In pick-and-place tasks, collaborative multi-manipulator control reduces task completion time by over 50% compared to single-manipulator operation, while maintaining high success rates (>83%) under dense workspace conditions. These results confirm the effectiveness and scalability of the proposed framework for real-time, collision-free multi-manipulator control.