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

Area of Science:

• Robotics
• Artificial Intelligence
• Control Systems

Background:

• Collision-free posture control for multi-manipulator systems in shared workspaces is complex due to high degrees of freedom and inter-manipulator interactions.
• Traditional motion planning methods lack scalability and efficiency for these demanding applications.
• Learning-based approaches are needed to address these challenges.

Purpose of the Study:

• To present a multi-agent deep reinforcement learning (MADRL) framework for real-time collision-free posture control of multiple manipulators.
• To develop an efficient state representation for cooperative collision avoidance.
• To enable scalable training and decentralized execution for real-time trajectory planning.

Main Methods:

• A multi-agent deep reinforcement learning (MADRL) framework is proposed.
• A line-segment representation of manipulator links is used for efficient interlink distance computation.
• A centralized training with decentralized execution (CTDE) paradigm is employed, leveraging global state during training and local observations during execution.

Main Results:

• The proposed method demonstrates faster learning convergence and superior computational efficiency compared to conventional state representations.
• In pick-and-place tasks, collaborative control reduced task completion time by over 50% versus single-manipulator operation.
• High success rates (>83%) were maintained under dense workspace conditions.

Conclusions:

• The framework provides a scalable solution for real-time, collision-free multi-manipulator control in dense industrial environments.
• The integration of efficient state representation and scalable training paradigms addresses key coordination challenges.
• The validated approach offers a principled foundation for advanced robotic coordination.