RoboBallet: Planning for multirobot reaching with graph neural networks and reinforcement learning.

Area of Science:

• Robotics
• Artificial Intelligence
• Operations Research

Background:

• Coordinating multiple robots in shared, obstacle-rich manufacturing workspaces for complex tasks is computationally challenging for traditional methods.
• Current industrial multi-robot systems often rely on manual, labor-intensive trajectory planning based on human expertise.
• Automated solutions are needed to overcome the intractability of joint task allocation, scheduling, and motion planning under spatiotemporal constraints.

Purpose of the Study:

• To develop a reinforcement learning (RL) framework for automated task and motion planning in multi-robot systems.
• To address the limitations of classical methods in handling complex, real-world robotic manufacturing scenarios.
• To enable efficient, collision-free coordination of multiple robots performing tasks in shared workspaces.

Main Methods:

• A reinforcement learning (RL) framework utilizing a graph neural network (GNN) policy was developed.
• The GNN policy was trained via RL on procedurally generated environments with diverse configurations.
• The approach uses a graph representation of scenes and a graph policy neural network to jointly solve task allocation, scheduling, and motion planning.

Main Results:

• The RL framework successfully achieved automated task and motion planning for eight robots performing 40 reaching tasks in an obstacle-rich environment.
• The trained policy demonstrated zero-shot generalization to unseen environments with varying robot placements, obstacle geometries, and task poses.
• The high-speed capability of the planner was shown to improve workcell layout optimization and enable fault-tolerant and online replanning.

Conclusions:

• The proposed RL framework offers a scalable and efficient solution for complex multi-robot coordination problems in manufacturing.
• This approach automates critical planning tasks, reducing reliance on manual intervention and improving efficiency.
• The framework's generalization capabilities and speed open new possibilities for dynamic and adaptive robotic systems.