AI Optimizes Industrial Robot Collaboration

Researchers at Google DeepMind, Intrinsic, and UCL have developed an AI capable of coordinating multiple industrial robots safely and efficiently. This approach represents an early step toward replacing the complex and time-consuming process of manual robot programming.

Researchers from Google DeepMind Robotics, robotics company Intrinsic, and University College London (UCL) have introduced an AI method called 'RoboBallet' to address the issue of manual robot programming. According to Alphabet-owned Intrinsic, programming the world's 4.3 million industrial robots currently takes over 100 million hours, a figure that resets whenever tasks or layouts change. The research, published in Science Robotics, aims to enable fully automated, collision-free coordination for multiple robots.

The team's approach utilizes a graph neural network (GNN) trained with reinforcement learning. Robots, tasks, and obstacles are represented as nodes in a graph, allowing the system to model complex relationships between them. By practicing in millions of simulated scenarios, the AI learns to find optimized, collision-free paths for the robots. The researchers state that it only requires CAD files and a rough task description to begin.

In lab tests, RoboBallet outperformed traditional methods and expert-designed solutions by approximately 25 percent. According to Intrinsic, the efficiency gains scale with the number of robots: when increasing from four to eight robots, the average task completion time fell by 60 percent.

Torsten Kroeger, Chief Science Officer at Intrinsic, describes the technology as a critical step toward adaptive, highly efficient planning in manufacturing. The goal is for humans to define only high-level tasks, while the system determines the optimal sequence, assigns actions, and generates collision-free paths for every robot. Paired with AI-driven perception, this could eventually enable real-time re-planning.

The system has not yet been tested on a real production line. Current limitations include an inability to handle robots with different capabilities or tasks that require a strict order.