Robots learn to work together like a well-choreographed dance

Scientists at UCL, Google DeepMind and Intrinsic have developed a powerful new AI algorithm that enables large sets of robotic arms to work together faster and smarter in busy industrial settings, potentially saving manufacturers hundreds of hours of planning time and unlocking new levels of flexibility and efficiency.

The system, called RoboBallet, has been designed to help teams of automated robots that work in shared, obstacle-filled spaces like assembly lines and factory floors, to plan their movements and tasks automatically – without colliding with each other or the surrounding environment.

This is a challenge that has long plagued manufacturers; the job is currently done manually by specially trained human programmers. It is a very tedious and error-prone process, that takes hundreds of hours for each set of tasks.

As described in a research paper in Science Robotics, RoboBallet trains a graph neural network-based robot brain using reinforcement learning (RL). In a RL framework, the robot brain learns by trial and error and is given a ‘reward’ when tasks are completed, with higher rewards for having completed them faster.  

The graph neural network is a neural network architecture that works natively with data in a graph form. Its use enables robots to understand and reason about their surroundings (treating each obstacle like a point in a network – in an organised manner) so they can work out the most effective way to work together. Both graph neural networks and reinforcement learning are AI techniques. 

In the research, after just a few days of training, RoboBallet was able to generate high-quality plans in just seconds – even for complex layouts it had never seen before, solving up to 40 tasks with eight robotic arms - far beyond the capabilities of previous systems.

Lead author Matthew Lai, a PhD researcher at UCL Computer Science and Google DeepMind, said: “RoboBallet transforms industrial robotics into a choreographed dance, where each arm moves with precision, purpose, and awareness of its teammates. It’s not just about avoiding crashes; it’s about achieving harmony at scale.

“For the first time, we can automate complex multi-robot planning with the grace and speed of a dance, making factories more adaptive, efficient, and intelligent.”

RoboBallet is able to plan robot movements hundreds of times faster than real-time. Researchers say this could allow factories to adapt instantly if a robot fails or if the layout changes. RobotBallet also enables layout optimisation, helping manufacturers decide where to place robots for maximum efficiency and throughput.

Researchers say the system’s scalability is a major breakthrough. Traditional planning algorithms struggle to handle more than a few robots due to the exponential growth in complexity. RoboBallet’s graph-based architecture allows it to learn general principles of coordination, rather than memorising specific scenarios, making it suitable for large-scale industrial use.

Co-author Associate Professor Alex Li from UCL Computer Science said: “In today’s factories, coordinating multiple robotic arms is like solving a moving 3D puzzle, every action must be perfectly timed and placed to avoid collisions. Right now, this planning takes specialists hundreds of hours and is costly to design manually.

“The name, RoboBallet, captures the elegance and what we can do with so many robots. Just as ballet dancers move in perfect harmony with each other, our robots can now coordinate their movements with a superhuman level of precision and grace – RoboBallet could instantly generate plans for brand-new layouts at large scales and speeds that are impossible for specialists to handcraft.”

What can this be used for?

As manufacturing continues to evolve toward more flexible and adaptive production, this technology could be used in car manufacturing, as well as electronics assembly or even building houses with robots. It’s especially useful in places where robots need to work closely together without getting in each other’s way.

What’s next?

While the current RoboBallet version focuses on reaching tasks, where a robot moves its arm to a specific point for tasks such as welding, researchers say it could be extended to more complex operations like pick-and-place, or painting. The researchers also envision future versions that handle task dependencies, heterogeneous robot teams, and more sophisticated obstacle geometries.

Limitations

The team acknowledges that RoboBallet doesn’t yet handle every possible factory scenario. For example, it doesn’t currently account for tasks that must be done in a specific order, or robots with different capabilities. But they believe these features can be added in future versions, and the system’s flexible architecture makes it well-suited to such enhancements.

The work was funded by Google DeepMind and Intrinsic, and the codebase has been open-sourced, allowing other researchers to continue building on it and accelerating the entire field forward.

The team has produced a video to explain their work and RoboBallet: here:  https://www.youtube.com/watch?v=uqw7hTlk_BQ