DeepMind’s robotic ballet: An AI for coordinating manufacturing robots

In today's world, robots play a crucial role in manufacturing, operating with precision as they move along conveyor belts in a well-orchestrated ballet of synchronized actions. Traditionally, programming these movements has been a labor-intensive task, often consuming hundreds to thousands of hours. Enter RoboBallet, an innovative AI system developed by Google’s DeepMind team, designed to empower manufacturing robots to autonomously determine their tasks. The challenge of directing these robots efficiently involves complex task allocation and scheduling—essentially deciding which robot should perform which task and in what sequence. This intricate process is akin to a more complex version of the famous traveling salesman problem. Additionally, ensuring that multiple robotic arms can operate without colliding with each other or surrounding equipment adds another layer of complexity. The team at DeepMind is addressing this trifecta of computational challenges: task allocation, scheduling, and motion planning. Matthew Lai, a research engineer at DeepMind, explains, “While there are existing tools for automating motion planning, task allocation and scheduling still require manual intervention. Our goal was to integrate all three aspects into a cohesive solution.” The team began by simulating environments known as work cells, where groups of robots execute their tasks on a designated product. In these work cells, a workpiece—specifically, a construction made of aluminum struts—was positioned on a table surrounded by up to eight Franka Panda robotic arms, each capable of seven degrees of freedom. These robots were tasked with completing as many as 40 different operations on the workpiece, with precision requirements that demanded the robotic arm's end effector to be within 2.5 centimeters of the target spot on a strut, approaching from the correct angle and holding steady for a brief moment to simulate actual work being done.