In an MIT experiment, four AI agents designed 12,774 robots for an obstacle course and ran 16,427 virtual tests, iXBT reports. The researchers built 135 of the designs that passed simulation, and only 66% of those passed the physical tests at every difficulty level. The course has 10 obstacles, each with 3 difficulty levels.

The system, called RoboBuilder, lets the agents design the physical structure of a robot as well as its software. They choose parts from a ready-made set, pick motors, create extra parts for 3D printing and write the programs that run the motors. The agents checked designs in a physics simulation, changed the geometry and settings, and sent the designs that passed to be built.

The main problem was the move from simulation to reality. Tyres slipped, joints came apart, gears disengaged and the load shifted on slopes. Training the simulator on real measurements and video did not remove the gap, and the 66% figure comes after that training.

Researchers gave the agents a feedback loop: after a failed attempt the agent received video of it, analysed it and could change the design or the simulator settings. On the first two difficulty levels, real-world success rose from 73% to 93% after this loop. The agents also came up with mechanisms that the automatic search algorithms in the experiment did not find, among them wheels with protrusions, lever systems, cable-climbing mechanisms and walking systems. On hard obstacles, and when asked to make one robot that clears the whole course, the AI did much worse.

Source: iXBT.com

Background

MIT

University

MIT is a private research university in Cambridge, Massachusetts, whose labs work on robots that understand what people mean by their instructions and on soft robots for medicine.

RoboBuilder

AI robot design benchmark

Benchmark from the Massachusetts Institute of Technology in which AI coding agents design a robot from standard parts, motors and one 3D-printed part.