Robots that can adapt like animals

The Intelligent Trial and Error Algorithm introduced in the paper ‘Robots that can adapt like animals’ (Nature, 2015): the video shows two different robots that can adapt to a wide variety of injuries in under two minutes. A six-legged robot adapts to keep walking even if two of its legs are broken, and a robotic arm learns how to correctly place an object even with several broken motors.

Recovery with closed-loop control

In this video, the six-legged robot uses the Intelligent Trial and Error Algorithm with a closed-loop controller to recover from damages.

Robot learns to walk on uneven terrain

Evolution of locomotion controllers for a six-legged robot using a generative encoding of Single-unit pattern generators, and recovery following a change in the environment.

Generic fault-detection for large-scale robot swarms

Robots in the swarm learn to tolerate each other (colored green), while simultaneously detecting and forming a consensus on abnormally behaving robots as being faulty (colored red).

In this video, the normal behavior of the swarm is dispersion, and the fault injected in the abnormally behaving robot prevents it from changing direction.

Robots in the swarm continue to tolerate each other despite a priori unknown changes in their behavior; the entire swarm shifts from performing aggregation to dispersion half-way though the experiment.

Division of labor in robot swarms

In the first few generations, the robots can hardly avoid obstacles. After 240 generations of selection, the robot swarm is able to forage for both small and large food tokens and cooperatively push them towards the region of the arena under the white wall.