Evolved Recovery From Damage: We are working on an evolutionary robotics algorithm for automated recovery of functionality following unanticipated robot damage. The algorithm contains two related genetic algorithms: one that evolves controllers for the robot, and another that evolves hypotheses regarding damage suffered by a physical robot. The hypotheses are tested on a simulated robot that resembles the physical robot, and the fitness function is proportional to the difference between the observed behaviour of the simulated robot after applying the damage hypothesis, and the actual behaviour recorded from the physical robot.

The animation above shows the behaviour of the robot after a controller that translates sensory signals into motor commands has been evolved for it.

The video above documents the behaviour of the 'physical' robot immediately after it sustains some unanticipated failure, which in this case is the complete separation of one of its joints. (In the work so far the 'physical' robot is also simulated.)

The animation above shows the behaviour of the 'physical' robot after the evolutionary algorithm has terminated. The algorithm evolved the correct hypothesis (that a particular joint had been separated), and then evolved a compensatory neural network controller that allowed the quadrupedal robot to locomote with only three legs.




Evolved Recovery From Environmental Change: An updated version of our algorithm allows the evolutionary algorithm to automatically determine whether the robot has suffered some failure, or whether the robot has encountered a novel environment. This determination is based solely on sensory data returned by the 'physical' robot.

The robot with its original evolved controller.	Robot encounters an environment in which the floor is tilted horizontally by 20 degrees, causing it to tip over when using the original evolved controller.	The algorithm correctly diagnoses: (1) that an evironmental change has occurred, as opposed to some internal failure; and (2) that the nature of the environmental change is a horizontal tilting of the floor by 20 degrees. Based on this diagnosis the original evolutionary algorithm evolves a compensatory neural network. Note that the new gait produced by the new controller is dramatically different from the original gait.