Evolving Controllers for Physical Multilegged Robots

Vinod Valsalam
Risto Miikkulainen




Evolved Controllers

Simulation


Trot gait evolved in simulation. This robot model approximates the morphology and dynamics of the target physical robot with sufficient accuracy to simulate it realistically. ENSO utilizes the resulting simulation to evolve effective controllers that transfer well to the physical robot.

Physical Robot


Physical robot reproduces the same trot gait evolved in simulation. It walks smoothly in a straight line even on very different surfaces such as linoleum and carpet, demonstrating that ENSO can evolve such robust controllers that transfer successfully from simulation to real robots.

Motor Speed Reduced by 10%


Behavior of the evolved controller when the maximum motor speed is reduced by 10%. Motor speed can reduce in this manner as a result of e.g. low battery charge or temperature. In contrast to the handcoded controller, the evolved controller generalizes well and continues to function robustly.

Motor Speed Reduced by 60%


Behavior of the evolved controller when the maximum motor speed is reduced by 60%. Even for such severe change in motor performance, the evolved controller remains robust by slowing down the legs automatically and keeping them synchronized. In contrast, the hand-designed controller fails for as less as 10% reduction in maximum motor speed.

Left Rear Leg Initialized with Max Error


Behavior of the evolved controller when the left rear leg is initialized with maximum error. This experiment simulates the leg getting obstructed by an obstacle. In contrast to the handcoded controller, the evolved controller takes only about two seconds to correct the same error by adjusting the behavior of all legs simultaneously.

Left Rear Leg Disabled


Behavior of the evolved controller when the left rear leg is disabled to simulate leg failure. Since the controller was evolved with all four legs, it produces the same trot gait as before. However, the asymmetric action of the functional and disabled legs causes the robot to curve to one side.

Controller Evolved with Left Rear Leg Disabled


Gait evolved with one leg disabled to simulate leg failure. Since this controller was evolved with only three functional legs, evolution adapted the gait accordingly to produce a straight walk utilizing only those three legs. Thus, this experiment demonstrates that ENSO can be utilized to design controllers for fault tolerance.