Evolving Controller Symmetry for Multilegged Robots

Vinod Valsalam
Risto Miikkulainen




Generalization on the Incline

ENSO


Generalization of controllers evolved by ENSO on inclined ground. Previously evolved controllers were tested for generalization by reducing ground friction, thus making the ground slippery. The robots still maintain their balance and orientation as they walk across the incline, slipping only a little. Thus, controllers evolved by ENSO generalize well and remain effective even on slippery ground.

S4 Symmetry


Generalization of controllers with handdesigned S4 symmetry evolved on inclined ground. Making the ground slippery makes it harder for these robots to maintain orientation and balance because their gaits are ill-suited for inclines. As a result, they flip over or move downhill instead of walking across the incline along the yellow line, i.e. controllers with S4 symmetry generalize poorly on inclines.

D2 Symmetry


Generalization of controllers with handdesigned D2 symmetry evolved on inclined ground. As in the case of S4 symmetry, the robots flip over or move downhill instead of walking along the yellow line, i.e. controllers with D2 symmetry also generalize poorly on inclines.

Random Symmetry Evolution


Generalization of controllers produced by random symmetry evolution on inclined ground. The result is similar to that of controllers evolved with S4 and D2 symmetries, demonstrating poor generalization.