Evolving Controller Symmetry for Multilegged Robots

Vinod Valsalam
Risto Miikkulainen

The videos on this website demonstrate the walking behaviors evolved by five neuroevolution methods, which differ only in the way they determine controller symmetry: (1) Evolving symmetry systematically using ENSO, (2) evolving symmetry randomly without using the group-theory mechanisms of ENSO, (3) using fixed S4 symmetry during evolution (i.e. maximal symmetry), (4) using fixed D2 symmetry during evolution, and (5) using direct encoding without symmetry constraints (which is equivalent to using the fixed trivial symmetry during evolution). The gaits in these videos demonstrate how ENSO outperforms the other methods.