We describe an evolutionary approach to the control problem of bipedal walking. Using a full rigid-body simulation of a biped, it was possible to evolve recurrent neural networks that controlled stable straight-line walking on a planar surface. No proprioceptive information was necessary to achieve this task. Furthermore, simple sensory input to locate a sound source was integrated to achieve directional walking. To our knowledge, this is the first work that demonstrates the application of evolutionary optimization to three-dimensional physically simulated biped locomotion. Index TermsBipedal walking, evolutionary algorithms, evolutionary robotics, physics, recurrent neural networks.
Originality: Seminal work; first successful application of evolutionary robotics methodology to realistic 3D bipedal walking. Rigour: First use of realistic rigid-body physics engine simulation in this domain. Sufficient runs undertaken to produce statistically significant results, providing very firm evidence of the applicability of the methodology. Significance: Extended evolutionary robotics techniques to control an inherently highly unstable system and introduced a new way of generating realistic (human-like) bipedal locomotion. Impact: Influenced significant amount of subsequent work in various countries and was developed into a leading commercial tool now used widely in animation and special effects (NaturalMotion.com). 36 Google scholar citations, 6 Web of Science.