Self-organized robotic locomotion by closing the propriosensory feedback loop*

From a dynamical systems viewpoint, robotic locomotion corresponds to attractor trailing in the phase space spanned by the variables describing the state of the controller, the robot's body, and the environment it is placed into. For top-down control the dynamics of the body is driven by a controller signal, allowing for the differentiation between the master and slave subsystems. However, when propriosensory input is also integrated into the controller scheme, self-organized motion patterns emerge due to the local feedback mechanisms.

We investigate how motion primitives are generated for wheeled and hexapod-type robots using a simple neural controller with propriosensory feedback. Each actuator is controlled by rate-encoding neurons with internal adaption, receiving in turn inputs via self-coupling and from proprioceptive feedback. We demonstrate that different hexapod gait patterns can be selected when changing the weight matrix of neural couplings. For wheeled robots, on the other hand, coexisting attractors emerge via the somatosensory loop even without central pattern generators, a control type termed here attractoring. Multistability in this case allows for autonomous direction reversal when colliding with obstacles.