Incremental embodied chaotic exploration of self-organized motor behaviors with proprioceptor adaptation
abstract
This paper presents a general and fully dynamic embodied artificial neural system, which incrementally explores and learns motor behaviors through an integrated combination of chaotic search and reflex learning. The former uses adaptive bifurcation to exploit the intrinsic chaotic dynamics arising from neuro-body-environment interactions, while the latter is based around proprioceptor adaptation. The overall iterative search process formed from this combination is shown to have a close relationship to evolutionary methods. The architecture developed here allows realtime goal-directed exploration and learning of the possible motor patterns (e.g., for locomotion) of embodied systems of arbitrary morphology. Examples of its successful application to a simple biomechanical model, a simulated swimming robot, and a simulated quadruped robot are given. The tractability of the biomechanical systems allows detailed analysis of the overall dynamics of the search process. This analysis sheds light on the strong parallels with evolutionary search.
2.1. Chaos in the Nervous System
There is a growing body of observations of intrinsic chaotic dynamics in the nervous system (Guevara et al., 1983; Rapp et al., 1985; Freeman and Viana Di Prisco, 1986; Wright and Liley, 1996; Terman and Rubin, 2007). Some studies indicate such dynamics in animal motor behaviors at both the neural level (Rapp et al., 1985; Terman and Rubin, 2007) and at the level of body and limb movement (Riley and Turvey, 2002). These seem particularly prevalent during developmental and learning phases (e.g., when learning to coordinate limbs) (Ohgi et al., 2008). The existence of such dynamics in both normal and pathological brain states, at both global and microscopic scales (Wright and Liley, 1996), and in a variety of animals, supports the idea that chaos plays a fundamental role in neural mechanisms (Skarda and Freeman, 1987; Kuniyoshi and Sangawa, 2006). Although the functional roles of chaotic dynamics in the nervous system are far from understood, a number of intriguing proposals have been put forward. Freeman and colleagues have hypothesized that chaotic background states in the rabbit olfactory system provide the system with continued open-endedness and readiness to respond to completely novel as well as familiar input, without the requirements for an exhaustive memory search (Skarda and Freeman, 1987). Kuniyoshi and Sangawa (2006) made the important suggestion that chaotic dynamics underpin crucial periods in animal development when brain-body-environment dynamics are explored in a spontaneous way as part of the process of acquiring motor skills. Recent robotics studies have demonstrated that chaotic neural networks can indeed power the self-exploration of brain-body-environment dynamics in an embodied system, discovering stable patterns that can be incorporated into motor behaviors (Kuniyoshi and Suzuki, 2004; Kinjo et al., 2008; Pitti et al., 2010). As indicated in the previous section, this work has been fundamentally extended to allow goal-directed (fitness-directed) search (Shim and Husbands, 2010; Shim and Husbands, 2012).
Incremental embodied chaotic exploration of self-organized motor behaviors with proprioceptor adaptation
abstract
This paper presents a general and fully dynamic embodied artificial neural system, which incrementally explores and learns motor behaviors through an integrated combination of chaotic search and reflex learning. The former uses adaptive bifurcation to exploit the intrinsic chaotic dynamics arising from neuro-body-environment interactions, while the latter is based around proprioceptor adaptation. The overall iterative search process formed from this combination is shown to have a close relationship to evolutionary methods. The architecture developed here allows realtime goal-directed exploration and learning of the possible motor patterns (e.g., for locomotion) of embodied systems of arbitrary morphology. Examples of its successful application to a simple biomechanical model, a simulated swimming robot, and a simulated quadruped robot are given. The tractability of the biomechanical systems allows detailed analysis of the overall dynamics of the search process. This analysis sheds light on the strong parallels with evolutionary search.
authos
links