Poramate Manoonpong, Florentin Wörgötter, and Ulrich Parlitz

How to generate smooth biological motion by the interplay of biomechanics and neuronal control is little understood. Successful bio-mimetic models have been implemented by us [Manoonpong et al. 2007, Steingrube et al. 2010], but remain limited using specific network design paired with targeted – and, thus, somewhat restricted – plasticity. Goal of this project is to achieve more flexible, transferable solutions and a better understanding of the general control & plasticity principles in neural motor function. For this we will combine novel biomechanical principles with improved learning mechanisms in the control networks. Specifically we will for the first time combine three important aspects: synaptic plasticity, forward control (efferent copies [von Holst et al. 1950]), and predictive motor models (motor planning [Matheson 2008, Kawato 1999]). Preliminary results on combining learning with forward models exist [Schröder-Schetelig et al. 2010]. Such a three-fold combination will allow us to create a system that can react proactively (in an anticipatory manner) to upcoming events, like a terrain change. Our main substrates are R1: two- and R2: six-legged walking robots. Specifically we will perform the following steps:

i) Improve the biomechanics of our walking robots (R1 and R2) by integrating elastic and passive joint properties to imitate muscle properties for better self-stabilization through mechanical feedback, called “preflexes” [Full et al. 1999]. Only the combination of appropriate biomechanics with advanced control will assure robust and smooth locomotion of the robots.

ii) Develop improved adaptive neural mechanisms for R1 and R2:
(1) Learn internal models from efference copies (motor signals) for self-adaptation, and disturbance anticipation [Schröder-Schetelig et al. 2010] using near (on-leg sensors) and far (camera) sensor systems.
(2) Implement neural motor memory to enhance learning capability through experience and long-term memorization of walking relevant events (walking event memory).
(3) Implement neural models of motor planning based on functions realized partly in the posterior parietal cortex [for “reaching” see [Gail et al. 2006] and cooperation with D2] and motor program disinhibition similar to the function of the basal ganglia.

iii) These advanced methods will be compared to the demands in orthosis and prosthesis with Otto Bock Healthcare addressing the difficult question of sensor driven walking event recognition and the thereby arising prospective motor planning also for prosthetic devices.

Building on large prior expertise in this field [Manoonpong et al. 2007, Steingrube et al. 2010, Schröder-Schetelig et al. 2010], this study will be the first to create a walking system based on neural control that reacts in an anticipatory way, that can correct errors during walking, and remember the context in which an error occurred. These properties would be vital for advanced prosthetic and orthopaedic devices helping patients in a pro-active way.



^{Fig. 1: Left, Developed walking robots. Middle: Basic modular neural control for locomotion of the robots where adaptive neural mechanisms can be integrated [Manoonpong et al. 2010]. Right: E-MAG Active orthosis and C-leg by Otto Bock Healthcare.}