Key result
Optimization of the compact model using reinforcement learning reduced the modeling error and improved the prediction of the compliant humanoid robot's behavior during walking.
Reinforcement learning can effectively optimize compact dynamic models of compliant humanoid robots to improve trajectory prediction accuracy without increasing model complexity.
May advance compliant robot modeling; leaves open translation to cardiovascular assistive devices.
COMAN is a compliant humanoid robot. The introduction of passive compliance in some of its joints affects the dynamics of the whole system. Unlike traditional stiff robots, there is a deflection of the joint angle with respect to the desired one whenever an external torque is applied. Following a bottom up approach, the dynamic equations of the joints are defined first. Then, a new model which combines the inverted pendulum approach with a three-dimensional (Cartesian) compliant model at the level of the center of mass is proposed. This compact model is based on some assumptions that reduce the complexity but at the same time affect the precision. To address this problem, additional parameters are inserted in the model equation and an optimization procedure is performed using reinforcement learning. The optimized model is experimentally validated on the COMAN robot using several ZMP-based walking gaits.
No takes yet. Share an insight, caveat, or question.
Colasanto et al. (2012) studied this question. Reinforcement learning optimization of a compact model vs. Original unoptimized model was evaluated on Mean squared error of the model prediction with respect to actual response trajectories. Optimization of the compact model using reinforcement learning reduced the modeling error and improved the prediction of the compliant humanoid robot's behavior during walking.
Synapse has enriched 5 closely related papers on similar clinical questions. Consider them for comparative context: