Minimizing Human-Exoskeleton Interaction Force through PILCO: A Probabilistic Learning Control Approach

This research aims to enhance the interaction between lower limb exoskeleton robots and human users, focusing on the swing phase of normal gait. The primary challenge is modeling the dynamics of the system and interactions, considering uncertainties and measurement noise. A control approach “Probabilistic Inference for Learning Control” (PILCO) initially introduced in a different context, has shown promise in optimizing control policies efficiently. To address the aforementioned challenge, this study adopts the principles of PILCO to the domain of lower limb exoskeletons. Within this framework, Gaussian processes (GPs) will serve as the foundation for extracting probabilistic models from data. These models will be integrated with reinforcement learning techniques to facilitate optimal control by minimization of interaction forces. Reducing interaction forces, not only enhances user comfort but also enables the robot to accurately follow the user’s movements. Our results show that the proposed control framework predicted the dynamic model of the robot, taking into account the measurement noise and disturbances, with an acceptable error. The root mean square error (RMSE) of identified model through GPs, respectively for hip, knee and ankle joints were 0.9864, 0.9027, and 1.4613. Meanwhile the proposed algorithm effectively converges the interaction disturbance to near-zero levels, which is considerable in comparison with other deterministic model-based control.