A Task-Invariant Learning Framework of Lower-Limb Exoskeletons for Assisting Human Locomotion

Kinematic control approaches for exoskeletons follow specified trajectories, which overly constrain individuals who have partial or full volitional control over their lower limbs. In our prior work, we proposed a general matching framework for underactuated energy shaping to provide task-invariant, energetic exoskeletal assistance. While the proposed shaping strategies demonstrated benefits such as reduced human torques during walking, it remains unclear how the parameters of these shaping strategies are related to different gait benefits. Meanwhile, research indicates that customizing assistance via online optimization can substantially improve exoskeleton's performance for each individual. Motivated by this fact, we com-bine derivative-free, sample-efficient optimization algorithms with our energy shaping strategies to propose a task-invariant learning framework for lower-limb exoskeletons. Through rapid online optimization, this framework enables exoskeletons to adjust shaping parameters for minimizing human joint torques across users and tasks. Simulation results show that shaping strategies with optimal parameters effectively reduce human joint torques and estimated metabolic cost during simulated walking. In addition, the optimal exoskeleton torques calculated using able-bodied subjects' kinematic data closely match the real human joint torques for different walking gaits.