Assistive Control of Knee Exoskeletons for Human Walking on Granular Terrains

TL;DR

This work addresses the challenge of human walking on granular terrains by integrating a stiffness-based model predictive control framework for knee exoskeletons with a learning-based, IMU-driven GRF estimator. The core idea is to estimate real-time ground reaction forces and human knee torque to drive an MPC that computes assistive knee torque, resulting in terrain-adaptive support. Key findings show that the MPC controller redistributes joint moments, reduces muscle activation, and yields a modest but meaningful reduction in metabolic cost on sand, though statistical significance for metabolism varied across conditions. The approach advances wearable robotics for real-world terrains and informs the design of energy-efficient, stabilizing assistive devices for rehabilitation and mobility augmentation.

Abstract

Human walkers traverse diverse environments and demonstrate different gait locomotion and energy cost on granular terrains compared to solid ground. We present a stiffness-based model predictive control approach of knee exoskeleton assistance on sand. The gait and locomotion comparison is first discussed for human walkers on sand and solid ground. A machine learning-based estimation scheme is then presented to predict the ground reaction forces (GRFs) for human walkers on different terrains in real time. Built on the estimated GRFs and human joint torques, a knee exoskeleton controller is designed to provide assistive torque through a model predictive stiffness control scheme. We conduct indoor and outdoor experiments to validate the modeling and control design and their performance. The experiments demonstrate the major muscle activation and metabolic reductions by respectively 15% and 3.7% under the assistive exoskeleton control of human walking on sand.

Figures (13)

• Figure 1: (a) Subject with IMUs and bilateral exoskeletons. (b) Schematic of the human lower limb and human-exoskeleton interaction during the swing (left leg) and stance (right leg) phases on sand.
• Figure 2: (a) Knee angle $\theta_k$ (top) and torque $\tau_h$ (bottom) profiles as function of gait phase $s$ of human walking on solid ground and sand. (b) The ground reaction forces on sand and solid ground: Longitudinal force $F_x$ (top) and vertical forces $F_z$ (bottom). (c) Knee joint stiffness curve during the walking gait. The arrows indicate the gait phase progressing direction. Heel strike (HS) and toe-off (TO) stand for heel-strike and toe-off events, respectively, and R and L for right and left legs, respectively.
• Figure 3: Schematic of the machine learning-based dual-pathway GRFs prediction scheme.
• Figure 4: The schematic of the overall exoskeleton control design.
• Figure 5: (a) Experimental devices and the interconnection among various wearable sensing, exoskeletons and embedded systems (top) and outdoor experimental setup (bottom). (b) Indoor experimental setup with sand and solid ground platform with motion capture system (not shown in the figure).