An ankle exoskeleton controller learns balance on a virtual human first: five untrained wearers withstand 11.8% larger backward perturbations
Synopsis
Balance-assistive exoskeleton controllers usually apply a predefined torque profile that does not adapt to the wearer's real-time response, and evaluating each candidate strategy requires repeated destabilizing perturbations. This approach first trains a muscle-driven neuromusculoskeletal virtual human on about 2 minutes of motion-capture data, then trains a reinforcement-learning controller that outputs bilateral ankle assistance from shank and foot angular-velocity sensors alone. On the physical exoskeleton, five healthy participants whose data were not used for training withstood 11.8% larger backward-translation perturbations without stepping, with peak center-of-mass displacement and velocity reduced by 13.5% and 28.2%.
Interpretation
A balance-assistive ankle exoskeleton controller can now be learned in neuromusculoskeletal simulation and deployed on the physical device without subject-specific tuning. Prior balance-assistive controllers mostly applied a predefined torque profile during perturbation, so assistance did not adapt to the user's real-time balance response. Earlier approaches used predefined torque profiles, such as brief plantarflexion torque pulses during standing perturbations or spline-based hip torque profiles during walking perturbations; these showed assistive torque helps, but the assistance was not adaptive to the user's real-time response. In five healthy participants whose data were not used for training, assistance raised the perturbation magnitude tolerated without stepping by 11.8% on average, a threshold increase of 2.8±1.7 cm.
The controller reads only a history of sagittal angular velocities from both shanks and feet and outputs one normalized torque command shared by both ankles, with no explicit perturbation detection. The method chains motion imitation and reinforcement learning in two stages: first train the virtual human to reproduce kinematic and kinetic responses to backward perturbations, then freeze it and train the exoskeleton branch. The virtual human was trained on 24 motion-capture trials of 5 s each from a single subject, tracking joint angles with a root mean square error of 0.6±0.2 degrees and joint moments with 0.07±0.04 Nm/kg.
In simulation the virtual human showed only a 4% threshold gain, below the 11.8% on hardware, and on hardware assistance began earlier and reached a slightly higher peak torque. Hardware gyroscope signals showed an earlier rise in foot angular velocity and a transient increase in shank angular velocity during torque delivery, pointing to sensor placement and mechanical coupling through the compliant cuffs affecting controller inputs. These differences did not prevent the controller from providing sufficient assistance, but they show that simulation-based control behavior and performance depend on the fidelity of the virtual human and the device model.
Perspective
The demonstrated setting is five healthy adults, backward-translation perturbations only, standing with arms crossed and no stepping, with the controller running in real time on a bilateral ankle exoskeleton and no subject-specific tuning.
The virtual human was built from one subject and only from successful trials up to the unassisted balance threshold, so it did not cover the anthropometric differences, recovery strategies and near-failure responses seen in hardware testing; the frozen human policy can respond to assistance but not adapt to it, and the device model omits cuff compliance and sensor coupling. Individual threshold gains were positively associated with ankle-dominant baseline recovery and inversely associated with body mass, and the authors propose body-mass-normalized torque limits as a direct test.
