Learning Ankle Exoskeleton Control for Reactive Standing Balance Through Neuromusculoskeletal Simulation

TL;DR An ankle-exoskeleton balance controller is trained with RL against a muscle-driven virtual human built from about 2 minutes of one person's motion capture, then run on hardware without tuning. In five new subjects it raised the treadmill perturbation they could withstand without stepping by 11.8% and cut peak center-of-mass velocity by 28.2%.
What they do
Ankle exoskeletons could help people recover balance after a push or slip, but tuning when and how hard to assist normally takes many perturbation trials on people. The authors move that search into simulation. A muscle-driven model learns to imitate one person’s recorded balance responses, and an RL exoskeleton controller is then trained against this “virtual human” and deployed on a physical bilateral ankle exoskeleton.
How it works
- Virtual human. A 2D musculoskeletal model (10 DoF, 18 muscles) is trained with SAC to track 24 motion-capture trials of backward treadmill perturbations from one subject, with DeepMimic-style rewards plus joint-moment tracking.
- Exoskeleton policy. With the human policy frozen, a second SAC branch sees only 200 ms of noisy shank and foot gyroscope history and outputs one torque for both ankles, capped at 16 Nm. Its reward penalizes stepping, center-of-mass (CoM) displacement and velocity, and torque; no timing or perturbation detection is prescribed.
- Hardware. Two brushless motors, shank and foot IMUs and an onboard Jetson Orin Nano; 1.1 kg per side plus a 1.2 kg belt.
Results
- The virtual human tracked joint angles with 0.6 ± 0.2° RMSE and moments with 0.07 ± 0.04 Nm/kg RMSE.
- In five healthy subjects not used for training, the balance threshold rose 11.8% (2.8 ± 1.7 cm) with assistance. Peak forward CoM displacement fell 13.5 ± 8.3% and peak CoM velocity 28.2 ± 5.3%.
- Assistance began 87 ± 12 ms after perturbation onset with no explicit cue. The model predicted biological ankle-moment offloading with R² = 0.81.
- Simulation underestimated the benefit (4% threshold gain vs. 11.8% on hardware). Gains were larger for ankle-dominant and lighter participants.
Limitations
The virtual human comes from a single subject and only successful trials, the frozen human policy cannot adapt to the device, and the device model omits cuff compliance and sensor coupling. Testing covered five healthy adults and backward perturbations only.
Figures



Abstract
Deficits in reactive balance can increase fall risk among older adults and individuals with neuromuscular impairments, motivating the need for interventions that improve balance recovery. Wearable robotic exoskeletons offer a promising approach to addressing this need by augmenting balance control. However, developing controllers that deliver appropriately timed and scaled assistance remains challenging, as evaluating each candidate strategy requires repeated exposure to destabilizing perturbations. We propose a simulation-to-real framework that combines motion imitation and reinforcement learning to develop balance-assistive ankle exoskeleton control. A muscle-driven neuromusculoskeletal model was trained using a small set of motion-capture data from a single subject to reproduce kinematic and kinetic responses to backward-translation perturbations. This model served as a virtual human for training a reinforcement-learning exoskeleton controller. The resulting controller generated bilateral ankle assistance using only inertial sensor measurements, without explicit perturbation detection. We implemented the controller on-board a bilateral ankle exoskeleton and evaluated it in five healthy subjects whose data were not used for training. Assistance increased the perturbation magnitude that subjects could withstand without stepping by 11.8% and reduced peak center-of-mass displacement and velocity by 13.5% and 28.2%, respectively. Individual benefits varied with baseline recovery strategy and body mass. These results indicate that a single virtual human can yield transferable balance assistance, and that more diverse virtual humans could further improve its effectiveness.
Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.
