A hip exoskeleton lowers walking energy versus a passive device in a small test
Multi-Agent Reinforcement Learning of Co-Adaptive Hip Exoskeleton Control (MARL-Exo) for Metabolic Energy Reduction · arXiv:2610.37441
In plain words
A hip exoskeleton can change a person’s walking, while the person’s changing gait also affects the device’s controller. This paper introduces MARL-Exo, which first trains a simulated person to walk, then trains that person and a hip-device controller together. The controller’s reward includes an estimate of hip-muscle energy use.
In a treadmill study, active assistance lowered measured net energy use for all eight participants compared with wearing the same device with zero commanded torque. The study tested only eight healthy adults at one speed.
The research
When a powered hip exoskeleton changes someone’s gait, that person’s response also changes what the controller must handle. How can training account for this interaction while guiding assistance by the wearer’s energy use?
The authors address both challenges with a simulated human model and an exoskeleton controller that learn together.
MARL-Exo trains a simulated walker and hip-exoskeleton controller together, using modeled hip-muscle energy to guide assistance before testing it with people.
MARL-Exo first trains a simulated human model to walk, then co-adapts it with a hip-exoskeleton controller whose reward includes modeled hip-muscle energy.
The authors compare two reward variants in simulation and test the deployed controller in eight adults, asking whether assistance reduces measured energy use.
With the same device worn, active assistance lowered net metabolic rate for all eight participants compared with passive mode. This supports a benefit from assistance beyond simply wearing the device.
Against walking without a device, the cohort mean reduction was variable and not statistically significant after correction.
The hardware result comes from eight healthy adults walking at one treadmill speed after brief familiarization.
The comparison with walking without a device was variable and not statistically significant after correction, so the study does not establish a benefit across other people or walking conditions.
What the evidence shows
- Train both sides
MARL-Exo is a two-stage framework that trains a simulated full-body human model and a hip-exoskeleton controller together, then deploys the controller to assist walking. The question is how to train assistance when the device changes the wearer’s gait and the wearer’s response changes what the controller encounters.
The authors also address a gap they identify in using a physiological estimate of muscle energy directly in the controller’s reward.
- Build in energy use
The framework first trains the simulated person to walk without a device, then introduces the exoskeleton and trains both controllers together. The device controller uses delayed hip movement measurements and recent commands, and its reward includes modeled energy use by hip muscles.
The deployed Metabolic controller was trained further from a converged Metabolic + Power controller after removing its explicit mechanical-power reward term; both retained the metabolic term and other listed reward terms.
- Compare reward choices
In one deterministic simulation episode, both reward variants reduced modeled energy across the analyzed muscles compared with walking without an exoskeleton. The reduction was 13.8% with Metabolic and 8.8% with Metabolic + Power.
This comparison tests the reward refinement in the reported training setup: Metabolic began from the other controller’s converged checkpoint, and their reported training-step counts differed. The values cover a partial muscle set, not whole-body human energy use.
- Measure assistance in people
The authors deployed the Metabolic controller without per-person tuning and measured net metabolic rate in eight healthy adults walking on a treadmill. Compared with wearing the same device in passive mode, active assistance lowered net metabolic rate for every participant, with a mean reduction of 18.3%.
This comparison isolates the benefit of assistance while the hardware is worn. Compared with walking without a device, the mean reduction was smaller and variable, and did not reach statistical significance after correction.
- Bound the conclusion
Together, the results show that a controller trained with a modeled muscle-energy reward can lower measured walking cost relative to the same device without active assistance in this small test. The no-device comparison also shows why the hardware matters: carrying it passively raised energy use, and the net result with assistance varied.
The human evidence is limited to eight healthy adults, one treadmill speed, and a short familiarization period.
Deep
- Original paper authors
- Yifei Yuan, Jakob Wolf, Ghaith Androwis, Xianlian Zhou
- Submitted to arXiv
- License
- arXiv non-exclusive