UniExo: Unified Multi-Skill Policies for Musculoskeletal Locomotion and Co-Adaptive Exoskeleton Control

Yifei Yuan, Jakob Wolf, Ghaith Androwis, Xianlian Zhou

September 2026

TL;DR UniExo distills four musculoskeletal locomotion experts into one skill-latent human policy that tracks 94.7% of unseen clips, then co-trains a single hip exoskeleton controller with it in simulation. On hardware the controller assisted six people at four treadmill speeds and one person through walking, turning, running and backward walking without mode switching.

What they do

Most exoskeleton controllers are built for one activity and need explicit mode switching. UniExo first trains one muscle-driven human policy (MyoLeg, 80 muscles) that walks, turns, runs, walks backward and moves between these skills. It then co-trains a single hip exoskeleton controller with that policy in simulation and deploys it on a custom device.

How it works

Unified human policy. Four single-skill imitation experts, trained with PPO on AMASS and LAFAN1 clips, are distilled with DAgger into one network with a 32-D skill latent. PPO fine-tuning on clips stitched through a 1.5 s standing pose teaches the transitions.
Hip moment predictor. A small MLP maps a 300 ms window of hip angle and velocity to the hip moment, run per leg with shared weights. Randomized input kinematics cover differences between users and between simulation and the IMUs.
Co-adaptation. The predictor initializes the exoskeleton policy, which trains alongside the human policy with a shared critic. Rewards favor positive assistance power, offloading the muscles’ hip moment, lower hip muscle activation and smooth torque.
Hardware. A 4.23 kg bilateral hip exoskeleton with two motors, a Raspberry Pi 5 and three IMUs; torque capped at 15 N·m.

Results

The unified policy tracks 94.7% of unseen clips (experts 97.9%, distilled only 84.3%). Under 40 N·s pushes it stays upright in 68.8% of episodes, against 43.9% for the experts.
Co-adaptation cuts negative exoskeleton power by 65-81% and raises the positive-work fraction by 11.5-25.2 points across skills.
On a treadmill, six participants at 0.76-2.25 m/s received 96.2 ± 2.1% positive work without per-user tuning.
Overground, one participant walked, turned, ran and walked backward with smooth torque across transitions; the torque period shortened from 1.4 to 0.7 s within the first running stride.

Limitations

Metabolic cost was not measured, and the multi-skill route used one participant. The policy covers four skills on flat ground only. The simulated exoskeleton is rigidly attached and aligned with the hip, unlike the real device, and simulated rectus femoris activation peaks in swing whereas EMG peaks at loading.

Figures

From Figure 1: Overview of UniExo. (A) Four single-skill experts are distilled with DAgger into one policy whose enco... — From Figure 1: Overview of UniExo. (A) Four single-skill experts are distilled with DAgger into one policy whose encoder maps the reference motion to a skill latent and whose decoder maps the latent and proprioception to muscle activations; PPO fine-tuning on single-skill clips and stitched transitions gives the unified human policy. (B) A predictor learns the filtered hip moment from 300 ms of randomized hip kinematics, initializes the exoskeleton controller, and the two policies are co-adapted with PPO and a shared critic. (C) The controller runs on a Raspberry Pi 5 with three IMUs and drives two motors.

From Figure 2: Skill latent of the unified policy. Top: a sequence of walk, turn, run and backward motions. Left: the... — From Figure 2: Skill latent of the unified policy. Top: a sequence of walk, turn, run and backward motions. Left: the motion sequence in the discriminant projection, normalized onto a sphere. Right: one stitched walk–turn–run–backward sequence in the same view, colored by time.

From Figure 8: One participant on the laboratory route with UniExo. Left: hip angle, hip angular velocity, exoskeleto... — From Figure 8: One participant on the laboratory route with UniExo. Left: hip angle, hip angular velocity, exoskeleton torque and power of both legs over one repetition, with bands marking transitions. Right: the route and the order of motions.

Abstract

Daily locomotion encompasses diverse activities and frequent transitions between them, yet most exoskeleton controllers are designed for a single activity or a narrow set of related movements. Changes in activity therefore typically require explicit mode switching and separately tuned or retrained controllers. Simulation-based learning reduces the need for hardware-based tuning but generally retains this limitation. Here we present UniExo, a framework that first constructs a multi-skill musculoskeletal human policy and then jointly trains an exoskeleton control policy with it. Four single-skill imitation experts for walking, turning, running and backward walking are distilled into a single network structured by a skill latent and subsequently fine-tuned through reinforcement learning on transition sequences. The resultant unified human policy achieves a mean tracking success rate of 94.7% on unseen clips of the four skills and exhibits greater robustness to perturbations than its constituent experts. A single hip exoskeleton controller (UniExo) is initialized from hip moment prediction of the human policy and co-adapted with it through multi-agent reinforcement learning across the four skills. This co-adaptation shifts the timing of the assistance torque and raises the fraction of positive work delivered to the hip. When deployed on a custom hip exoskeleton, the controller generalizes across four treadmill speeds in six participants and assists one participant through a continuous route of all four skills and their transitions, without skill labels or explicit mode switching. UniExo thus provides a step towards replacing activity-specific controllers with unified, user-specific controllers that support diverse locomotor activities and the transitions between them.

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.

Matt Strong