Continual Online Personalization of Exoskeleton Control via Manifold-Aware Experience Replay

Changseob Song · Inseung Kang
Carnegie Mellon University
Paper arXiv

Presentation video

Why this matters

Continual online adaptation

Updates the gait phase estimator in real time with sparse sensors, so assistance timing remains personalized as the user and task change.

Task-aware replay without labels

Uses a gait manifold to group similar locomotor states, avoiding explicit task labels during replay selection.

Reduced catastrophic forgetting

Preserves previously learned assistance behavior across task transitions by replaying the most error-prone prior bins.

Users encountered a series of scenarios designed to replicate continual online personalization during daily ambulation.

Scenario 1: Personalization across multimodal tasks

The user transitions across multiple locomotor modes while the controller updates online. The replay mechanism saves ongoing task's data and labels within the gait manifold.

Scenario 2: Forgetting in long-exposured task

The user is exposured to a single task for 5 minutes. The replaying framework keeps repalying previously seen data, while the baseline continuously update with current task.

Scenario 3: Testing on forgotten tasks

The controller is tested on other tasks than the one for forgetting. The replaying framework preserves its experiences on other tasks, while the baseline suffers from forgetting.

Framework at a glance

The manifold-aware experience replay framework for online adaptation (OA) consists of a latent-space replay buffer (blue), an OA module (red), and a main control loop (black) that predicts gait phase from sensor inputs to command torque via mid-level control splines.

Method

Exoskeleton HW & SW

Exoskeleton hardware and software diagram
Robotic hip exoskeleton designed to assist the user’s bilateral hip motion and the real-time control flow within the edge controller.

Hemiplegic gait emulation

Hemiplegic gait emulation setup
Hemiplegic gait was emulated using metronome cue and knee brace. This caused temporal asymmetry of user's gait, as well as kinematic asymmetry.

Main results

40% lower torque RMSE vs. baseline without replay
60% better gait phase tracking vs. baseline
~1.5s average update latency after replay stabilizes
After the subjects experienced each task (personalizing phase), they walked for 5 minutes for a single task (forgetting phase). When tested on different tasks other than the forgetting task (testing phase), significant decrease in both torque and gait phase tracking errors were observed for all task transitions.

Citation

BibTeX

@article{song2026continual,
  title={Continual Online Personalization of Exoskeleton Control via Manifold-Aware Experience Replay},
  author={Song, Changseob and Kang, Inseung},
  journal={arXiv preprint arXiv:2606.17455},
  year={2026},
  url={https://arxiv.org/abs/2606.17455}
}