CCPL: Smoother Exoskeleton Personalization with Less Data

Summary

Source: arXiv cs.RO, posted Sept. 24, 2026

The authors introduce Context-Continuous Preference Learning (CCPL), a method designed to address the high physical and temporal cost of collecting user feedback for exoskeleton personalization. The core hypothesis is that a user's preference landscape changes smoothly as operating conditions shift. To test this, the team developed a Gaussian-process preference model that shares observations across nearby contexts while still allowing for context-specific utility estimates. This approach aims to make learning more efficient when feedback is limited.

Evaluation was conducted through simulations and retrospective analyses of preference data from nine healthy adults using ankle and elbow exoskeletons. In simulation environments, CCPL outperformed independent learning in reconstruction and preference-based Bayesian optimization when preferences were smooth, though it suffered from negative transfer when continuity was weak. In the human studies, full-data reference landscapes estimated separately for each participant and context showed greater similarity between nearby operating conditions. With only five exposures per context, CCPL increased the mean reconstruction correlation with these references from 0.644 to 0.720 for ankle assistance and from 0.476 to 0.526 for elbow assistance compared to independent learning. The five-exposure budget was approximately 37% lower for ankle and 17% lower for elbow than the estimated independent-learning budgets required to match these correlations. CCPL also improved held-out response prediction relative to independent learning, while benefits over pooled learning varied.

Why it matters

Personalizing robotic exoskeletons has traditionally relied on extensive trial-and-error sessions, which are physically demanding for users and time-consuming for clinicians. Existing approaches often treat each operating condition in isolation or pool all data together, ignoring the structured nature of human preference. By leveraging the assumption of continuity, this work builds on established Bayesian optimization techniques but adds a spatial prior that reflects how human motor preferences actually evolve. This positions CCPL as a more efficient alternative for real-time adaptation, potentially reducing the barrier to entry for personalized robotic assistance.

Robot's take

The strength of this approach lies in its data efficiency; reducing the feedback budget by up to 37% is a significant practical advantage for clinical settings. However, the evaluation relies heavily on retrospective analysis of healthy adults, which may not fully capture the variability seen in clinical populations with neurological impairments. The simulation results showing negative transfer when continuity is weak suggest that the model's performance is highly dependent on the underlying assumption of smooth preference landscapes. It remains to be seen whether these benefits translate to online personalization in real-time human-robot interaction, as the current human studies are retrospective. Future work should focus on prospective trials with diverse user populations to validate the robustness of the continuity assumption.