IMPORTANT: yes
When Preferences Vary Smoothly: Learning Exoskeleton Assistance from Limited Feedback
Personalizing an exoskeleton to an individual user is hard enough. Personalizing it across multiple operating conditions, each requiring a different set of assistance parameters, is harder still. Every new condition demands fresh feedback: the user must experience the assistance settings, form judgments about comfort and support, and communicate those judgments through ordinal ratings or pairwise comparisons. This data-collection bottleneck grows with the number of operating conditions, from walking speed to lifting load, and ultimately constrains how broadly a personalized exoskeleton can be deployed.
Sunin Baek, Sungwoo Park, and Daekyum Kim propose a solution grounded in a simple but empirically important question: do a user's preferences vary smoothly across operating conditions? If so, then feedback collected in one context should inform estimates in nearby contexts, reducing the number of exposures needed to achieve good personalization. Their paper, submitted September 23, 2026, introduces Context-Continuous Preference Learning (CCPL), a Gaussian-process preference model that shares observations across nearby contexts while retaining context-specific utility estimates, and validates it through simulations and retrospective analyses of ankle and elbow exoskeleton data from nine healthy adults.
The Core Idea: Sharing Observations Across a Continuum of Contexts
The CCPL framework models a user's preference landscape as a latent utility function defined jointly over assistance parameters and operating context. For each context c, the utility function f(x, c) assigns a real-valued preference score to each candidate assistance parameter vector x, with larger values indicating stronger preference within that context. The learning objectives are to reconstruct this joint utility surface and to identify the optimal assistance settings at each context.
The model uses a Gaussian-process prior over the joint action-context space, with a covariance kernel that factorizes into separate action and context components. The action component captures how preferences change as assistance parameters vary; the context component captures how preferences change as operating conditions shift. By tuning the length scale of the context covariance, the model controls how quickly correlations decay with differences in context: short length scales mean observations from one context are informative only about very similar contexts, while long length scales encourage sharing across a broader range.
Three learning strategies are compared. Independent preference learning fits a separate Gaussian-process model for each context, sharing no information. Pooled preference learning fits a single model across all contexts but discards the context structure entirely, forcing the same utility function regardless of operating condition. CCPL sits between these extremes: it shares observations across nearby contexts through the context covariance while retaining the ability to produce distinct predictions in each context.
Evidence for Context Continuity in Human Data
The first empirical question is whether the smoothness assumption holds. The researchers analyzed retrospective data from two human studies. In the ankle study, six participants evaluated 101 distinct assistance profiles at each of three walking speeds (1.00, 1.25, and 1.50 m/s). In the elbow study, six participants performed loaded curls across five contexts defined by load level and fatigue-induction dose.
For each participant, the researchers computed full-data reference landscapes, estimated separately for each context, and then measured how similar these landscapes were as a function of the distance between contexts. In both the ankle and elbow studies, full-data reference landscapes tended to be more similar between nearby operating conditions than between distant ones. The negative slope of similarity versus context distance provides direct evidence that the smoothness assumption is empirically grounded in these exoskeleton domains.
This finding is not trivial. The authors note that some prior work has optimized assistance parameters separately for different walking speeds without attempting to share information across conditions. Their results suggest that such separation may discard useful signal, but that sharing must be done carefully through a model that respects the context structure.
Limited-Feedback Reconstruction: Fewer Exposures, Better Estimates
The central practical claim is that CCPL reduces the feedback burden. The researchers evaluated reconstruction accuracy using Pearson correlation between posterior mean utility estimates and full-data reference landscapes, with varying numbers of exposures per context. With five exposures per context, CCPL increased mean reconstruction correlation from 0.644 to 0.720 for ankle assistance and from 0.476 to 0.526 for elbow assistance, relative to independent learning.
These improvements translate directly into feedback-budget savings. To match CCPL's reconstruction correlations at five exposures, independent learning would require approximately 37% more exposures for ankle assistance and 17% more for elbow assistance. For the ankle study, this means that a user could achieve the same reconstruction quality with roughly 3.3 exposures per context instead of 5, a substantial reduction in the time and physical effort required.
Simulation experiments provided a systematic test of the continuity assumption's role. By varying a mixing parameter γ that controlled the proportion of smoothly varying versus independently generated utility components, the researchers showed that CCPL improved reconstruction and preference-based Bayesian optimization when preferences varied smoothly (high γ), but showed negative transfer when continuity was weak (low γ). This demonstrates that the method's benefits are not universal but depend on the underlying structure of the preference landscape.
Beyond Reconstruction: Bayesian Optimization and Held-Out Prediction
Reconstruction accuracy is only one criterion. The ultimate goal of preference-based personalization is to find the best assistance settings efficiently, which is a problem for Bayesian optimization. CCPL also improved preference-based Bayesian optimization relative to independent learning under smooth preference conditions, finding preferred actions with fewer exposures.
Held-out response prediction, which evaluates how well the model predicts new preference comparisons from held-out data, also favored CCPL over independent learning in both the ankle and elbow studies. This suggests that the context-sharing structure captures genuine regularities in user preferences rather than merely interpolating more densely within each context.
The benefits over pooled learning, however, were mixed. Pooled learning, which discards context structure, sometimes performed comparably or even better than CCPL, depending on the study and the evaluation metric. This is an important nuance: sharing across contexts is beneficial only when the context covariance structure correctly models the relationships, and when the number of contexts is not so large that the shared observations overwhelm the context-specific signal.
The Source-Reuse Analysis
An additional analysis examined how observations from other operating contexts could be reused when personalizing for a new context. For each target context, the researchers varied the number of source contexts and measured how much this improved reconstruction. The results showed that even a single source context could provide meaningful benefit, and that adding more sources yielded diminishing returns. This finding has direct implications for practical deployment: a user who has been profiled in one operating condition could have their preferences partially transferred to a new condition with minimal additional data collection.
Limitations and Open Questions
The study has several important limitations. The human data are retrospective, meaning the experimental designs were not optimized for evaluating CCPL's learning strategies. The ankle study used only three walking speeds, and the elbow study used five contexts defined by discrete load and fatigue levels; the true continuity of preferences across a continuous range of conditions remains to be tested.
The simulations used a specific generative model for preference landscapes, and the results depend on how well that model captures real preference structures. The context length scales were fixed rather than learned, and their sensitivity to this choice warrants further investigation.
Most importantly, the paper's conclusion is careful: the benefits of CCPL for online personalization in humans remain to be established. All evaluations were retrospective, using data from completed sessions. Whether CCPL can genuinely reduce the number of exposures needed during live personalization, where the model's choices influence what data is collected, is an open question that requires online experimentation.
The authors also note that the current framework assumes a single user. Extending CCPL to share information across users, perhaps through a hierarchical prior, could further reduce the data burden, particularly for new users who have no prior personalization data. This direction connects to the broader literature on multitask learning and cross-user transfer in preference-based optimization.
Why This Matters
The paper addresses a practical tension in wearable robotics: the need for personalization across many operating conditions collides with the physical and temporal cost of collecting user feedback. By showing that preferences do vary smoothly and that this smoothness can be exploited through Gaussian-process modeling, CCPL provides a principled approach to reducing the feedback burden.
The quantitative results are substantial. A 37% reduction in the feedback budget for ankle assistance means that a clinical or home-based exoskeleton fitting session could cover more operating conditions in the same time, or the same conditions with less user fatigue. For the elbow study, even the smaller 17% reduction is meaningful given the additional burden of fatigue induction in each context.
The methodological contribution, a Gaussian-process preference model with factorized action-context covariance, is relatively simple but effective. Its transparency, compared to deep learning approaches, makes it easier to diagnose when sharing is helping versus when it introduces negative transfer. The simulation experiments that deliberately vary the continuity parameter provide a template for evaluating sharing-based methods under controlled conditions.
These findings support context continuity as a basis for sharing preference observations under limited feedback, offering a practical path toward exoskeletons that can be personalized more efficiently across the range of conditions they must operate in. The next step is to validate these benefits in online personalization, where the model's choices actively shape the data collection process.