When an exoskeleton changes how it helps a person walk faster or lift a heavier load, the user may have to try many assistance settings before the device finds a comfortable fit. A new preference-learning model reuses feedback from nearby conditions to reduce how much data it needs to estimate what a person prefers.
The method, called Context-Continuous Preference Learning, treats a user’s preferences as related across conditions such as walking speed or lifting load, but not identical. Its Gaussian-process model shares information between nearby conditions while keeping a separate estimate for each one. In effect, feedback from one setting can guide the search in a similar setting without assuming the same assistance profile will suit both.
Researchers tested the model retrospectively on ankle- and elbow-exoskeleton data from nine healthy adults. They compared it with learning each condition separately, using the same recorded user responses. With five exposures per condition, CCPL’s estimates correlated more closely with full-data model references than estimates from learning each condition separately did: 0.720 versus 0.644 for ankle assistance and 0.526 versus 0.476 for elbow assistance. These references were models fitted to all available observations, not direct measurements of users’ true preferences.
At that five-exposure budget, the separate-learning method needed an estimated 7.94 ankle exposures or 6.02 elbow exposures to reach the same reconstruction scores. That corresponds to 37% and 17% fewer exposures in the retrospective analysis. An exposure meant trying one assistance profile and recording a user assessment; the paper did not test whether CCPL reduces actual appointment time or improves movement in a live system.
The approach depends on a condition that may not always hold: nearby settings must have sufficiently similar preference patterns. In simulations where preferences changed independently between conditions, sharing data could perform worse than learning each condition separately. The human data also came from small studies of healthy adults, and the authors say prospective trials are still needed to show whether the model can choose better assistance online.
Sources
- Baek, Park and Kim, “Context-Continuous Preference Learning for Exoskeleton Personalization,” arXiv:2609.28427
- Lead image: Paulo O, “Robotic Exoskeleton”, CC BY 2.0, via Wikimedia Commons.
.jpg){kind=link}