June
2026
User Intent-Driven and Context-Aware Personalization for Assistive Exoskeletons
Authors:
Abstract:
Personalizing exoskeleton control to individual users remains a central challenge to real world deployment across healthy and clinical populations. Existing approaches optimize personalization parameters for biomechanical objectives such as metabolic cost or joint loading, requiring resource and time-intensive, user-specific data collection, without accounting for user comfort, intent, or task context. Prior work shows that users can perceive and report their preferred assistance, yet no lightweight method maps this expressed intent to quantitative control parameters in real time.
We present a vision-language model (VLM) guided human-exoskeleton interface that translates natural user feedback and egocentric visual context into parameter updates for a hip exoskeleton controller, and treats user feedback as reward to train a contextual bandit on user preferences. The framework separates high-level intent parsing performed by the VLM, from low-level action selection and exploration performed by a user-specific contextual bandit that refines its policy with each interaction, requiring no lab setup or biomechanical instrumentation. This system runs on a lower-limb hip exoskeleton, in real time, tested with users in a pilot study.
We evaluated the framework on a five-task locomotion track across 2 laps, spanning level ground, ramps, and stairs, against a single-shot interpreter and a few-shot recommendation system as personalization baselines. Our contextual bandit framework reduced no improvement personalization instances by 31\% and raised terminal user satisfaction by the second lap (5.15 to 5.67), with mean within-episode gains consistent with maximum learning across the first lap (+3.71 gain) and sustained learning by the second (+2.75). Our pilot results validate our framework, supporting the use of VLMs as context-aware interpreters for real-time, user-intent driven assistance personalization for lower limb exoskeletons, and contextual bandits as lightweight online learners for modeling user preferences. This framework highlights the potential of incorporating user preferences through natural-language interfaces for personalizing exoskeleton assistance to ensure user comfort and support long-term adoption of these devices.
We present a vision-language model (VLM) guided human-exoskeleton interface that translates natural user feedback and egocentric visual context into parameter updates for a hip exoskeleton controller, and treats user feedback as reward to train a contextual bandit on user preferences. The framework separates high-level intent parsing performed by the VLM, from low-level action selection and exploration performed by a user-specific contextual bandit that refines its policy with each interaction, requiring no lab setup or biomechanical instrumentation. This system runs on a lower-limb hip exoskeleton, in real time, tested with users in a pilot study.
We evaluated the framework on a five-task locomotion track across 2 laps, spanning level ground, ramps, and stairs, against a single-shot interpreter and a few-shot recommendation system as personalization baselines. Our contextual bandit framework reduced no improvement personalization instances by 31\% and raised terminal user satisfaction by the second lap (5.15 to 5.67), with mean within-episode gains consistent with maximum learning across the first lap (+3.71 gain) and sustained learning by the second (+2.75). Our pilot results validate our framework, supporting the use of VLMs as context-aware interpreters for real-time, user-intent driven assistance personalization for lower limb exoskeletons, and contextual bandits as lightweight online learners for modeling user preferences. This framework highlights the potential of incorporating user preferences through natural-language interfaces for personalizing exoskeleton assistance to ensure user comfort and support long-term adoption of these devices.