Bayesian Preference Elicitation: Human-In-The-Loop Optimization of An Active Prosthesis

Sophia Taddei; Wouter Koppen; Eligia Alfio; Stefano Nuzzo; Louis Flynn; Maria Alejandra Diaz; Sebastian Rojas Gonzalez; Tom Dhaene

cs.ROFeb 26, 2026

Bayesian Preference Elicitation: Human-In-The-Loop Optimization of An Active Prosthesis

Authors: Sophia Taddei, Wouter Koppen, Eligia Alfio, Stefano Nuzzo, Louis Flynn, Maria Alejandra Diaz, Sebastian Rojas Gonzalez, Tom Dhaene,

Organizations: Federated Labs Robotics & AI, Vrije Universiteit Brussel. · Human Physiology and Sports Physiotherapy Research Group, Vrije Universiteit Brussel. · IDLab, Ghent University – imec. · Robotics & Multibody Mechanics Research Group, Flanders Make.

Abstract

Tuning active prostheses for people with amputation is time-consuming and relies on metrics that may not fully reflect user needs. We introduce a human-in-the-loop optimization (HILO) approach that leverages direct user preferences to personalize a standard four-parameter prosthesis controller efficiently. Our method employs preference-based Multiobjective Bayesian Optimization that uses a state-or-the-art acquisition function especially designed for preference learning, and includes two algorithmic variants: a discrete version (EUBO-LineCoSpar), and a continuous version (BPE4Prost). Simulation results on benchmark functions and real-application trials demonstrate efficient convergence, robust preference elicitation, and measurable biomechanical improvements, illustrating the potential of preference-driven tuning for user-centered prosthesis control.

Figures & tables

Figure 1: Mean $\log_{10}$ simple regret over 11 runs for all benchmark functions. Each algorithm was run for 50 iterations. Shaded regions indicate the standard error. Across both the 2D problem (Branin) and the 4D problems (Ackley, Alpine1, Hartmann), BPE4Prost consistently achieves lower regret compared to the other methods.

Figure 2: Runtime comparison of the three algorithms on Ackley (4D) and Branin (2D). BPE4Prost scales approximately linearly with the number of queries, while EUBO-LineCoSpar remains near-constant due to its line-restricted evaluations.

Figure 3: [A] Active prosthetic knee system used in this study. [B] High-level control framework based on a finite state machine (FSM) with three states: Stance (S), Swing Flexion (SF), and Swing Extension (SE). Threshold rules between states are indicated with tunable parameters written in red. [C] Mid-level control framework. A position controller is used in S, while impedance controllers are employed in SF ans SE. Tunable parameters are indicated in red.

Subject	Config.	$t_{\mathrm{SF}}$ (%)	$k_{\mathrm{SF}}$ (Nm/ ∘ )	$\theta_{\mathrm{SF}}$ ( ∘ )	$k_{\mathrm{SE}}$ (Nm/ ∘ )
AB1	$\hat{x}_{N,1}$	45	0.4	45	0.35
	$\hat{x}_{N,2}$	50	0.4	40	0.35
	$\hat{x}_{N,3}$	55	0.4	50	0.35
AB2	$\hat{x}_{N,1}$	40	0.4	55	0.35
	$\hat{x}_{N,2}$	40	0.4	45	0.35
	$\hat{x}_{N,3}$	40	1.8	45	0.35

Table 1: Converged configurations obtained for each participant. Each participant completed three trials, yielding preferred parameter vectors

\{\hat{x}_{N,i}\;|\;i\in\{1,2,3\}\}

of the form

\left[t_{\mathrm{SF}},\,k_{\mathrm{SF}},\,\theta_{\mathrm{SF}},\,k_{\mathrm{SE}}\right]

. For participants AB1, AB2, and TF1, optimization was conducted over a discrete grid with five points per parameter, whereas TF2 was optimized in a continuous space.

Figure 4: Comparison of knee angle trajectories for preferred solutions and their corresponding challengers. Each subplot displays the mean $\pm$ std of the knee angle over the normalized gait cycle (0–100%) across steps. The number of gait cycles used to compute each mean trajectory is indicated.

Figure 5: Evolution of the estimated optimal configuration for TF2,

\hat{x}=\arg\max_{x}\mu_{N}(x)

, for the three trials, shown for each dimensions. The final estimate for each trial is highlighted. Trials 1 and 2 converge rapidly in most of the dimension, whereas Trial 3 exhibits increased variability, likely due to noisy user feedback.

Preference Elicitation

Preference elicitation studies how systems learn people’s values and trade-offs from questions, choices, rankings, and feedback.

See all 69 papers

Multi-Objective Human-in-the-Loop Bayesian Optimization of a Lower-Limb Exoskeleton
Sep 25, 2026Neil Janwani, Matthew T. Lerner, Aaron J. Young +1Human-in-the-Loop AI Bayesian Optimization
Georgia Institute of Technology
Personalized Lower-limb Exoskeleton Assistance via Preference-based Bayesian Optimization
Aug 10, 2026Xiao-Yin Liu, Guotao Li, Weiqun Wang +1Bayesian Optimization Ankle Exoskeleton Control
State Key Laboratory of Multimodal Artificial Intelligence Systems, Institute of Automation, Chinese Academy of Sciences, Beijing 100190, China · School of Artificial Intelligence, University of Chinese Academy of Sciences, Beijing 100049, China · CASIA-MUST Joint Laboratory of Intelligence Science and Technology, Institute of Systems Engineering, Macau University of Science and Technology, Macao, China
Efficient Controller Learning from Human Preferences and Numerical Data Via Multi-Modal Surrogate Models
Mar 25, 2026Lukas Theiner, Maik Pfefferkorn, Yongpeng Zhao +2Bayesian Optimization Multi-Fidelity Optimization
Control and Cyber-Physical Systems Laboratory, Technical University of Darmstadt, Darmstadt, Germany · Volkswagen AG, Wolfsburg, Germany

Sep 25, 2026J/K move · Enter open · S save

Multi-Objective Human-in-the-Loop Bayesian Optimization of a Lower-Limb Exoskeleton

Neil Janwani, Matthew T. Lerner, Aaron J. Young, Maegan Tucker

Georgia Institute of Technology

Human-in-the-Loop AI Bayesian Optimization Assistive Robotics Ankle Exoskeleton Control

Open paper AskPDF arXiv

Human-in-the-loop optimization (HILO) is a common approach for optimizing the control of assistive devices to account for the wearer's unique biomechanics and subjective preferences. However, despite research suggesting that a person may have a different prioritization of objectives depending on time-varying factors such as the environment, their mood, or energy levels, existing HILO approaches only consider a single objective or enforce a fixed weighting on a set of objectives. Neither approach is capable of representing an individual's preferences over objectives. In this work, we propose Multi-Objective Human-in-the-loop Bayesian Optimization (MO-HILBO), which builds on explicit multi-objective Bayesian optimization to efficiently infer a personalized set of Pareto-optimal controllers. We compare our approach with an existing multi-objective HILO method and experimentally demonstrate MO-HILBO on a lower-limb exoskeleton across two objectives: metabolic cost (efficiency) and ordinal human feedback (comfort). We find that MO-HILBO (1) discovers Pareto-optimal controllers, and (2) that the pairwise ordering of points on the Pareto front itself is consistent with validation trials. Lastly, we open-source mohilo, a Python package for running both HILO and MO-HILBO on wearable devices: https://dynamicmobility.github.io/mohilo/.

Abstract

Preference Elicitation

Explore similar work

Multi-Objective Human-in-the-Loop Bayesian Optimization of a Lower-Limb Exoskeleton

Personalized Lower-limb Exoskeleton Assistance via Preference-based Bayesian Optimization

Efficient Controller Learning from Human Preferences and Numerical Data Via Multi-Modal Surrogate Models