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

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

September 2026

TL;DR MO-HILBO uses multi-objective Bayesian optimization to learn each wearer's set of Pareto-optimal hip exoskeleton controllers, trading metabolic cost against self-rated comfort. With 24 three-minute trials per person (3 subjects), its Pareto controllers dominated anti-Pareto ones in validation, and it predicted 17 of 18 pairwise orderings of held-out controllers.

What they do

Human-in-the-loop optimization usually tunes an exoskeleton for one objective, or for a fixed weighted sum of several. MO-HILBO instead learns each wearer’s Pareto set: the controllers where neither comfort nor metabolic cost can improve without worsening the other. The authors compare it with an evolutionary baseline (NSGA-II) in simulation, run it on a hip/knee exoskeleton with three participants, and release it as the Python package mohilo.

How it works

Surrogate. One Gaussian process per objective, with hyperparameters fit under a prior on measurement noise. The Pareto set is read off the GP means.
Acquisition. BoTorch’s qLogNParEGO draws a new random weighting of the objectives at every query, so successive trials probe different trade-offs.
Decoupled feedback. Each 3-minute trial yields one metabolic estimate and up to three 1-5 comfort ratings, since pilot subjects often changed their minds within a trial.
Controller. Hip torque tracks a biological joint moment that a neural network estimates from IMUs and insoles. MO-HILBO tunes its delay (0-0.375 s) and the scale factors for positive and negative moment phases (0-0.3 each); the knee motors give zero torque.

Results

Simulation (50 runs per setting): MOBO reaches a lower IGD+ than NSGA-II within 40 queries at every noise level tested, and stays ahead for 1-4 action dimensions and 2-5 objectives.
Human study: 24 trials per subject walking at 1.2 m/s, 3-4 hours per session. In 5-minute validation trials, three Pareto controllers dominated three anti-Pareto ones for every subject (hypervolume 4.02 vs 1.300, 1.55 vs 0.366, 1.02 vs 0.128).
The surrogate predicted 17 of 18 (94%) pairwise orderings of held-out Pareto controllers; the one miss was a metabolic-cost ordering.
All subjects found high delay neither comfortable nor efficient. The two first-time users found low torque most comfortable; the experienced user found high flexion and extension torque both comfortable and efficient.

Limitations

Three participants, one session each, and only three hip parameters, with the knee left unpowered to keep sessions short. In simulation, four action dimensions were already harder for MOBO. The method assumes a person’s objectives stay fixed, which may not capture adaptation, fatigue or shifting preferences.

Figures

From Figure 1: MO-HILBO on a lower-limb exoskeleton, optimizing (a) metabolic cost, measured by indirect calorimetry,... — From Figure 1: MO-HILBO on a lower-limb exoskeleton, optimizing (a) metabolic cost, measured by indirect calorimetry, and (b) comfort, rated on a tablet. Top right: the Pareto front of the learned surrogate; bottom right: the matching Pareto set of control parameters.

From Figure 2: The MO-HILBO loop. Comfort ratings from a tablet and metabolic cost from indirect calorimetry update o... — From Figure 2: The MO-HILBO loop. Comfort ratings from a tablet and metabolic cost from indirect calorimetry update one Gaussian process per objective, and a multi-objective acquisition function picks the next exoskeleton controller to try.

From Figure 5: (a) The fitted model for subject MB02, whose training data lie near the inferred Pareto front. (b) Pre... — From Figure 5: (a) The fitted model for subject MB02, whose training data lie near the inferred Pareto front. (b) Predicted (stars) versus measured (squares) validation controllers from the Pareto and anti-Pareto sets of each subject. (c) The three subjects' Pareto sets in action space: they mostly do not overlap, and none has a high delay.

Abstract

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/.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong