Generating Personalized Lower-Limb Kinematics Across Walking Speeds Using Subject-Conditioned Diffusion

Jul 16, 2026

Personalizing exoskeleton assistance requires subject-specific gait data across many locomotor tasks, but collecting it involves repeated motion-capture sessions that are costly, time-intensive, and especially burdensome for clinical populations. The challenge is acute across walking speeds, where gait changes markedly and often diverges further in stroke. This work proposes a subject-conditioned residual diffusion framework to generate personalized lower-limb kinematics at unseen speeds from a single seen-speed sequence, transforming sagittal hip, knee, and ankle trajectories via a conditioned residual and transformer denoiser with feature-wise modulation, reducing MAE by over 70%.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

Introduction

Motivation: need for personalization in exoskeletons

Data bottleneck and clinical limitations

Generative models and diffusion for motion synthesis

Prior work and identified gaps

Proposed approach and main hypothesis

Generating Personalized Lower-Limb Kinematics Across Walking Speeds Using Subject-Conditioned Diffusion Diya Dinesh* , Adrian Krieger* , Changseob Song, Dongho Park, Aaron J. Young, Senior Member, IEEE, Inseung Kang, Member, IEEE arXiv :2607 .07533v 1 [ cs .RO] 8 Jul 2026 Abstract—Personalizing exoskeleton assistance requires userspecific gait data across many locomotor tasks, yet collecting this data demands repeated motion capture sessions that are costly, time-intensive, and especially burdensome for clinical populations. This challenge is most acute across walking speeds, where gait changes substantially and deviates further in clinical gait. This work introduces a subject-conditioned residual diffusion framework that generates personalized lower-limb kinematics at unseen walking speeds from a subject’s gait sequence at a single seen speed. Given sagittal-plane hip, knee, and ankle trajectories at a seen speed and a desired unseen speed, the model generates a residual that transforms the seen trajectory into the unseen one, using a transformer denoiser conditioned on the subject’s gait and the two speeds through feature-wise linear modulation. Trained only on able-bodied data, the model achieved a mean absolute error (MAE) of 3.4◦ on held-out able-bodied subjects. Without any strokespecific fine-tuning, it achieved a 6.0◦ MAE on out-of-trainingdistribution stroke subjects, retaining subject identity for clinical gait. The framework reduced the MAE by over 70% relative to supervised feed-forward baselines, and a single seen speed matched the accuracy of four speeds within 0.4◦ . These results demonstrate that subject-conditioned residual diffusion can synthesize personalized gait across speeds from minimal data, reducing the collection burden for downstream exoskeleton personalization. Index Terms—Subject-conditioned diffusion, exoskeleton personalization, domain generalization, stroke gait, personalized gait generation, lower-limb kinematics. I. INTRODUCTION Exoskeletons are gait-assistive devices with strong potential to improve mobility for individuals with motor impairments. [1]–[3] However, their effectiveness depends heavily on adapting to each user, since human gait varies substantially across individuals, walking speeds, impairment types, and biomechanical constraints [4]–[6] . Therefore, a one-sizefits-all controller may provide suboptimal assistance, reduce comfort, or destabilize walking. This challenge is especially important for clinical populations such as stroke survivors, where gait patterns can deviate substantially from ablebodied individuals [7]–[9] . Therefore, personalization is vital *These authors contributed equally to this work. (Corresponding author: Changseob Song, [changseob@cmu.edu](changseob@cmu.edu)) D. Dinesh is with the School of Computer Science, Carnegie Mellon University, Pittsburgh, PA, 15213 USA. A. Krieger, C. Song and I. Kang are with the Department of Mechanical Engineering, Carnegie Mellon University, USA. D. Park and A. J. Young are with the Woodruff School of Mechanical Engineering and the Institute for Robotics and Intelligent Machines, Georgia Institute of Technology, Atlanta, GA 30332 USA. This research was supported by the NIH R21 Award 1R21EB037268-01 and NIH DP2 Award 1DP2HD111709-01 . to provide safe and effective assistance while improving the user’s functional outcomes such as energy [8]–[11], fatigue [12], [13], and stability [8], [14]–[16] . A major bottleneck for current data-driven personalization is the large amount of data required. Since each user’s gait varies across locomotor contexts [4], [5], data from a small subset of tasks may be insufficient for effective personalization. In practice, collecting gait data requires repeated motion capture sessions across tasks, such as different walking speeds. While these sessions yield high-fidelity kinematic data, they are time-consuming and exhausting for clinical populations, making it infeasible t

FAQs

为什么跨步速的个性化对穿戴式外骨骼特别困难？

因为步速变化会显著改变步态形态，而临床人群（如中风患者）的步态与健康人群差异更大，导致采集与适配成本更高、外推更难。

该方法如何从“单一已见步速”的数据生成“未见步速”的个性化下肢运动学？

通过主题（受试者）条件的残差扩散框架：在给定髋、膝、踝的矢状面轨迹与目标未见步速时，模型学习一个残差来把已见步速的轨迹转换到未见步速，并使用以受试者步态和两个步速为条件的transformer去噪器。

实验结果体现了哪些关键性能特征？

仅在健康体数据上训练时，模型在未参与受试者上达到3.4° MAE；无需针对特定中风个体的微调，在训练分布外的中风受试者上实现6.0° MAE，并在准确性与受试者身份保持方面取得改进。

Jul 16, 2026

Upload by Evangeline