"> Generating Personalized Lower-Limb Kinematics Across Walking Speeds Using Subject-Conditioned Diffusion - Diya Dinesh, Adrian Krieger, Changseob Song, Dongho Park, Aaron J. Young, Inseung Kang | Academic Research

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

Diya Dinesh1,*, Adrian Krieger2,*, Changseob Song2†, Dongho Park3,
Aaron J. Young3, Inseung Kang2
1School of Computer Science, Carnegie Mellon University
2Department of Mechanical Engineering, Carnegie Mellon University
3Woodruff School of Mechanical Engineering and the Institute for Robotics and Intelligent Machines, Georgia Institute of Technology
*Indicates Equal Contribution
†Corresponding Author
Code arXiv

Given a user-specific gait trajectory at one walking speed, our model generates the residual change needed to synthesize that subject's gait at a new target speed.

Motivation

Why personalized gait data?

Exoskeleton assistance needs to adapt to each user because human gait varies across individuals and locomotor tasks.

Why is data collection hard?

Per-subject data collection requires repeated motion-capture trials across diverse speeds and tasks, which is time-consuming and costly, particularly challenging for clinical populations.

What do we propose?

We generate personalized gait kinematics across diverse target speeds from a single source-speed sample using subject-conditioned residual diffusion, which preserves subject-specific gait characteristics.

Architecture

Model architecture diagram

Subject-conditioned residual diffusion framework. Given a source gait trajectory at one walking speed and a desired target speed, the model predicts the residual change needed to synthesize the subject's gait at the target speed. Subject-specific gait features and speed information are used to condition the denoising process, allowing the generated trajectory to preserve individual motion characteristics.

Results

Main Results

First result visualization

Our model generated accurate target-speed gait for unseen able-bodied subjects and generalized to stroke subjects without fine-tuning, while preserving subject-specific gait patterns.

Representative Able-Bodied Subjects

Representative gait trajectories for the able-bodied subject with median error among all held-out subjects showing the best, median, and worst source-to-target speed pairings.

Representative Stroke Subjects

Representative gait trajectories for the stroke subject with median error among all subjects showing the best, median, and worst source-to-target speed pairings.

Error Heatmaps

First result visualization

MAE across source-target speed pairings for (a) able-bodied and (b) stroke subjects. The model maintains low errror across a wide range of speed transitions.

Performance under various source speed conditions

Second result visualization

(a) Able-bodied MAE per source speed across evaluated models.
0.41° range of average error so the model remains robust regardless of source speed

(b) Impact of source speed availability on able-bodied data.
Only a 0.38° decrease between one and four speeds, showing one source speed is sufficient for personalized generation

(c) Average able-bodied MAE across source-target speed distance.
Error below 5° for non-diagonal pairings with only 1.5° of variability

Personalization Heatmaps

Third result visualization

We compute the MAE between the generated trajectory and each subject's real trajectory at the same target speed, then rank the subjects by error. Lower rank indicates stronger subject identity preservation.

Cadence Analysis

Fourth result visualization

Generated and ground-truth cadence are shown for (a) able-bodied and (b) stroke individuals
6.8% average error for able-bodied subjects and 7.4% average error for stroke subjects indicating temporal consistency in the generated data.

BibTeX


        @misc{dinesh2026generatingpersonalizedlowerlimbkinematics,
          title={Generating Personalized Lower-Limb Kinematics Across Walking Speeds Using Subject-Conditioned Diffusion}, 
          author={Diya Dinesh and Adrian Krieger and Changseob Song and Dongho Park and Aaron J. Young and Inseung Kang},
          year={2026},
          eprint={2607.07533},
          archivePrefix={arXiv},
          primaryClass={cs.RO},
          url={https://arxiv.org/abs/2607.07533}, 
    }