5th Year Master's Thesis Presentation - Diya Dinesh

About this event

Speaker: DIYA DINESH, 5th Year Master's Student, Computer Science Department, Carnegie Mellon University

Talk Title: Personalized Gait Kinematics and Kinetics Across Walking Conditions

Personalized gait modeling is essential for individualized assistive technologies, biomechanical models, and rehabilitation strategies, but requires user-specific gait data across many locomotor tasks that is costly and time-intensive to collect and especially burdensome for clinical populations. This challenge extends to both lower-limb kinematics and kinetics, which describe joint motion and the associated demands across walking speeds and inclines. This thesis investigates conditional residual generative models for producing personalized lower-limb kinematics and kinetics given limited subject-specific measurements.

First, a subject-conditioned residual diffusion framework generates personalized lower-limb kinematics at target walking speeds not observed for a held-out subject given their sagittal-plane hip, knee, and ankle trajectories at a seen source speed. The model uses a transformer denoiser conditioned on the source gait and source and target speeds to generate residual changes in the joint kinematics. Trained only on able-bodied data, the model achieved a mean absolute error (MAE) of 3.5 degrees on held-out able-bodied subjects and 6.6 degrees MAE on out-of-training-distribution stroke subjects without fine-tuning, retaining subject identity for clinical gait. It outperformed a deterministic transformer with the same backbone and conditioning, while a single source speed achieved accuracy within 0.1 degrees of using four available source speeds.

Second, a residual flow-matching framework generates lower-limb joint moments at target walking speeds using source kinematics, source kinetics (joint moments), and measured target kinematics. The model learns residual changes in the joint moments conditioned on the available subject-specific measurements and walking speeds. It achieved an MAE of 0.0618 Nm/kg on held-out able-bodied subjects, outperforming source-copying and linear speed-scaling baselines. The kinetic framework was further adapted to stroke gait and incline walking. Fine-tuning reduced stroke MAE from 0.1199 to 0.0807 Nm/kg and improved accuracy on inclined walking.

Together, these results demonstrate the potential of conditional residual generative models to generate personalized lower-limb biomechanics across walking conditions while reducing the need for repeated subject-specific measurements, with potential applications in assistive technology, personalized biomechanical modeling, and rehabilitation.

Thesis Committee:

Inseung Kang, Chair

Aran Nayebi

Additional Information