GaitDynamics Demo

This demo uses GaitDynamics, a diffusion model, to generate running dynamics. The runner's forward speed is prescribed as the input ; the full-body kinematics and ground reaction force (GRF) are generated by the model as outputs. Here we hold the speed fixed and vary the cadence to see how it leads to different gait parameters.

The demo has three steps:

  1. Explore real data — Browse the real kinematics and GRF recorded during the experiment.
  2. Prescribe cadence — Set a low and a high cadence at the same fixed speed. The sliders preview how the stride timing is re-timed.
  3. Generate — Run the model to synthesize the motion for both cadences and compare them over the gait cycle.
1

Explore real data

This is a single real running trial (speed = 4.0 m/s). The subject height is 1.74 m and weight is 69.3 kg.

On the left, pick any kinematic signal (joint angles, in degrees) to see how it evolves over time; on the right, pick a GRF or center-of-pressure (CoP) signal. GRF components are shown in Newtons (N), and CoP positions in meters.
Kinematics
Force / CoP
2

Prescribe cadence

Speed is fixed, identical to the recorded trial. Cadence is what we change. A cadence scale s simply resamples the stride: the trial is temporally resampled to T / s of its original duration, so the runner takes faster, shorter steps when s > 1 and slower, longer steps when s < 1. Low = 85–99%, high = 101–115%.

The previews below show the resampled left-knee flexion: the curve's shape and amplitude are unchanged, only its timing compresses or stretches with cadence.
85 99
101 115
3

Generate kinematics for each cadence

Now we run GaitDynamics, the diffusion model, to synthesize the full-body motion for each of the two cadences.

Because the process is stochastic, we draw several samples — each one is called a seed. A seed is just one random starting point, so different seeds give slightly different but equally plausible motions. The status box reports progress as seed k / N; the plots below show the mean ± standard deviation across seeds, so you can see both the typical motion and how much it varies. Generation runs the model many times and may take a couple of minutes.
3D skeleton overlay — low cadence / high cadence (kinematics + GRF arrows)

The two generated cadences are overlaid on one skeleton so you can compare the strides directly. The arrows represent the generated GRF vectors.
Run the model to view the 3D skeleton overlay.
Compare cadences over the gait cycle

After generating, pick any kinematic or force / CoP channel to compare the low- vs. high-cadence conditions over the gait cycle (mean ± SD across seeds).
Kinematics
Force / CoP

Citation

Tan, T., Van Wouwe, T., Werling, K. F., Liu, C. K., Delp, S. L., Hicks, J. L., & Chaudhari, A. S. GaitDynamics: a generative foundation model for analyzing human walking and running. Nat. Biomed. Eng (2026). https://doi.org/10.1038/s41551-025-01565-8