GaitDynamics

A camera loses sight of a pelvis marker, a wireless packet drops, or a study never instrumented the trunk — and a conventional gait model that maps a fixed set of joint angles onto ground reaction forces has nothing to say. The rigidity cuts the other way too: such models estimate what was not measured, but cannot say what would happen to knee loading if a patient walked differently. Measuring that means a force-plate laboratory and hours of expert time, or a physics-based simulation that can run for hundreds of hours per trial.

GaitDynamics dissolves both problems by refusing to designate inputs and outputs at all. It learns a single distribution over a 1.5-second window whose 63 channels hold pelvis linear velocities, joint angles, joint angular velocities, ground reaction forces, and centres of pressure together. Any subset can be declared known and the remainder generated by diffusion inpainting, so one set of weights covers force estimation from whatever kinematics happen to exist, synthesis of gait that was never recorded, and prediction of how dynamics shift when a movement pattern is deliberately altered. It was developed at Stanford University by the Delp, Hicks, and Chaudhari labs, released as a preprint in March 2025 and published in Nature Biomedical Engineering in January 2026.

That makes it a different object from wearable-signal foundation models such as RelCon and SSL-Wearables, which learn representations of raw accelerometry for downstream probing. GaitDynamics is generative over the mechanics themselves, and emits kinetics — forces and the joint moments computed from them — not embeddings.

#Key Features

• Inpainting over a joint kinematics-kinetics window: Missing channels are generated by the diffusion model rather than imputed. Masking both hips — the worst case, at three degrees of freedom each — costs 2.2% of body weight in peak vertical force accuracy; masking the trunk or both ankles costs 0.1%. Filling the same gaps with median values costs up to 36.4%.
• A dedicated force refinement model: A second transformer maps completed full-body kinematics to forces, which is more accurate than reading forces straight out of the diffusion window.
• Experiment-free prediction of gait modifications: Guided diffusion steers generation toward synthetic kinematics — medial-lateral trunk sway scaled to 2.0-3.0x normal, say — while a penalty keeps the rest of the body realistic, and knee adduction moment follows from the generated force and lever-arm vectors.
• Transfer to cohorts it never saw: Seven of the fifteen source studies were withheld from training entirely, and accuracy on them matched held-out participants from training studies.
• Speed: A 1.5-second trial generates in under a second on a laptop-class NVIDIA RTX 3050 GPU.

#Technical Details

The diffusion transformer uses 4 self-attention blocks with 4 attention heads, a 256-dimensional latent space, and 1,024 feedforward units, with rotary position embeddings along the temporal axis and a 1,000-step cosine noise schedule sampled via DDIM over 50 inpainting iterations. The force refinement transformer uses 6 blocks with 8 heads and 512 feedforward units. Both trained with the Adan optimizer in 30 hours on one NVIDIA RTX A6000. Training data come from the AddBiomechanics dataset — 15 harmonized motion-capture and force-plate studies, pruned to 34.8 hours from 270 participants, with 8 studies (169 participants, 10,352 trials) used for training and testing on 101 held-out participants. With full-body kinematic input, mean absolute errors were 3.2%, 1.2%, and 0.7% of body weight for the vertical, anterior-posterior, and medial-lateral force profiles and 3.7% for peak vertical force — below published minimal detectable change thresholds, and better than CNN and RNN baselines on the vertical and medial-lateral profiles.

#Applications

The flexible-input formulation targets settings where force plates are unavailable: remote rehabilitation, exoskeleton control, and running injury screening from whatever kinematics a partial setup can supply. The gait-modification pathway supports knee osteoarthritis work, where trunk sway retraining offloads the medial compartment and each candidate strategy would otherwise need its own experiment. A Hugging Face Space and a Colab notebook take an OpenSim .osim model and .mot kinematics and return forces and filled-in kinematics, so a lab can use the model without writing training code.

#Impact

GaitDynamics moves gait biomechanics from task-specific regression toward a shared pretrained substrate, a shift that waited on a corpus of synchronized kinematics and kinetics large enough to support it. The release is unusually complete: MIT-licensed code, both checkpoints in the repository, and a training corpus itself public under CC BY 4.0. The authors are candid about the boundaries. Training used healthy participants only; accuracy held for individuals with osteoarthritis but degraded for a post-stroke individual with markedly lower cadence. The model will not generate running above 5 m s⁻¹, its fastest training data being 4.7 m s⁻¹, and the fixed-window guided-diffusion formulation cannot express large cadence changes. Turning gaits were pruned, and all kinematic inputs were marker-based motion capture, leaving robustness to IMU- and camera-derived kinematics for others to establish.