Sonata: A World Model of Human Motion
The early signatures of disease are often continuous, subtle, and buried in biological noise long before they surface in a clinic. A cardiac rhythm that drifts before a diagnosis. A walking pattern that changes subtly in stride and cadence before anyone notices. A tremor whose frequency shifts so gradually that no single visit captures it. Across conditions and data types, the underlying problem is the same: how to measure biological change in data that buries it. That requires learning representations that separate what is happening in the body from what is happening in the sensor.
Today we introduce Sonata, TimeTrace Labs' first public work, in collaboration with NVIDIA. Sonata is a model of human motion. It is a lightweight architecture that learns how people move by predicting the future state of the body in a learned latent space, rather than reconstructing the raw sensor signal. It is the first application of world models to inertial kinematics, and the first of several pieces of work at TimeTrace on the measurement of biological systems.
The Method
A world model's distinction from a conventional time-series architecture is not cosmetic. A model that learns to predict the raw signal must reproduce the high-frequency noise, the thermal drift, and the device-specific artefacts alongside the biomechanics. Every parameter is shared between signal and noise. When data are abundant the nuisance can be absorbed without cost. Clinical data are rarely abundant, and a model that reproduces artefacts ends up shaped around the sensor, not the disease. Sonata inverts this. By predicting the future state of motion in a learned latent space, the training objective rewards the model for internalising what persists across windows, such as stride cadence, turning mechanics, and the characteristic asymmetries of gait disorder, and discards what does not. The model learns what matters by learning what persists.
Sonata is trained on open source six-axis motion data from nine datasets spanning Parkinson's disease, multiple sclerosis, stroke, cerebellar ataxia, and prospective fall-risk cohorts. Each dataset meets three physics-informed inclusion criteria: full six-axis sensing of acceleration and rotation, trunk placement near the body's centre of mass, and retention of gravity in the accelerometer signal. These constraints exclude much of the available wearable data, but ensure the model is trained on a form of measurement common in healthcare. Across a series of comparisons outlined in the technical report, latent-state prediction yields representations that are consistently stronger for clinical discrimination, fall detection, noise robustness, and data-efficient transfer than a matched raw-signal baseline trained on the same architecture and data. The model is small, matched to the physical dynamics of human motion. It does not need scale; it needs every parameter pointed at biomechanics.
The Deployment
At 3.77 million parameters, the model is small enough to run on the wearable device itself rather than on a remote server. Patient data can stay on the device, with only the outputs that matter leaving it. This keeps sensitive biometric information private by default and reduces the data governance overhead that comes with continuous off-device transmission. Continuous monitoring becomes viable in settings where transmitting raw signal is impractical, including home use and long clinical trials. And because inference happens locally and immediately, the model can support applications where latency matters, such as predicting falls before they happen or flagging clinical deterioration as it unfolds.
The Outlook
Sonata is one public component of a broader programme at TimeTrace Labs: we model the evolution of human physiology (how biology evolves over time and across data types) under the conditions that define healthcare data. Motion is one of the clearest continuous readouts of health, altered in neurological diseases, cardiovascular decline, post-surgical recovery, frailty in ageing, and the functional toxicity of oncology treatment. It sits among several continuous records of physiological state, cardiac rhythms, respiratory patterns, electrical activity of the brain, glucose, voice; each observed through different instruments but governed by related dynamics. Clinical datasets can often be fragmented, small, irregularly sampled, subject to missingness, and governed by regulatory constraints. Our methods are designed around these conditions rather than in spite of them.
That discipline is inherited from the physical and quantitative sciences, where measuring under adversarial conditions is the default, not the exception. This allows our approach to scale: larger cohorts, cleaner data, and richer modalities sharpen the output rather than change the method, but the properties that matter are established at small scale, not rescued at large.
Biology, measured under constraint, yields its structure more honestly than biology averaged over scale.
These are not refinements of existing measurements, they are new ones. Kinematics is one modality and neurology is one application; the same principles extend across data types and therapeutic areas. We have applied our platform across over 7 million people and 30 different conditions.
In practice, this means detecting disease progression in days and months rather than years, stratifying patients in clinical trials with the precision that today's biomarkers cannot deliver, and measuring drug effects on timescales that existing instruments cannot resolve.
This work was done in collaboration with NVIDIA, and is one of several threads of work at TimeTrace Labs we look forward to sharing in due course.
The full technical report is available on arXiv: https://arxiv.org/abs/2604.18058
Please cite this work as:
TimeTrace Labs and NVIDIA. Sonata: A World Model of Human Motion. TimeTrace Labs Blog, May 2026.
BibTeX:
@article{timetrace2026sonata,
author = {{TimeTrace Labs} and {NVIDIA}},
title = {Sonata: A World Model of Human Motion},
journal = {TimeTrace Labs Blog},
year = {2026},
url = {https://www.timetracelabs.com/blog/sonata}
}