SpatioTemporal is building foundation models for motion and intent.

Most AI systems understand the world as static frames, images, or coordinates. That works for recognition, but it breaks down when systems need to reason about how things move, interact, and change over time - especially around humans.

We take a different approach.
SpatioTemporal models motion itself as the primary signal.

Our work focuses on learning a vocabulary of movement - kinematic tokens that compress space and time into reusable elements. These tokens capture how people, vehicles, and objects move, not just where they are. Once motion is tokenised, transformer-style models can be applied to prediction, anomaly detection, and intent inference in dynamic environments.

Our current research model, LSTM-01 (Large SpatioTemporal Model), is trained on motion tokens rather than pixels. It learns patterns of movement, predicts near-term futures, and infers likely intent under partial observability. The architecture is designed for robustness through overlapping recomputation, with an emphasis on real-time and edge-adjacent deployment.

This work is motivated by a growing gap in modern AI: the lack of reliable world models that understand physical dynamics. We believe motion is a missing compression layer in today’s systems, and that modelling it explicitly is critical for safer human-robot interaction, autonomous systems, and real-world AI.

SpatioTemporal operates at the intersection of research and application, publishing openly while building toward practical safety-critical use cases in mobility, robotics, and human-machine teaming.

SpatioTemporal

Org chart

Teams

Offices

SpatioTemporal

Org chart

Teams

Offices

Related companies