Motion Intelligence.
A new fundamental domain of AI.
Artificial intelligence has learned to read, write, see and speak. Now we’re teaching it to understand movement.
SpatioTemporal is pioneering Motion Intelligence, building foundation models that learn how the physical world moves, how things interact, and what might happen next. Moving towards a deeper understanding of space, time, intent and causality.
We believe understanding motion will become as fundamental to intelligent machines as understanding language has become to AI.
The intelligence we take for granted.
Humanity’s oldest intelligence may be its most overlooked.
Before we learn to read or write, we learn to navigate the world around us.
We learn how to avoid danger, anticipate movement, judge distance, negotiate shared spaces and respond to the actions of others.
A child navigating a busy house. Two strangers passing on a footpath. Thousands of commuters moving through a railway station. These are remarkable demonstrations of intelligence, yet we perform them so naturally that we barely notice.
Over millennia, humans have developed extraordinary instincts for understanding movement, recognising risk and anticipating consequences.
These instincts underpin our ability to live and work together. They help make our homes, workplaces, roads and cities function. And they’re among the most difficult capabilities to give machines.
Movement is information.
Seeing the world is not the same as understanding what it’s doing.
- A pedestrian hesitating at a crossing.
- A cyclist beginning to wobble.
- A vehicle drifting towards another lane.
- A crowd changing direction.
Each movement carries information beyond position, speed or direction. It offers clues about behaviour, intent, emerging risks and possible futures.
Humans recognise these signals instinctively. We don’t simply observe movement. We interpret what it means.
Artificial intelligence has made extraordinary progress in understanding language, images, speech and video.
We believe motion deserves the same foundational treatment.
Motion Intelligence is a new domain of AI dedicated to learning the structure and meaning of movement.
Not simply recognising what exists, but understanding how the physical world behaves.
The grammar of movement.
Compressing space and time into motion tokens.
Language models learn from words and the relationships between them. Our foundation models learn directly from movement.
SpatioTemporal transforms continuous motion into compact, discrete representations called motion tokens. Each token captures a short pattern of movement. Combined over time, these patterns form a learned vocabulary of behaviour.
Acceleration. Hesitation. Turning. Yielding. Changing direction. Together, they provide the foundations for recognising more complex movement and anticipating what may happen next.
This approach allows Motion Intelligence to operate on movement itself, rather than depending on the original images or sensor data used to observe it – creating a path from understanding individual motion to understanding interactions between multiple moving agents.
One foundation model. A growing understanding of a world in motion.
One intelligence. A world of applications.
Understanding movement changes what machines can do.
Motion Intelligence isn’t limited to robots, vehicles or any particular industry. Movement is universal. The intelligence required to understand it has applications wherever the physical world is observed, simulated or navigated.
Understand.
Read the movement of the real world.
Interpret behaviour and interactions from cameras, sensors and other observations.
Move beyond identifying objects towards understanding how they’re behaving, how they relate to one another and which movements may matter.
From workplaces and infrastructure to transport networks and entire cities.
Simulate.
Create worlds that behave more like ours.
Generate realistic, varied movement and interactions in virtual environments.
Enable simulations to represent not only different places and conditions, but different behaviours, intentions and responses.
From digital twins and autonomous-system training to virtual worlds and complex multi-agent simulation.
Anticipate.
Understand what may happen next.
Recognise emerging risks, predict possible movements and provide intelligence that helps machines make better decisions.
Give autonomous systems the ability to respond to behaviour, rather than simply react to position.
From robots that read the room to cars that read the road.
From Research to Reality.
A new domain of AI. Already taking shape.
SpatioTemporal has moved beyond the initial thesis. We’ve built and trained our first Motion Intelligence foundation models, developing a learned vocabulary of 10,000 motion tokens.
We’ve demonstrated compact inference, including models capable of running locally in a web browser.
We’ve begun testing how motion understanding changes the behaviour of larger Physical AI systems.
In controlled NVIDIA Cosmos 2 simulations, introducing Motion Intelligence reduced robot-human collisions from 24% to 2% in our most challenging test scenario. We observed earlier yielding, and smoother shared-space negotiation.
And in Cosmos 3 trials, we’ve shown that adding our Motion Intelligence foundation model to NVIDIA’s flagship robotics world model, we improve understanding of complex moving systems by nearly an order of magnitude – and give Cosmos a completely new modality: motion.
When machines understand more about movement, they can make better decisions about how to behave.
We’re continuing to develop the models, strengthen the benchmarks and explore applications across robotics, simulation and physical-world intelligence.
Intelligence for a world in motion.
The next generation of machines must learn to share our world.
Human civilisation is built on movement. People, vehicles, goods and entire societies move through shared environments, continuously adapting to one another.
Our homes, workplaces, roads and cities depend on countless unspoken interactions.
We anticipate. We negotiate. We yield. We cooperate. Most of the time, without even thinking about it.
As intelligent machines become increasingly capable, they will enter these same environments.
They will work alongside us, move amongst us, observe our infrastructure and participate in the systems that support modern society.
Their success will depend on more than their ability to see, reason or perform tasks – they will need to understand the dynamic world we inhabit.
To recognise when to act, when to wait, when to intervene and when to give way. To anticipate how their actions may affect others. And to operate in ways that make the human world safer, more productive and more natural to share.
This is why we’re building Motion Intelligence. Not simply to make better robots. But to establish a fundamental capability for the next generation of intelligent machines.
Understanding movement is the beginning.
Understanding its consequences is the ambition.
