Three Months Into Our Innovate UK Project
Where our research on causal world models stands, what we've learned, and what comes next in the new year.
We're three months into our Innovate UK AI Champions project.
In September we shared that we'd won Innovate UK backing to work on causal world models. The problem we set out to solve is transfer: an AI system learns a task in one place, moves somewhere else, and stops working reliably. We argued that this is baked into how these models are trained, and that the fix is to teach a model the structure of a task separately from the habits of the place it learned it.
Since starting the research, we have three things to show:
- A way of testing cause and effect that we can trust.
- Early evidence that a model can learn cause and effect rather than coincidence.
- A robotics setup that measures exactly how badly a trained system breaks when its world changes.
First, a fair test
You can't tell whether a model understands cause and effect unless you can ask it a "what if" question and check the answer. What if that object had been pushed? Would the two balls still have collided?
To check an answer, you need two versions of the same scene that differ by exactly one change. The standard benchmark, CLEVRER, couldn't give us that, so we built our own simulator. It plays out a scene, then forks it: one future where nothing changes, and one where a single object is removed or nudged. Every frame, every contact and every question comes from the same source, and nothing in it is generated by the model being tested.
This is the foundation for everything we are working on throughout the project.
Cause, not coincidence
With a fair test in place, we asked the question that matters. Does a model that's told about the intervention predict differently from one that isn't?
It does. On scenes it had never seen, our model answered "what if" questions correctly 79% of the time. A twin model shown the same scenes, but kept blind to the intervention, managed 61%. The gap is the part the model has learned about cause and effect.
We pushed on this to make sure it wasn't a lucky shortcut. Feed the model the wrong intervention and its accuracy collapses. Reverse the direction of a push and it gets worse in the way physics says it should. The model is using the intervention, not memorising the scene.
These are fresh scenes from the same simulator, not the real world. Real-world transfer is the goal, and we're working towards it.
Watching a robot break
The second half of the project moves from prediction to action. If world models are becoming the way robots and industrial systems decide what to do, we need to watch the transfer problem happen to a robot.
Our testbed is a simulated robot arm learning jobs from recorded demonstrations, the way most industrial robots are trained today. Then we change the conditions. Different operators recorded the demonstrations. Commands arrive with noise, or smoothed, or with the camera rotated a couple of degrees.
A recorded demonstration, replayed in simulation: the arm pushes the blue cube onto the target. Demonstrations like this are what the arm learns from.
The result is the transfer problem in miniature. Under the conditions it trained on, the robot succeeds about 72% of the time. Under conditions it hasn't seen, that falls to about 54%. Heavy smoothing of its commands nearly halves its success rate. Nothing about the task changed. Only the surroundings did.
That gap is what we're trying to close. We now have a reproducible way to measure it, with every attempt and every failure kept on record. Our first-pass fixes, which added more context or more varied training data, didn't close it. Better imitation of the demonstrations didn't mean better control. That's a useful negative result, because it tells us the answer isn't simply more data.
Testing in drone navigation
This month we began testing the research on drone navigation. The aim is to try a stronger class of policy, then move the same measurements onto new tasks to see whether what the drone learns under one set of rules carries over to another.
Left: four simulated runs, each with a new layout of boxes and low objects, as the drone flies its route between the lettered waypoints. Right: the demonstrated experiment, where the drone flies a series of moves after training in simulation.
With this in place, a robot could check its own forecast before it moves, and hold back when the world has shifted beyond what it knows. That matters most for platforms that work outside a controlled space, where the surroundings change constantly and a wrong move means a breakage or an expensive mistake.
Drone navigation is a starting point. In the coming months we'll take what the drone has learned further, towards robots that can perceive the space around them and find their way through it, and we'll try the same approach on other kinds of robots. We're excited to keep building, and we'll share more as the work progresses.
Run the decision before you commit to it.
Bring a decision. We'll discuss how to model it and what data a simulation would need.
Talk To An Expert