AI simulation infrastructure: what you need to train, test and trust a simulated world

"AI simulation infrastructure" covers a lot of ground, from data centres running robot training to a browser game that teaches a safety procedure. Underneath, the same parts keep showing up. We have built several simulations in the lab, and these are the layers we have found we need.

The layers

• The world: a model of a real place. We capture ours as Gaussian splats from drone or phone video, which is quick and photo-real. A housing site took one 44-second drone pass to place 113 plots in 3D.
• Physics and state: what moves, what holds pressure, what swings, and the rules that decide what happens next.
• Scenarios: repeatable starting conditions, so two runs can be compared fairly.
• The actors: trainees, operators or AI agents acting inside the world.
• Measurement: what is scored, and whether that score actually means anything.

Measure the right thing

In our crane signals trainer, the load hangs on a physics joint and swings like a real pendulum. We set out to fault trainees whose load swung too much, and twice we measured the wrong thing. Peak swing was 9.8 degrees whether the move was good or bad. What mattered was when the move stopped: a traverse that ended on a full pendulum period left 1.0 degree of swing, while one that ended half a period out left 18.3. The simulation was easy to build. Knowing what to score took the experiments.

Record unsafe acts instead of preventing them

In our isolation trainer, valves take sustained effort, spring back if released early, and the line only bleeds down once both ends are shut. Unsafe acts are recorded as faults, not blocked. So a trainee can complete five of five steps and still fail. A simulation that blocks every mistake teaches the interface. One that records them teaches the job.

Repeatability and evidence

For operations twins, we prefer deterministic simulation: the same inputs give the same outputs, so a result can be checked, argued with and reproduced. AI then sits on top, explaining and searching, rather than quietly changing the numbers underneath. We wrote about that split in digital twins as command interfaces and what makes a digital twin useful to operators.

Running it where people are

All of our trainers run in a web browser on ordinary hardware. That rules out some physics, but it means a team can try a scenario on a work laptop with nothing to install. The heavy jobs, such as capture, reconstruction and AI reading, run once on one desktop graphics card. That is the infrastructure most organisations actually need first: not a cluster, but a reliable path from a real place to a scenario people can run and a score they can trust.