Colossio
Blog · Oct 2026

Measuring the gap from simulation to steel

Simulation is where policies learn. Hardware is where they are judged. The gap between the two should be a number, not a feeling.

Policies trained in simulation see millions of episodes a real robot never could. But a simulator is a model, and every model is wrong somewhere: friction, latency, backlash, sensor noise.

Pair every sim metric with a real one

For each task we track the same success metric in simulation and on hardware, on the same scenarios. The ratio between the two is the gap, and it is tracked over time like any other metric.

Close it from both sides

Some of the gap is fixed by better simulation: identified actuator models, measured latencies. Some is fixed by hardware that behaves more like the model. Owning both ends means choosing whichever fix is cheaper.