News Research
Fitting motor curves on the bench
The least glamorous work in sim-to-real may be the most valuable: measuring the real motor and putting it in the digital twin.
"All models are wrong. Some are useful."
A digital twin is only as good as its motor model. If the simulated motor accelerates faster than the real one, every policy trained on it will be wrong in the same way.
Measure it on the bench
We fit the motor's behavior on the bench, low level, and put those curves into the twin. Even buyers of expensive humanoid and quadruped robots still have to fit these curves themselves. On GOAT RACER it is already done. We think it is almost the most valuable thing in the stack, because it is what makes sim-to-real work.
A sensored motor
GOAT RACER PRO replaces the stock motor with a sensored motor for closed-loop RPM control. That is where its precise low-speed control comes from, and why experiments on it repeat.
Real to sim, and back
"The real point of sim to real is taking that run and refitting your model. You come up with your own digital twin."
Every real drive is logged. Real and simulated observations pass through the same ONNX network, so we can compare them directly. Then the logs flow back to refit the twin. Sim to real, then real to sim, every lap.

Draft note: add a real motor curve plot (sim vs bench) if we have one.
Small changes add up
- The RealSense D436 adds about 10 degrees of field of view over the D435, which better matches the simulated camera.
- The battery's low-voltage cutoff moved into software, which improved run time.