A simulator ("digital twin") of a robot is biased, but it is free. Real trials are unbiased but scarce. Our research programme sim2real-inference asks how to spend the scarce ones. This step tested a natural idea: run the real trials on the starting conditions where the twin is predicted to be most wrong, and weight them so the estimate stays unbiased.
On 21 MuJoCo twin-versus-"real" pairs (a pendulum arm, one omitted effect each: friction, delay, sensor noise, gain, payload mass), 1,000 replications per cell:
- With 120 real trials it cut the mean squared error against uniform sampling by 17 to 35% on the three pairs where the twin's error is predictable from the starting state.
- On the weakly predictable pairs it was 8% worse in the median (32% at worst).
- With 30 or 60 trials it lost on 19 and 18 of 21 pairs, because a third of the trials go to learning where the twin is wrong.
- Even a sampler trained on all 4,000 contexts only reaches about 0.8 of uniform's MSE in the median, so this is a refinement, not a way to make the twin pay.
- The nominal 95% interval covered 92.4% at worst, so one pre-registered coverage hypothesis was refuted.
Model-versus-model pairs on one task: the numbers say how the method behaves, not how a real robot would. Code, pre-registration and outcome: research/sim2real-inference/experiments/m5a_active/ and LOG.md. This was written with AI assistance.