Evaluating a robot policy on a simulator alone is biased; real trials are unbiased but scarce. Can many twin runs plus a few real ones give a valid estimate with fewer real trials? We tested this on 21 MuJoCo model-versus-model pairs (a pendulum arm where the "real" model adds a disturbance and one omitted effect: friction, delay, sensor noise, actuator gain or payload mass, at swept magnitudes; not real-robot data).
- Pairing is the lever. With 30 real runs and 270 extra twin runs, a control variate on twin runs of the same initial context multiplied the effective real sample by up to 3.2x, within 6% of the closed form 1/(1 - rho^2 (1 - n/N)). Twin runs on independent contexts gave 0.96 to 0.98x: no help, a small loss.
- It only pays when the twin is correlated. Below a paired correlation of about 0.75 the gain falls under 2x; at 0.17 it is nil.
- Intervals undercover. Pre-registered hypothesis H1 (95% intervals stay within 2 points of nominal) is refuted:
cvandppi_ppdrop to 88.8% and 86.7% at n = 30. Part of that is the skewed bounded outcome (the real-only interval also falls to 91.2%), so the next step is a finite-sample interval, not a better twin. - Detectability predicts value too, here. We expected a twin that is easy to tell apart from reality to be no worse as a control variate. In this benchmark every effect both shifts and decorrelates, so detectability tracked value (Spearman -0.95) almost as well as correlation (0.999). A fair test needs a pure-offset pair.
Code, pools and the pre-registered log: research/sim2real-inference/ (milestone 3). Produced with AI assistance (Claude Code).