Simulator misspecification breeds overconfidence: sim-to-real study on SPEED+

Nikolay Dolgov · hf · 2026-11-19

This paper studies directional robustness on the SPEED+ sim-to-real benchmark. A simple monocular pose estimator trained on synthetic spacecraft images transfers poorly to the real world, showing overconfident pose errors and uneven gains from backbone adaptation and photometric augmentation under different lighting. The takeaway: simulator misspecification can yield unreasonably high confidence that standard domain-randomization tricks don't fix.

Original post →

More from Embodied

Embodied channel →