Three approaches to closing the sim2real gap on humanoids — and their trade-offs
KyleMorgenstein · x · 2026-09-11
Answering how to minimize the sim2real gap for a policy trained in sim on a humanoid, KyleMorgenstein lays out three general approaches:
- Domain randomization + history: hope the policy learns implicit system identification (bad because conservative);
- Actual system identification (bad because hard);
- Black-box optimizers like actuator nets and ASAP (bad because hard to debug).
A concise map of the current sim2real methodology landscape and its trade-offs.
Related event: Three Routes to Closing the Sim2Real Gap for Humanoid Robots(2 posts)→
More from Embodied
- FADA (CoRL 2026) open-sourced: humanoid robots adapt to new conditions from 2 minutes of experience — GuanyaShi · 2026-09-12
- whurley jokes that Tesla's robotaxi account came in a little strong — whurley · 2026-09-12
- Rumor: OpenAI quietly expanding its humanoid robot project — DJiafei · 2026-09-12
- MolmoAct2 fine-tuned on VLA-Replica, benchmark against OpenAI VLA coming — YuXiang_IRVL · 2026-09-12
- Minotaur raises $100M to crack mixed-case palletizing with defined TAM — MatthewChang · 2026-09-12
- HRI 2027 abstract deadline Sep 11; conference heads to Santa Clara — petitegeek · 2026-09-12