Sim-to-real lessons: RL-trained robots vastly outperform behavior cloning, UW work shows
chris_j_paxton · x · 2026-10-12
A robot trained with RL at UW shows impressive recovery behaviors, highlighting the day-and-night difference between RL-trained and behavior-cloned models. Still, scaling RL to new use cases is hard.
Key ingredients for making sim-to-real work, per the author:
- Assets must look good and simulate well: visually simple parts (no textures, simple materials) make realistic rendering easy; the reflective board was the only tricky object and needed taping to match real-world rendering.
- Physical assumptions matter: recovery behaviors are too wild if the board isn't bolted down, showing how strongly the policy assumes a fixed setup.
More from Embodied
- He built a cat activity dashboard with Opus 5.5 by sniffing Ring, Petlibro and Whisker app traffic — indragie · 2026-10-12
- Maker's 3D-printed robot comes to life with servo legs and custom frame — shekitup · 2026-10-12
- Maniformer paper cuts egocentric hand tracking error to 4.29mm for robot learning data — heyshrutimishra · 2026-10-12
- First look at ByteDance's new World Model: text-to-navigable 3D worlds with real-time editing — Uncanny_Harry · 2026-10-12
- Humanoid startup UMA opens its lab to Numerama for a deep dive on the US-China robot race — RemiCadene · 2026-10-12
- Builder uses Codex and a Raspberry Pi 5 to make his ceiling AC iPhone-controllable — paw_lean · 2026-10-12