Sim-Trained, Language-Free Arm Policy Wins Third Place at CoRL

carlosdponx · x · 2026-11-17

#3 in the author's CoRL 2025 demo ranking: machenta's one-shot example-based arm policy, which attendees could walk up and try on their own tasks.

The model is trained entirely in simulation (the author is bullish on this) and, he believes, not language-based at all — a direction he finds promising: if anything will beat frontier LLMs at robotics, it will be models that skip language entirely and find gains that way.

Related event: CoRL 2025 demos: consumer robot dog and sim-trained arm policy turn heads(2 posts)→

Original post →

More from Embodied

Embodied channel →