Sitzmann clarifies: only learned systems generalize to unknown in-the-wild objects

vincesitzmann · x · 2026-09-20

Responding to the Wiggle and Go authors, Vincent Sitzmann clarifies his ECCV remark: if you know the rope will be attached to your robot and can make a bunch of assumptions, system identification and physics simulation work reasonably well.

His actual point: for unknown, in-the-wild objects, the only systems that can hope to generalize are learned systems — whether video world models or agentic trial-and-error systems — and any specialized model will perform worse in that setting.

Related event: Robot Learns Rope Dynamics with a Single Wiggle, Sparking Debate(3 posts)→

Original post →

More from Embodied

Embodied channel →