MessyMem (CoRL 2026): persistent memory for robots via 3D scene graphs and VLM analysis

leto__jean · x · 2026-09-17

MessyMem, accepted to CoRL 2026, gives mobile manipulation robots learning-from-doing persistent memory: remembering where objects are or which drawer is locked across large spaces, interactions, and fine-grained detail.

The underlying 3D scene graph isn't just a map — it's updated by what the robot does. A VLM reads each interaction and records what it revealed, and linked keyframes preserve details the graph never stores, like the name written on a cup.

Related event: Stanford's MessyMem Gives Robots Persistent Cross-Scene Memory(3 posts)→

Original post →

More from Embodied

Embodied channel →