AIMessyMem aims to give robots persistent memory
The team behind MessyMem says its 3D scene graph changes as a robot acts, while linked keyframes preserve details such as the name on a cup.
TLDR
MessyMem’s team describes a robot-memory system combining a 3D scene graph, interaction analysis and linked keyframes. A vision-language model records what each interaction reveals, while the keyframes retain details the graph does not store. The team reports 58% on cluttered picking without keyframes and 43% on locked cabinets without interaction analysis. With all three components together, it reports 84% on cluttered picking and 99% on locked cabinets.
Combined views
477
1 Source, first seen 1d ago
MessyMem aims to give robots persistent memory
The team behind MessyMem says its 3D scene graph changes as a robot acts, while linked keyframes preserve details such as the name on a cup.
TLDR
MessyMem’s team describes a robot-memory system combining a 3D scene graph, interaction analysis and linked keyframes. A vision-language model records what each interaction reveals, while the keyframes retain details the graph does not store. The team reports 58% on cluttered picking without keyframes and 43% on locked cabinets without interaction analysis. With all three components together, it reports 84% on cluttered picking and 99% on locked cabinets.