MIT builds spatiotemporal memory for robots that recalls past actions in plain language

MIT researchers led by AeroAstro associate professor Luca Carlone published a long-term memory framework letting mobile robots build rich, queryable mental models of large-scale environments. A robot can answer natural-language questions like "where did we leave the half-assembled component last night" and act on the answer in real time. The method fuses 3D scene-graph representations with language descriptions accumulated over time, then exposes them as a language-based map the LLM can reason over. It beats state-of-the-art baselines on accuracy while running fast enough for live deployment, with applications extending to AR maintenance and commuter wayfinding.

View full digest for June 17, 2026