Stanford's MessyMem gives robots long-horizon memory: 80% progress over 25 tasks spanning 3+ hours
leto__jean · x · 2026-09-17
Stanford researchers introduce MessyMem, a persistent memory system for mobile manipulation accepted to CoRL 2026, which lets robots learn from doing and reuse that knowledge instead of re-exploring each task.
- Architecture: a globally queryable, spatially grounded 3D scene graph, updated with properties and outcomes learned through interactions, plus linked visual keyframes for fine-grained recall.
- Ablations: removing keyframes drops cluttered pick to 58%; removing interaction analysis drops locked cabinets to 43%; combining scene graph + interactions + keyframes reaches 84% and 99%.
- Main result: 80.0% task progress across 25 continuous simulation tasks spanning 3+ hours, beating the strongest ablation by 14.8 points and the best external baseline by 28.9 points, retrieving relevant evidence from thousands of keyframes up to an hour old.
Validated both in simulation and on a real mobile manipulator.
Related event: Stanford's MessyMem Gives Robots Persistent Cross-Scene Memory(3 posts)→
More from Embodied
- Drone-captured 400-ft guyed tower rebuilt as 3DGS via Metashape and LichtFeld Studio — janusch_patas · 2026-11-02
- Physical AI's real story: a 4-layer map of 15 companies behind humanoid robots — pdamodaran · 2026-11-02
- Autonomous's Harness Pro uses an ESP32-P4, not a mini PC, to monitor AI coding agents — dee_hw · 2026-11-02
- KAIST's QCP lets robot policies control execution quality in real time — kaist-ai · 2026-11-02
- ANYbotics begins customer deliveries of ANYmal X, first legged robot certified for explosive atmospheres — lukas_m_ziegler · 2026-11-02
- InSpatio opens China's first embodied simulation data center targeting 1M hours/month — 智东西 · 2026-11-02