Stanford's MessyMem (CoRL 2026): persistent memory lets mobile robots reuse experience
leto__jean · x · 2026-11-10
Stanford researchers (Anuva Banwasi et al., with Jeannette Bohg's group) unveiled MessyMem, a persistent memory system for mobile manipulation, accepted to CoRL 2026 with a main-conference poster and a SISREN workshop spotlight.
The problem: today's robots treat every task as new — discovering a locked cabinet or an object in a drawer doesn't carry over. MessyMem maintains a spatially grounded 3D scene graph, augments it with properties and outcomes learned through interaction (learning-from-doing), and links keyframes for fine-grained recall so robots reuse experience across rooms and visits.
Results: over a continuous 25-task, 3+ hour simulation, MessyMem hit 80.0% task progress, beating its strongest ablation by 14.8 points and the strongest external baseline by 28.9 points, while retrieving task-relevant evidence from thousands of keyframes over an hour into the past. It was also validated on a real mobile manipulator.
More from Embodied
- Six-legged robot doubles as a crypto hardware wallet, est. $700–$1,000 — R4nsack · 2026-11-10
- Granola's AI hardware Stream starts shipping today — thesephist · 2026-11-10
- Google DeepMind ships MuJoCo 4.0, first major robotics engine upgrade since 2023 — lukas_m_ziegler · 2026-11-10
- EPFL spinout's dual-arm robot AIR hits 95% task success in paid pilots — lukas_m_ziegler · 2026-11-10
- Humble Robotics raises $51M Series A for cabless electric freight robot — lukas_m_ziegler · 2026-11-10
- Boston Dynamics launches Stretch Gen 2: 1,000 cases/hour unloading robot — lukas_m_ziegler · 2026-11-10