MicroSLAM tops LaMaria benchmark, turning cheap egocentric video into robot training environments

lukas_m_ziegler · x · 2026-09-10

microagi's Zurich CV team released MicroSLAM, monocular SLAM from human egocentric video. It ranks first on the LaMaria monocular SLAM benchmark and second among all SLAM systems, including multi-camera+IMU setups. The pitch: monocular egocentric video is the cheapest robotics data — one person, one camera — and MicroSLAM converts it into interactive 3D scenes and RL environments where behavioral foundation models can train, so robots learn to navigate real facilities from footage a human recorded. Existing SLAM breaks under dynamic egocentric video's collapsing static-world assumption and scale issues.

Original post →

More from Embodied

Embodied channel →