Skip to content
Bud's Wiki

Gianni Hodel — microSLAM: monocular SLAM for scalable egocentric robot data

TL;DR

Section titled “TL;DR”

Gianni Hodel (microagi, Zürich) announces microSLAM, a monocular SLAM system built by microagi’s Zürich CV team with ETH Zurich’s Computer Vision and Geometry Lab. Reported as first on the LaMaria benchmark for monocular SLAM and competitive with systems carrying extra cameras and IMUs, while running on a single RGB stream. Product framing: robot learning is bottlenecked on data; a head-mounted camera on a worker records how a plant actually operates, but the footage is only useful if geometry can be recovered from one moving camera in a dynamic scene. Solving it monocular is what makes head-worn glasses (rather than rigs) the scalable data-collection form factor, and microSLAM converts the footage into interactive 3D environments where behavioral foundation models can be trained.

Key claims

Section titled “Key claims”
  • microSLAM is a monocular SLAM system running on a single RGB stream, developed jointly by microagi’s Zürich computer vision team and ETH Zurich’s Computer Vision and Geometry Lab [Post].
  • microSLAM ranks first on the LaMaria benchmark for monocular SLAM [Post].
  • microSLAM’s monocular result “holds up against” systems that carry additional cameras and IMUs [Post] — an approximate rather than a metric comparison in the tweet.
  • The stated product thesis: monocular geometry recovery is what makes head-worn glasses (not multi-sensor rigs) viable as a scalable data-collection form factor for worker footage [Post].
  • microSLAM turns egocentric worker footage into interactive 3D environments intended as training substrate for behavioral foundation models [Post].

Method

Section titled “Method”

The tweet does not describe microSLAM’s algorithm. Framing it as “monocular SLAM from a single RGB stream” that recovers geometry in dynamic scenes places it in the same slot as WildPose (WildPose: A Unified Framework for Robust Pose Estimation in the Wild) and the feed-forward reconstruction line (VGGT-Ω: Scaling Feed-Forward Reconstruction for Static and Dynamic Scenes, Depth Anything 3: Recovering the Visual Space from Any Views), but no architectural details, backbone identity, or scale-recovery mechanism are disclosed. The first reply on the thread flags the standard monocular ambiguity — a single moving camera recovers geometry only up to an unknown scale — and asks where microSLAM gets its metric scale [Post reply]; the tweet does not answer.

Results

Section titled “Results”
  • First place on the LaMaria monocular-SLAM benchmark [Post]. No per-benchmark numbers, no comparison table, no ATE/RPE figures in the tweet.
  • Qualitative comparison to multi-camera + IMU systems is asserted but not quantified [Post].

Why it’s interesting

Section titled “Why it’s interesting”

Two positions this tweet stakes out on the wiki. First, it is a second 2026 monocular-camera-SLAM datapoint alongside WildPose: A Unified Framework for Robust Pose Estimation in the Wild — both target the dynamic-scene monocular case that the Pose Estimation and Motion Capture cluster previously had only WildPose in. Second, it is an infrastructure-play instance of the “labels-on-real-egocentric-video is the bottleneck” thesis that Synthetic Training Data tracks: instead of scaling per-worker sensor rigs (ActiveMimic ActiveMimic: Egocentric Video Pretraining with Active Perception, HumanScale HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining, EgoScale EgoScale: Scaling Dexterous Manipulation with Diverse Egocentric Human Data) or paid-collector networks with wearable devices (RekaDaily-10k RekaDaily-10k: Collecting 10,000+ Hours of Egocentric Household Manipulation Data, EgoSuite-Open100K Lightwheel AI open-sources EgoSuite-Open100K — 100K-hour egocentric human dataset with hand and body pose), microagi bets that a strong-enough monocular geometry stack lets consumer glasses become the collection form factor. Complements the “video-model-as-data-multiplier” recipes (Worlds in One Demo: A Synthetic Data Engine for Learning Open-World Mobile Manipulation WANDA, Atlas: A World Model for Spatial Intelligence Atlas) by targeting the upstream geometry-recovery step rather than the downstream video generation. Also relevant to the “interactive 3D environments as training substrate” line the World Foundation Models cluster tracks — microSLAM’s stated output shape is closer to Atlas / SimFoundry twins than to raw reconstruction.

See also

Section titled “See also”