GroundedSLAM is an unparalleled technique demonstrated by the RoboCap product that aims to be the the visual-inertial geometric layer of Grounded Superintelligence’s Grounded API. It estimates the motion of a camera rig through a metrically scaled three-dimensional world, while the broader API combines that trajectory with stereo depth and, for RoboCap’s original human-manipulation use case, hand tracking. For drone analytics, this means GroundedSLAM could provide the coordinate transform that relates observations across frames, whereas DVSA’s detectors, trackers, semantic models, and retrieval services would still determine what each observation represents. RoboCap is a useful architectural precedent for converting synchronized commodity-grade imagery and inertial measurements into structured spatial data, rather than a ready-made drone product 

RoboCap itself is a 250-gram, six-camera, dual-IMU head-worn capture rig developed with BitRobot. Its six global-shutter RGB cameras record 1920 by 1080 images at 30 Hz; two IMUs sample at 200 Hz; two overlapping fisheye pairs provide calibrated stereo baselines; and two lateral cameras extend visual coverage. The cameras share a hardware trigger, camera-to-camera offset is reported as zero at timestamp granularity, and per-device calibration reduces the residual camera-to-IMU timing offset to no more than 3 ms. These properties, rather than the cap form factor, are central to the reported metric estimation: simultaneous exposures limit motion-dependent inter-camera inconsistencies, global shutters avoid rolling-shutter geometry during rapid motion, stereo supplies directly measured scale, and the IMUs constrain rotation and acceleration when image evidence weakens. 

The Robocap technical report paper describes one multi-camera visual-inertial estimator that accepts all available cameras rather than running independent trajectories and reconciling them afterward. A GPU-accelerated patch-based optical-flow front end supplies visual tracks to a square-root sliding-window estimator. Overlapping cameras establish stereo landmarks with metric depth from their first observation, while non-overlapping cameras track separate landmarks monocularly and continue constraining motion when the main stereo views are occluded. Appearance-based place recognition identifies revisits. During capture, pose-graph loop closure corrects the causal trajectory; after capture, an offline visual-inertial bundle adjustment optimizes all keyframes, merged landmarks, inertial constraints, and relative-pose constraints associated with recognized revisits.