Grounded API launch — hand tracking, depth, and SLAM from one RoboCap recording

TL;DR

Grounded SI (GroundedSI) launches Grounded API, a hosted data-enrichment product for robot-learning teams. From one RoboCap recording it emits metric 3D hand tracking, metric depth maps, and camera trajectory + SLAM at claimed sub-cm accuracy, viewable interactively in Rerun (RRD export) or exportable into downstream robotics stacks. The company positions this as a productization of internal robot-learning research and states two benchmark claims: multi-view hand tracking trained on a proprietary pretraining corpus that “tops all of Meta’s released datasets,” and a GPU-accelerated SLAM that beats ORB-SLAM3 and NVIDIA’s cuVSLAM on several public benchmarks. Ships alongside a BitRobotNetwork-built RoboCap (8 synced head + wrist cameras, wireless, worker-scale ergonomics) and LeRobot / Hugging Face integration for policy training.

Key claims

One RoboCap recording, processed by Grounded API, yields metric 3D hand tracking, metric depth maps, and camera trajectory + SLAM with sub-cm accuracy [Post].
Outputs are inspectable interactively in 3D via Rerun (.RRD export) and consumable in LeRobot-compatible workflows for large-scale policy training [Post].
Grounded’s multi-view hand tracker is trained on a proprietary pretraining corpus that “tops all of Meta’s released datasets” — no numbers or benchmark named in the tweet [Post].
Grounded’s GPU-accelerated SLAM beats ORB-SLAM3 and NVIDIA cuVSLAM on several public benchmarks — no benchmark names or metrics in the tweet [Post; linked blog gsi.company/blogs/the-robocap-for-robot-learning was unreachable at filing time].
RoboCap (co-designed for a year with BitRobotNetwork) carries 8 total synced head + wrist cameras, no wires, “ergonomic for extended real-world collection,” and is priced to deploy at scale [Post].
Target users: (a) data vendors scaling collection with research-grade enrichment, (b) enterprises building proprietary datasets for their environments, (c) researchers running controlled real-world experiments [Post].

Method

Not described. The tweet frames Grounded API as a hosted service that ingests a RoboCap recording (8-camera synchronized head + wrist rig) and returns multi-view 3D hand pose, metric depth, and camera trajectory + SLAM as separate output streams, plus a .RRD bundle for Rerun. The linked research blog (gsi.company/blogs/the-robocap-for-robot-learning) contains the algorithmic detail per the announcement but was unreachable at filing time. No architecture, training-data composition, or benchmark numbers are stated in the tweet itself.

Results

Two headline benchmark claims — SOTA multi-view hand tracking on the authors’ proprietary pretraining corpus (comparison target: “all of Meta’s released datasets”) and GPU SLAM ahead of ORB-SLAM3 + cuVSLAM on unnamed public benchmarks. Both are qualitative in the tweet; no metrics, benchmark names, or ablation are provided in the announcement itself [Post].
Sub-cm accuracy is claimed for the combined hand + depth + SLAM output on a single RoboCap recording [Post] — again unquantified.

Why it’s interesting

Two positions for Luma. First, this is the productized end of the same monocular-geometry-as-data-infra bet that Gianni Hodel — microSLAM: monocular SLAM for scalable egocentric robot data (microSLAM, ETH CVG + microagi) and WildPose: A Unified Framework for Robust Pose Estimation in the Wild argue at the research layer — but with a specific hardware form factor (RoboCap, 8-camera synchronized head + wrist, wireless) rather than head-worn glasses or per-camera rigs like Ego-OSCAR: Egocentric Open source Stereo CAptuRe System and Ego1 (Ego1 — egocentric capture headset for Physical AI (General Intelligence Labs)). If the ORB-SLAM3 / cuVSLAM claim holds up numerically, it is a direct progress marker over SLAM on Robocap via cuVSLAM, packaged with rattler-build (Claude Code build report) (Rerun’s cuVSLAM-on-Robocap integration, which is visual-only and acknowledges loop-closure limits). Second, the multi-view-hand-tracker + LeRobot export path fits the “curated pose stack as the load-bearing upstream stage for VLA pretraining” argument that Compute-Optimal Scaling Laws for Human Motion Generation makes for motion generation and that HumanEgo: Zero-Shot Robot Learning from Minutes of Human Egocentric Videos makes for manipulation policy — Grounded is offering the enrichment layer as an API rather than as a paper. All numeric claims are asserted without a benchmark until the linked blog becomes accessible.