Hub

Hub is best understood as a capture network with a data pipeline bolted on, and the ordering matters. The four dataset lines — egocentric, teleoperation, hand tracking, tactile — are the shelf product, each described with concrete instrumentation rather than vague modality labels: stereo RGB with dense metric depth, IMU synced at 200 Hz, 6-DoF VIO, 21-keypoint hand pose with MANO parameters, contact pressure maps with shear and slip events. That level of specificity is what a robotics engineer needs to judge whether the data will align with their rig and their training format. The custom collection flow is where Hub differentiates hardest against Scale AI and Defined.ai. The site publishes an explicit timeline: SOW ingested at 00:00, itemised quote by modality, volume and spec at 00:42, network matching against 10k environments and 150k contributors at 02:15, first samples with frame-level QA in your S3 bucket at 08:30 — with your approval gate before the full run — and delivery complete at 36:00 in MCAP, LeRobot, or a custom format. A 42-minute quote is the number that matters commercially, because it turns a procurement unknown into a line item before your planning meeting ends. The supply side is the moat. 730+ verified commercial sites across 100+ industries — farms, warehouses, construction sites, restaurants, auto repair, hotels, pharmacies, wood workshops, metal fabrication, commercial cleaning — plus 150,000+ active contributors in 150 countries. Environments like a tire shop or a wood workshop are exactly where synthetic and simulated data runs thin, and Hub's pitch is that the long tail of physical work becomes training signal instead of a recruiting problem. Strengths: modality breadth that most data vendors don't attempt (tactile alongside vision and pose, aligned in time), published turnaround numbers rather than sales-deck promises, diversity-cap enforcement as a first-class step, and delivery into your own storage in the format your training stack already reads. Weaknesses and honest boundaries: this is a data supply business, not a modelling one. There is no pre-trained model, no inference endpoint, no annotation-only service for data you already have, and nothing here for text or tabular projects. Custom datasets and samples go through a request and approval flow, so there is a human step before you hold pixels. The site itself is the only source for the pipeline numbers — there is no public changelog, docs portal or integration marketplace surfaced in this pass, so treat the published timings as vendor-stated service levels and confirm them in your own SOW. Where it fits: frontier labs training embodied and manipulation policies, robotics teams that need depth-synced and IMU-tracked teleoperation trajectories, and research groups that need tactile and hand-tracking data aligned with vision but have no capture network of their own. Where it doesn't: anyone whose roadmap is simulation-only, anyone shopping for real-time