Inside the Blue Projects AI Data Collection Studio

Most AI data vendors either run purely in the field, with all the variability that involves, or purely through crowdsourced remote contributors, with all the quality-control problems that involves. Blue Projects runs a third model: a dedicated, calibrated studio for controlled, repeatable capture, backed by a field network for the environmental diversity a studio alone can't provide. This post covers what the studio actually is.

The Space

1,200 sq ft, split into four reconfigurable zones, each built around a specific task category:

• Zone A — Kitchen / daily-task set (320 sq ft) — overhead camera, wrist camera, counter rig, for cooking, cleaning, and household manipulation tasks
• Zone B — Workbench / teleop rig (320 sq ft) — leader-follower robot arms and force sensors, for precision manipulation and teleoperation capture
• Zone C — Warehouse / shelving (320 sq ft) — motion-capture volume and depth cameras, for picking, sorting, and logistics-style tasks
• Zone D — Control room and storage (240 sq ft) — capture PCs, NAS, and a calibration bench

The space is built to a specific baseline: 10–12 ft clear ceiling height for overhead rigs and full-arm teleop clearance, matte non-reflective non-slip flooring with calibration grid markings, dedicated power circuits with UPS backup, wired LAN for sensor sync with a separate redundant line for cloud upload, blackout-capable flicker-free lighting, sound-dampened acoustic treatment for clean audio capture, and badge-entry access control.

What "Calibrated" Actually Means Here

The same capture rig runs every session — head-mounted ego cameras, a motion-capture suit and gloves, a bimanual teleoperation rig with leader and follower arms, depth cameras, and a handheld capture unit — all time-synchronized and distortion-checked. That consistency is deliberate: dataset variance should come from the task being performed, not from which equipment happened to be used that day.

The Task Catalogue

• Manipulation — grasp, sort, stack, fold, assemble, tool-use, pouring, packing
• Whole-body and navigation — staged walking, obstacle negotiation, reaching, load carrying
• Daily-living tasks — cooking, cleaning, laundry, appliance operation, tidying
• Industrial / warehouse — shelving, picking, scanning, inspection on a mock warehouse floor
• Gesture and expression — gesture vocabulary, micro-expression, pointing and signalling
• Speech and interaction — scripted and natural dialogue, instruction-following, multiple languages

If a task isn't on this list, it's a scoping call, not a blocker — the zones are built to be reconfigured around a client's actual brief.

From Session to Dataset

Every session moves through the same four stages: capture (multi-stream recording staged to task spec), sync (all sensor streams timestamp-aligned into one episode record), annotate (pose, depth, segmentation, and action-language labelling), and deliver (exported in RLDS, WebDataset, HDF5, or a custom schema on request).

How Engagements Are Structured

Sample-to-scale, deliberately de-risked at each step: a small free sample (5–10 episodes) to evaluate fit, a paid pilot batch (100–500 episodes) to validate at working scale, then full-volume capture once both sides are confident. On-site walkthroughs and live session viewing are available for scoping visits.

Consent and Compliance

Every session is preceded by logged, double opt-in participant consent. Data handling practices are built around alignment with GDPR, India's DPDP Act, and CCPA/CPRA depending on where participants and clients are based, along with the data-governance expectations found in the EU AI Act and the NIST AI Risk Management Framework. Formal certifications (ISO/IEC 27001, ISO/IEC 42001, ISO 9001) are a work in progress — we'll confirm exact status directly with any client who needs it for their own diligence process, rather than represent it as complete before it is.

Where This Fits In

The studio doesn't replace our field data collection work — it complements it. Controlled, repeatable capture for tasks that need consistency; field sourcing for the environmental diversity a fixed studio can't provide on its own. Most engagements use both.

Frequently Asked Questions