VLA data operations outsourcing for AI and robotics, from $1,000 a month, all in

Sourcefit’s VLA data operations rates are $1,000 a month for entry level, $1,355 for mid level and $1,710 for senior VLA data analysts, per full-time employee working from home. Office-based staff add $100 a month. Each rate is all in: salary and benefits at cost plus the $395 management fee. There are no setup, recruitment or exit fees. Where the work suits it, output-based pricing is also available, built on the same transparent framework: we show how the unit price is calculated, so you know exactly what you are paying for.

No. The fee is already inside every rate on this page and on every quote. A $1,000 rate is about $605 of salary and benefits at cost plus the $395 fee. You will never see the fee added to a rate you were quoted.

Legal employment and compliance with an HR team on the ground, recruitment, an i5 computer with dual monitors and headset, a 24/7 IT help desk, a secure managed network, our security certifications and insurance, training and quality, client success management and the Knit workforce platform. For office-based staff it also covers the seat, office network and utilities.

Most VLA data operations roles are recruited, interviewed and onboarded in two to four weeks. Calibration on your taxonomy adds one to two weeks before production volume. You interview and choose the VLA data analysts.

That is your choice. Most clients manage the team directly: the VLA data analysts report to your managers and work inside your processes and tools, with Sourcefit handling employment, HR, equipment, IT and quality support behind them. If you would rather not manage day to day, our Sightline model builds an autonomous team with its own Sourcefit team lead and quality layer that delivers against agreed SLAs, while still reporting directly to you and working to your KPIs, not ours. Sightline is quoted separately.

Yes. VLA data analysts work in your annotation platforms, project dashboards and data formats, such as Label Studio, CVAT, custom tooling, RLDS or LeRobot exports, under your access rules. You keep ownership of every account, and access is revoked the day someone leaves.

Sourcefit holds ISO 27001:2022 and ISO 27701 certification directly, is PCI DSS validated, has a SOC 2 report, signs HIPAA business associate agreements where needed and operates GDPR-aligned data handling. Home-based staff work on company-managed computers over a VPN with traffic routed through Cloudflare.

The team is yours. If you decide to move them to your own entity or take them with you, they go with you. There is no transfer fee, buyout or conversion charge.

Vision-language-action models learn from video of a robot or a person performing tasks, paired with language instructions and action labels. Sourcefit teams turn raw episodes into model-ready data: segmentation into sub-tasks, action and object labels, instruction and caption writing, success, partial and failure tags, idle and event detection, and QA. We annotate and validate data you collect; we do not operate robots or run teleoperation.

A dedicated analyst at a flat monthly rate with salary at cost and the management fee shown separately, or output-based pricing per episode for defined, high-volume work, built on a transparent framework with the calculation shown so you know exactly what you are paying for. Price per episode depends on video length, number of cameras and modalities, action granularity and QA layers.

Taxonomy and action-space design, embodiment definitions, collection on your hardware and policy evaluation. Everything downstream of collection, segmentation, labeling, language, outcome tags and QA, works well with a dedicated team trained on your taxonomy.

Inter-annotator agreement against a gold set you own, with 95 percent or better on action labels as a common production threshold; language-to-action alignment checks; segmentation consistency; and per-episode scoring so weak sources can be excluded. New analysts complete a calibration battery before production.

Instructions are written against the video rather than copied from operator notes, with multiple phrasings per instruction type and atomic step boundaries, so the model learns to follow language instead of ignoring it. Your style guide for instructions is part of onboarding.

Yes. Outcome tags for success, partial and failure, missed grasps, wrong-object actions, contact moments and recovery segments are standard, because models trained only on successes do not learn recovery.

Yes. Analysts work in your annotation platform and dashboards and deliver in your format, including RLDS and LeRobot-style exports, with timestamp synchronization across modalities and embodiment metadata preserved.

Access only to your systems under named accounts, company-managed computers over a VPN, ISO 27001:2022 and ISO 27701 controls, IP assignment in the terms, and deletion terms at the end of a project. Consent documentation for human-demonstration footage stays with you.

VLMs pre-label and flag low-confidence frames cheaply, and they cannot say why an instruction and a scene do not match. Humans catch mismatched language, verify outcomes and review edge cases. Our analysts work AI-assisted queues, and the human share shifts to the hard cases as your automation improves.