Capturing synchronized expert demonstration data for Physical AI.
We record master practitioners at real work, and turn those sessions into training data for robot foundation models. Every session is consented, fairly compensated, and auditable.
Six aligned signals from one real episode.
Every stream is synchronized to one timeline and delivered with task metadata, practitioner consent, and episode-level provenance.
- Scene and wrist-perspective videoThe full visual field of the work.
- Body and hand movementFull-body and per-finger kinematics during real work.
- Gaze and attentionWhere the expert looks before they act.
- Contact and force contextWhat the hand applies at the point of contact.
- Audio and expert explanationReasoning spoken aloud, time-aligned to the action.
- Outcomes, errors, and recovery labelsExpert-graded results with corrections marked.
EPISODE AB-HOSP-0114 TASK Fine-dining plating, four covers DURATION 14m 06s STREAMS 6 synchronized · one clock RECOVERY 2 events marked PROVENANCE measured / derived, per value CONSENT practitioner-licensed EXPORTS LeRobot · RLDS · OXE
Built from real work, with the people who do it.
We record skilled work as it actually happens, alongside the people who do it. Each session becomes clean, documented training data, delivered in the formats robotics teams already use. Genuine expertise, not staged or crowd-collected demonstrations.
We focus on domains rich in scarce expertise.
Fields where deep, hard-won skill is concentrated in a few master practitioners.
Hospitality
Skilled kitchen, service, and housekeeping work in professional operations.
Heritage & craft
Master artisans and heritage construction: textiles, jewellery, traditional printing, masonry, and fine finishing.
Aviation MRO
Certified inspection, maintenance, and recovery procedures.
Surgical & clinical
Programme in design with hospital partners.
How robotics teams use Anthrobase data.
Our datasets support work across the full lifecycle of a robot policy, from its first training through to evaluation in the real world.
Pre-training
Generalist policies pre-trained on long-horizon demonstrations from real workplaces, not lab setups.
Post-training
Expert trajectories with graded outcomes and recovery behaviour, the signal imitation and RL depend on.
Simulation
Recordings that serve as physical ground truth, narrowing the gap between simulation and reality.
Evaluation
Expert-graded sessions that benchmark a policy against the standard a master practitioner sets.
Every dataset is documented and licensed with care.
Each dataset ships with its provenance, consent terms, and licence scope documented. The full specification, an illustrative EDS schema, and an episode manifest are shared with labs under NDA, and live evaluation samples will follow the first capture sessions. The first licensed dataset release is planned for H2 2026.
We work with master practitioners, not crowd workers.
A small network of genuine experts, not anonymous crowd labour. Every session is consented, every practitioner fairly compensated, every record auditable.
Consent
Explicit, scope-limited, opt-in by default. Any practitioner can withdraw their data and request deletion at any time.
Compensation
Above-market session rates plus a share of licensing royalties. We never pay by the piece.
Attribution
Practitioners retain authorship. Partner institutions are named on every data card.
Provenance
An audit-grade chain of custody on every record, held in India under the DPDP Act.
What we will not record
We publish this because the trust of the people we work with matters more than any single dataset.We do not record people who cannot meaningfully refuse.
We work in no carceral facilities and with no bonded labour, and we capture minors only within heritage programmes that carry both guardian and institutional consent.
We do not record patients without explicit informed consent.
Any surgical capture proceeds only with a patient-facing consent process, full anonymisation, and the right to revoke consent before publication.
We do not record a technique without the practitioner's agreement.
Where a method is the practitioner's livelihood, the practitioner sets the licensing terms, not the institution that employs them.
We do not permit resale outside the contracted use.
Buyer licences are scoped to specific AI-training purposes, and resale, or any sale to defence end-uses outside that scope, is contractually forbidden.
Loom
Internal platformEvery production session will run through Loom, our capture-to-corpus pipeline. It aligns every stream to a single clock, stamps provenance on every value, and checks each episode’s integrity before it can enter the corpus.
A session that does not meet the standard does not ship. As the platform matures, we plan to open parts of Loom to the robotics community.
A vetted review network sits behind every session.
Independent technical reviewers affiliated with IIT Delhi and IIT BHU are being onboarded to validate every captured session against the EDS schema. Behind them, an anonymised network of master practitioners across nine partner sites forms the corpus.
Meet the network →We are hiring in New Delhi.
Six roles hiring now across capture engineering, computer vision, and field operations, plus four internships, with five more roles opening through 2026.
View open roles →Ready to talk?
Tell us your vertical, task, and embodiment, and we will respond within one working day with a feasibility note and a timeline.
Scope a custom dataset.
Tell us the vertical, task, and embodiment you train on, and we will respond within one working day with a feasibility note and a timeline.
datasets@anthrobase.ai →Become a capture site.
We work with hotels, hospitals, MRO facilities, craft collectives, and training institutes, and we supply the hardware, the operators, and the legal frameworks.
partnerships@anthrobase.ai →