Research · Open
/ 05Research

Frontier research on real physical work.

Eru Labs operates as a research initiative: eight active tracks, public technical memos, and open partnerships with AI labs, universities, and industrial research groups. Permissioned data, consent-first, worker-aligned.

/ 05.ATracks

Eight active tracks.

Each track has a lead, a research agenda, and at least one upcoming memo. Inbound partnerships and visiting researchers are welcome.

01 / RT-PWOPhysical Work OntologyA shared schema for tasks, tools, materials, roles, dependencies, and outcomes across trades and industries.Active
02 / RT-EAUEgocentric Action UnderstandingLearning from hands, tools, contact states, and worker POV. Fine-grained skill representation.Live · v0.4
03 / RT-ESIExocentric Site IntelligenceUnderstanding environments, crews, flows, zones, and coordination from third-person site context.Live · v0.4
04 / RT-OCMOutcome-Causal ModelingConnecting actions to rework, safety, quality, inspection, and economic outcomes.Active
05 / RT-AKTAction Knowledge TransferTransferring human skill into machine-usable representations for robotics, copilots, and training.Active
06 / RT-PMSPhysical Memory SystemsPersistent memory of sites, companies, workers, machines, and workflows. Long-horizon retrieval.Open
07 / RT-HRIHuman-Robot Work InterfacesHow humans and robots coordinate inside real operations. Mixed-fleet dispatch, escalation, and trust.Active
08 / RT-TGRTrust, Governance, Worker RightsPermissioned data, consent, role-based access, certification, privacy, and safety. Non-negotiable.Active
/ 05.BTechnical Memos

Papers, notes, and field memos.

We publish short, opinionated technical memos before formal papers. Drafts go out to partners early.

/ 05.CLab Principles

How we work.

Eru Labs operates as a research initiative inside PIN. Independence, technical clarity, and worker alignment are operating constraints.

P·01

Worker consent comes first.

No worker is captured without explicit, revocable consent. Identity is opt-in. Faces and audio are filtered by default at the Aegis layer.

P·02

Outcome-grounded, not vanity-trained.

Eru is evaluated against real outcomes — inspections, rework, claims, certifications — not just label accuracy on a held-out test set.

P·03

Permissioned data, role-based access.

Every record has a provenance chain, a permission scope, and a retention policy. Partner access is contractually bounded and audited.

P·04

Publish before pitch.

We share technical memos before commercial pitches. If a claim can't survive a public memo, it doesn't ship into a customer deck.

P·05

Embodiment-agnostic.

Eru is not married to any robot, any sensor, any vendor. The model is the layer above embodiment; partners come and go, the model persists.

P·06

Field-bound research.

Every track has a deployment counterpart on PIN OS. If it works on paper but breaks at a jobsite, it isn't done.

Research collaboration.

Visiting researcher slots open for AI labs and universities working on physical-world models, action understanding, and outcome-causal learning.