Model · v0.4 · Learning
/ 02The Model

A model of actions, sequences, and outcomes.

Eru is a continuously-learning foundation model of real physical work — embodiment-agnostic, outcome-causal, and trained on the messy reality that breaks scripted automation.

Eruv0.4 · 2.4B
Perception
Workflow
Outcome
Action
/ 02.AArchitecture

Four interlocking representations.

Each layer is useful on its own. Together they form a single physical world model that other systems can query.

L·01

Physical Perception

What is happening?

Tools, hands, objects, materials, equipment, environments, hazards, gestures, contact states — the moment-to-moment state of a worksite at frame level.

obj.detecthand.posetool.idmaterial.classhazard.clsscene.context
L·02

Task Structure

Where does it fit?

Steps, dependencies, phases, checkpoints, missing actions, incorrect sequences, blocked tasks, parallel work, and crew coordination.

seq.detectphase.clsdep.graphmissing.stepblocker.idhandoff.detect
L·03

Outcome Causality

Did it work?

Rework, inspection results, schedule impact, safety incidents, claims, quality scores, certifications, and downstream economic effect.

inspect.resultrework.causeclaim.linkschedule.∆safety.eventquality.score
L·04

Action Knowledge

How should it be done?

Human demonstrations, tool use, contact states, grip, force, timing, body position, correction strategies, and expert behavior.

tool.usecontact.stateforce.profiletiming.deltacorrectionexpert.prior
/ 02.BData Flow

One model. Many surfaces.

Eru consumes signals from the entire PIN network and partner integrations, exposing them as task-, action-, and outcome-grade representations.

Inputs

12 streams
Aegis first-person videovideo
Beacon third-person contextvideo
Mobile / site uploadsvideo
Partner sensorsimu/force
Robotics partner dataaction
Project managementtasks
Inspection recordsaudit
Change ordersaudit
Safety logsevents
Claims data∂outcome
Worker certificationsidentity
Supervisor correctionslabel
Eruv0.4 · 2.4B
Perception
Workflow
Outcome
Action

Outputs

12 surfaces
Task recognitionapi
Sequence predictionapi
Risk detectionstream
Quality verificationstream
Progress truthapi
Worker skill profilesidentity
Robot task priorsapi
Simulation scenariosbatch
Human guidancecopilot
Robot evaluationapi
Outcome predictionapi
Physical-world memorystore
/ 02.CDifferentiation

How Eru differs.

Eru is not a language model, a vision model, or a robotics control policy. It is a physical work model.

CapabilityLLMVisionRobotEru
Physical task recognitionNoBasicLimitedDeep · 1,420 types
Workflow sequence modelingText onlyNoNoYes · graph-native
Outcome causalityNoNoNoYes · business truth
Action knowledge / priorsNoNoPolicy-onlyYes · human demo
Real-world failuresNoNoSim-onlyYes · claims, rework
Continuous learningPeriodicPeriodicPeriodicLive · 14,892 hrs/day
Physical memoryContext onlyNoNoYes · persistent ∞

See how it works.

For deployment and integration details, explore the full technical stack.