NemoFleet Ops Console
← NemoFleet overview

NemoFleet Ops

— · --:--:-- Connecting…
Stop Cosmos endpoint ↗

Scenario run

Run scenario steps through each event with real calls: Cosmos on camera frames, then Nano or Ultra through the escalation gate.
No run in this session yet.

Live floor

Simulated motion · quarantine zones and halts from live KV tags
Moving (simulated) Knows a hazard (peer) Halted (KV tag) Handshake Person

Where decisions are settled

no data yet

Alarms

0 open ·

Actions

written to D1 and KV
Active fleet tags (KV)
—
Recent actions (D1)
—

Models and services

NVIDIA modelNebius service

Robot cameras · Cosmos perception

Cosmos3 Super Reasoner · Token Factory dedicated endpoint · frames are simulation renders, results are stored real responses

Night Shift

Replay runs live on Token Factory. Dream, Gate and Consolidate need the Serverless Jobs runner, which is not built yet.
0.62—

Learning curve

Ultra share of model decisions, % by run

Fleet memory

D1 precedents · KV tags

Usage today

tokens by model (UTC day)
Cosmos is billed per minute while its dedicated endpoint runs ($0.37/min), not per token.

Model call log

every Token Factory call, from D1 · newest first
TimeTierModelRobotEventLatencyTokensResult
Alarm rules and thresholds (design)

Raised live today: person or safety event (critical), hazard sighting (warning), operator quarantine and pause (info). The other rules are the design for the full system.

AlarmSeverityFires whenAutomatic responseOperator action
Person in laneCriticalPerception or sensors detect a person in a robot laneNearest robots halt; always escalates to UltraConfirm zone clear, release
Collision riskCriticalPlanner cannot find a conflict-free path within 2 sAffected robots hold positionQuarantine zone, reassign tasks
Schema rejectsWarningMore than 2% of model outputs fail validation in 10 minRetries, then falls back to safe stopInspect the call log
Ultra latencyWarningUltra p95 above 30 s for 5 minLive escalations route to Nemotron SuperCheck Token Factory status
Escalation stormWarningUltra calls per 100 events double versus the 7-night averageNoneReview τ, check for a new hazard type
Confident mistakesWarningReplay finds Nano wrong at confidence ≥ 0.9 on more than 0.3% of decisionsLower τ for next shiftReview flagged episodes
Promotion gate failedInfoCandidate Nano regresses normal operationsProduction adapter stays in placeRead the MLflow run
Night Shift stalledWarningAny stage exceeds its time budget or a job fails twicePipeline pauses, keeps current adapterRetry stage
Idle dedicated endpointWarningCosmos endpoint running with no calls for 15 minNoneStop endpoint
BudgetWarningDaily spend passes 80% of budgetNoneStop idle endpoints, defer dreams
Stale hazard tagInfoA tag expires while robots are still inbound to that cellRe-request perception on next sightingNone needed
Robot offlineWarningNo heartbeat for 10 sPlanner reserves its last cellDispatch technician

Episodes ·

Loading…

TimeTierRobotEventLatencySchemaDecisionNight Shift replay (Ultra)