Platform Workflow
NexCore runs eight stages along a single forward pipeline, with two feedback loops that return to curation. Every stage shares the same model for assets, versions, and permissions.
01Collection taskModel + SN → upload mode → task sheet
02Data assetsFiltered by task · integrity checks
03Curation projectNon-destructive curation · pipeline decision
04Pipeline processingSelect operators · asynchronous run
05Dataset versionHDF5 / LeRobot · frozen on export
06Training and modelsUnified strategy · compute · checkpoints
07DeploymentCloud / edge · NexBench toggle
08Evaluation and feedbackNode scoring · bad case attribution
Stages at a Glance
| Stage | Input | Key actions | Output | Gate |
|---|---|---|---|---|
| 01 Collection task | Device inventory, scene, target count, collector | Select device model and SN → select upload mode → enter task details | Task sheet (format locked automatically, SN bound) | Device registration check |
| 02 Data assets | Upload receipts / feedback archives | Integrity checks, archiving by task | Asset records | G1 Upload integrity |
| 03 Curation project | Selected asset records | Create project; decide whether to run a pipeline | Curation project + processing steps | — |
| 04 Pipeline processing | Curation project with original RAW / HDF5 data | Select operators, configure parameters, start processing | Structured outputs (tagged with pipeline job ID) | G2 Pipeline complete |
| 05 Dataset version | Curation results that passed quality checks | Choose a format and export | Dataset version (frozen) | G3 Dataset-ready / G4 Version frozen |
| 06 Training and models | Frozen dataset versions | Select model family, apply unified strategy, allocate compute | Training job + model versions / checkpoints | — |
| 07 Deployment | Model version, target device | Deploy; toggle NexBench per job | Deployment record + inference record | G5 Pre-deployment |
| 08 Evaluation and feedback | Inference execution data (built-in feedback) | Node scoring, bad case attribution | Evaluation results + bad cases | G6 Feedback review |
When something goes wrong at any stage, this table tells you whether an input, an action, or a gate is missing.
Production Loops
| Loop | Path | Value |
|---|---|---|
| Loop A · Data loop | Collect → assets → curation / pipeline → dataset → training → deployment → inference → bad cases → back to curation | Each iteration produces a dataset that covers the nodes and scenarios that failed in the last round. Failed samples are never discarded; failure data is ingested too. |
| Loop B · Version loop | Dataset frozen on export → training bound to a fixed version → model records its training lineage → deployment records the model version → inference records carry the version → bad cases record their source version | Every result can answer: which samples were used, what curation was applied, and which version it is pinned to. |

