Glossary
| Term | Definition |
|---|---|
| Collection task | The task sheet for a collection job, binding a device model, device SN, upload mode, scene, target count, collector, and due date. |
| Device model | The type of device, which determines the storage format and curation path. Station Lite, MOS, Robot-free Ego, and Robot-free UMI are currently supported. |
| Device SN | The serial number of an individual device. The model identifies the type of device; the SN identifies the exact unit. |
| Robot-based / teleoperation | Demonstrations performed by teleoperating a real robot. The data includes robot state and is stored as HDF5. |
| Robot-free / egocentric | Collection that doesn't involve a robot (for example, Robot-free Ego or Robot-free UMI). The data is stored as RAW. |
| Storage format | The format in which uploaded data is ingested: HDF5 or RAW. Set automatically by device model and can't be selected manually. |
| Upload mode | How data is sent from the device to the platform: full, screened, or segmented. Set when the task is created and locked afterward. |
| Asset pool / data assets | The complete set of uploaded data. Each row in the asset list is one episode. |
| Data ID | The unique identifier assigned at ingestion. All later curation, exports, and feedback reference the same ID, and it never changes. |
| Curation project | An editable workspace that references read-only source data in the asset pool and records labels, ranges, and processing steps. |
| Data workbench | The interface for working on a curation project, with preview, timeline trimming, quality labeling, review, and the pipeline entry point. |
| Non-destructive curation | Curation produces only metadata and export instructions. Source files are read-only and processing results are stored separately. |
| Anomaly label | The single quality label type, shown as Anomaly_<note>, for example Anomaly_black frames. |
| Review | A person's confirmation that they have inspected an item. Adding a label doesn't mark an item as reviewed. |
| Exclude episode | Leaves the entire item out of the export without deleting the source file. |
| Excluded range | A time span trimmed out at export; the source file is unchanged. |
| Pipeline | An asynchronous processing chain made of operators that turns RAW data into curatable structured data, and can also enrich HDF5 data. |
| Operator | A processing unit in a pipeline, such as a SLAM, vision, depth, or tactile operator. Uploaded and configured by R&D engineers. |
| Dataset version | A fixed snapshot exported from a curation project, in HDF5 or LeRobot format. Frozen on export. |
| Parent version | The source version recorded when a new dataset version is created by merging or adding new data. |
| Training job | A training run bound to one or more frozen dataset versions, using the unified long-horizon strategy. |
| Checkpoint | A snapshot of the weights at a given training step. It can be registered as a model and deployed for testing without interrupting training. |
| Model version | A model produced by training or registered from a checkpoint, recording its training job, dataset versions, strategy, hyperparameter snapshot, and step. |
| Model family | The model architecture used for training, such as ACT, EasyDP, Pi 0.5, or R0. |
| Deployment record | The record of one deployment, including the model, device, NexBench setting, and status. |
| Edge / cloud | Where inference runs. Edge inference runs on an on-site inference node; cloud inference runs on platform compute. |
| NexBench | Evaluation built into each inference job. It scores execution nodes and steps to locate failures and compare model versions. |
| Bad case | A failure segment from an inference run, including the failed node, execution step, failure window, and original clip. |
| Inference feedback | Execution data from inference jobs that flows back to the platform automatically, with no toggle required. |
| Continuous operation data | Complete data from robots running all day, archived in 5-minute segments. |
| Quality gates (G1–G6) | The six checkpoints in the pipeline. Data that doesn't pass a gate stays at its current stage. |

