Skip to content

FAQ ​

Product Overview ​

What problem does NexCore solve? ​

It connects embodied AI data from collection, curation, and dataset versioning through training, model versioning, deployment, and inference feedback. This addresses three problems: fragmented workflows, results that can't be traced, and repeated manual data preparation.

What is NexCore, in one sentence? ​

An integrated data and model production line for collection, training, and inference: demonstration data goes in, and trainable dataset versions, deployable model versions, and traceable inference records come out.

What objects does the platform manage? ​

Four: data, datasets, models, and inference records. For any training run, you can answer which samples were used, what curation was applied, and which version it is pinned to.

What is the main workflow? ​

Create a collection task → collect and upload → ingest into the asset pool → create a curation project (with an optional pipeline) → export a dataset version → train → deploy for inference → evaluate with NexBench → return bad cases to curation.

How does NexCore relate to MOS? ​

They are separate products. MOS 2.0 is the on-site robot system that handles the chassis, arms, map waypoints, obstacle avoidance, and the task center. NexCore provides device records, model deployment, and data curation, and only links to MOS from the device page.

We have our own robot. Can we use NexCore? ​

Yes. Robots connect through the device integration layer, and third-party robots are supported through adapters, so you don't need to replace your robot. The scope of each integration is assessed per project.

Can we move to another platform later? ​

Datasets export in the open HDF5 and LeRobot formats, so there is no format lock-in.

Is our data safe in the cloud? ​

Private deployment is supported. In the public cloud, data is isolated by project, source data is read-only, and every operation is logged.

Roles and Permissions ​

Who can create and edit collection tasks? ​

FAEs and R&D engineers can create and edit tasks and target counts within the projects they have access to. Data collectors can only run tasks assigned to them.

Can data collectors change the target count? ​

No. Target counts are changed by an FAE or R&D engineer, and the platform records who changed it and when. If the task has already been claimed, the collector is asked to reconfirm.

Can data collectors change the upload mode? ​

No. The upload mode is set by an authorized role when the task is created and is locked afterward. Collectors follow it as specified.

Who can configure operators? ​

Only R&D engineers. Other roles can open the pipeline dialog and see the chain and selected operators, but parameters are read-only, with a message that R&D engineer permission is required.

Who can start pipeline processing? ​

FAEs and R&D engineers, within their projects. The entry point isn't visible to data collectors.

Can one customer see another customer's data? ​

No. Data is isolated by project. Searches never return another project's data, and the platform never indicates that such data exists but is inaccessible.

Collection Tasks and Devices ​

How do I create a collection task? ​

Go to Data Collection → Collection Tasks → New. First select a device model, then a specific device SN from the registered devices of that model, then an upload mode, and finally enter the scene, target count, collector, and due date. See Collection Tasks.

Why do I select the device model first? ​

The model determines the storage format and the curation path: teleoperation data is stored as HDF5 and can be curated directly, while robot-free data is stored as RAW and must go through a pipeline first.

Why do I also need to select a device SN? ​

The model only says what kind of device it is; the SN identifies the exact unit. Data assets need to answer which device recorded each episode, so every task must be bound to a specific device.

The device list is empty after I select a model. What should I do? ​

No devices of that model are registered in the device inventory, so task creation is blocked. The page offers Register Device and Go to Device Management. Register the device and its SN, then come back to create the task.

Can I select a device that is under maintenance or offline? ​

No. The list shows only devices of the selected model, and those that are under maintenance, offline, or decommissioned are grayed out.

Can one device run two tasks at the same time? ​

A device can be bound to several open tasks, but it shows "Bound to N open tasks" when you select it, so you can tell whether it is already committed. On-site scheduling is coordinated by the collection lead.

How do I choose between the three upload modes? ​

Full upload suits small batches of high-value tasks. Screened upload suits high-volume tasks with experienced collectors. Segmented upload suits all-day operation and continuous operation data, with a minimum segment length of 5 minutes.

Can the target count be 0 or a decimal? ​

No, it must be a positive integer. When the target changes, the progress denominator updates and the completed count never decreases.

When is a task considered complete? ​

When it has reached the target count and all uploads have finished. Reaching the count with uploads still pending doesn't count as complete.

Can a task start without a teleoperation device or if the checks fail? ​

No. If the teleoperation device ID is missing or the pre-collection checks fail, the task can't be set to ready.

Can I change the device, SN, or format after creating a task? ​

No. Create a new task instead. The original task is kept so historical data definitions aren't rewritten.

Data Assets and Curation ​

What does one row in the asset list represent? ​

One uploaded episode, not one collection task. Each row can be traced to its task, device, format, scene, and curation status.

Why can't I select some items? ​

RAW items that haven't been through a pipeline are in the Awaiting Pipeline state and can't be selected, not even with Select All. Run a pipeline first.

My filter returned no results. Is something broken? ​

No. A new task that hasn't uploaded any data yet returns an empty list, and the page shows an empty-state message.

Can the same item be in more than one curation project? ​

Yes. Different projects can reference the same read-only source asset, and items are deduplicated by data ID within each project.

Can I add all the data from a source to a project in one click? ​

No. Filters only narrow down the candidates; items must be explicitly selected to be included, to prevent accidental inclusion.

Why doesn't the workbench have a Save button? ​

Every change saves automatically, and the platform keeps the full history and audit trail.

Does curation modify the original data? ​

No. Curation only produces metadata and export instructions. Source files are read-only, and processing results are stored separately.

How do I fill in a quality label? ​

The type is always Anomaly, and the note describes the specific issue, for example Anomaly_black frames. Black frames, dropped frames, and missing joint data are just examples, not a fixed list.

Are labeling and reviewing the same thing? ​

No. A label describes a quality issue; a review confirms that a person has inspected the item. Adding a label doesn't mark the item as reviewed.

What's the difference between excluding an episode and excluding a range? ​

Excluding an episode leaves the entire item out of the export; excluding a range trims a time span out of it. Neither deletes the source file.

Batch processing shows "Completed with Errors". Did the whole batch fail? ​

No. At least one item failed and the rest were processed normally. Check the logs to find the failed items, then retry or review them manually.

Pipeline Processing ​

Where do I start pipeline processing? ​

From the Pipeline button in the data workbench header, which opens a full-page dialog. It isn't a top-level item in the left navigation.

Why isn't the pipeline required for all data? ​

Teleoperation HDF5 data already includes robot state and can be curated directly, while robot-free RAW data must go through a pipeline to produce structured data. So the pipeline is optional in general but required for RAW.

Can I leave the page while processing? ​

Yes. Pipelines run asynchronously. The workbench header keeps showing the status and progress, so you can come back to check.

If some operators fail, does the job still succeed? ​

No. A batch completes only when every required operator succeeds. If any fails, the whole batch is marked failed, with the failed step, logs, and a retry option.

Does a successful pipeline mean the data passes acceptance? ​

No. A successful run only means the outputs were generated; their quality still needs curation and human review.

Datasets ​

Which export formats are supported? ​

HDF5 and LeRobot. Each export is frozen as a version, so later changes to the project never overwrite it.

How are failed samples shown in an export? ​

As "succeeded (failed)", for example 99 (1). The two add up to the number of items included in the export, and the log lists the failed IDs and reasons.

Can I create a training job while an export is in progress? ​

No. Only fixed versions in the Exported state can be used for training.

How do I handle newly collected data? ​

Either export it as a separate dataset and train on it on its own, or merge it with a curated dataset in the same format to create a new version, deduplicated by sample ID with the parent version recorded.

Training and Models ​

Are there fast, standard, and fine training tiers? ​

No. Every job runs the full step budget (currently 80,000 steps) with the unified "long-horizon, quality-first" strategy. There are no tiers to choose from, so results from different people are always comparable.

Won't running the full schedule be slow? ​

It takes longer than a short run, but you can export checkpoints at any time to check direction without waiting for the end. Actual training time depends on GPU, batch size, and queueing.

What happens if no compute is available? ​

The job is queued, and the page shows that honestly; it is never shown as started before it is.

Can I get a model while training is still running? ​

Yes. Export the current checkpoint at any time; it is registered as a separate model and can go straight to deployment and inference testing while the main training job continues.

How are models counted? ​

By model version, not by training job.

Deployment, Evaluation, and Feedback ​

What do I need before an edge deployment? ​

Four things: hardware compatibility, the approved client, matching inference code, and disk space and network authorization. The deployment can't be submitted until all four are ready.

Does Station Lite have built-in inference compute? ​

Not by default. It needs external edge compute; don't assume the device alone is sufficient.

Why is a safety confirmation required before inference? ​

Running a robot on site carries safety risks. Confirm that no one is nearby, the path is clear, and the emergency stop works before continuing. If you cancel the confirmation, nothing starts.

Does a failed deployment affect the version in production? ​

No. A failure never replaces the current stable version, you can roll back, and the record is kept.

Where is the NexBench toggle? ​

On the inference job. It's off by default, and its setting is saved in the deployment record.

Do I need to turn on data feedback? ​

No. Execution data from every inference job flows back automatically. The NexBench toggle only controls whether the run is scored.

Can bad cases be used for training directly? ​

No. They must be reviewed first and then added to curation or imported into a new curation project. Only after curation and export do they become part of a training version.

Will bad cases end up in the production training set? ​

Not automatically. Bad cases are managed separately and enter training data only after they are explicitly added to curation and exported in a new version.

Why is continuous operation data archived every 5 minutes? ​

It turns a continuous stream into archives that can be uploaded, curated, and reused. Uploads resume after interruptions, and duplicate or out-of-order uploads never create duplicate samples.

Trials ​

How does a trial account work? ​

A trial runs through eight stages: application and assessment → environment and accounts → data preparation → curation and processing → training → deployment and validation → feedback and wrap-up → extension or close-out. See Request Access and Contact.

Will trial data be mixed with production data? ​

No. Trial workspaces are isolated from production, access is revoked when the trial ends, and data is handled according to the retention policy with an audit record.

Data Delivery and Model Download ​

How do I download annotated data? ​

In Dataset Management, find an exported dataset version and click Download to get the data files and export manifest as a multi-part archive. To formally deliver data to others, use a delivery order. See Annotated Data and Delivery.

Can I download models? ​

Yes. Select a model version in Model Management and click Download to get an archive containing the weights, config, normalization statistics, and inference config. See Downloading Model Artifacts.

Accounts and Permissions ​

How do I get an account? ​

Accounts are provisioned by an administrator; self-registration is not available. See Accounts and Members. If you don't have an account yet, request access.

Why can't I see data in a certain project? ​

Data is isolated by project, and each account sees only the projects it has been granted. Ask an administrator to adjust your project access.

NexCore is evolving rapidly; the latest version prevails.NexCore Docs · Embodied AI Skill Learning Platform by Lumos Robotics