Skip to content

Quick Start ​

This page gives you a working overview of NexCore and walks through a minimal example of the full collect → train → deploy loop.

Before You Begin ​

RequirementDetails
Platform URLNexCore website and sign-in: nexcore.lumosbot.tech
AccountAccounts are provisioned by an administrator; self-registration is not available. See Accounts and Members. If you don't have an account yet, request access.
Project accessData is isolated by project. When your account is created, the administrator assigns your role and the projects you can access.
Collection devicesBefore collecting, each device must be registered in the device inventory with its serial number (SN). See Device Management.
BrowserThe latest Chrome or Edge, with a screen at least 1440 px wide so the data workbench timeline and charts display in full.

The Console ​

After you sign in, the console has three main areas:

AreaContents
Top barAssistant and personal settings
Left navigationGrouped by workflow: Data Collection, Data Curation, Model Training, Model Deployment, Model Evaluation, Device Management, and System. Only modules available to your role are shown.
HomePlatform asset statistics (total data collected, datasets, models, devices, inference runs), workflow overview, training loss monitor, recent activity, device overview, and latest models

The shortcut buttons on the home page (New Collection, New Training, Inference Test, and New Deployment) take you straight into each workflow.

Start by Role ​

RoleMain responsibilitiesStart here
AdministratorAccount provisioning, role and project assignment, compute quotasAccounts and Members
Field application engineer (FAE)Create collection tasks, curate data, export datasets, train and deploy modelsTutorial B
Data collectorClaim assigned tasks and collect data as specifiedRunning a collection task
R&D engineerUpload and configure operators, maintain training templates, deploy modelsDeveloping Operators · Tutorial A
Data ownerDefine quality criteria and sign off on dataset versionsData Workbench · Datasets and Versions

Minimal Example: One Complete Loop ​

This example uses a registered Station Lite dual-arm workstation and an example task: placing a cube in a box.

  1. Create a collection task. Go to Data Collection → Collection Tasks and click New. Select device model Station Lite, the device SN, and Full upload, then enter task name pick_place_demo, a target count of 20, a collector, and a due date.
  2. Collect and upload. The collector claims the task and records 20 demonstrations through teleoperation. Each episode uploads automatically when it ends and is stored as HDF5.
  3. Curate. In Data Assets, filter by the collection task pick_place_demo, select all items, and create a curation project. In the data workbench, step through each episode, label failed grasps as Anomaly_grasp failed, mark those segments as excluded ranges, and review each episode.
  4. Export a dataset. Click Export in the workbench and choose HDF5. This creates dataset version v1.
  5. Train a model. Go to Model Training → Task Management, click Create Training Job, select dataset v1, a model family, and a compute tier, then submit.
  6. Check direction early. While the job is running, export a checkpoint from the job details page; it is registered as a model.
  7. Deploy for inference. Go to Model Deployment → Inference Deployment, select the checkpoint model and target device, complete the safety confirmation, and start inference.
  8. Review feedback. Inference data flows back automatically. In Data Feedback, review the failure segments and add them to a curation project to prepare the next dataset version.

Teleoperation data vs. egocentric data

HDF5 data from Station Lite and MOS can be curated directly. RAW data from Robot-free Ego and Robot-free UMI must first be processed by a data pipeline.

Next Steps ​

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