Skip to content

Deployment Overview ​

Deployment runs a model version in the cloud or on an on-site edge inference node to control a real robot. Each deployment creates a deployment record, and a running deployment is called a deployment instance.

Where: Model Deployment → Inference Deployment in the left navigation.

Deployment Types ​

TypeRuns onBest forPrerequisites
Cloud inferencePlatform computeQuick validation and demos, with no local hardwareTarget device is online
Edge deploymentOn-site edge inference nodeLong-term production and on-site operation with real-time requirementsAll four pre-deployment checks pass

Network and latency

Cloud inference depends on the network between the device and the platform, and latency affects control quality. Use edge deployment for latency-sensitive scenarios.

Deployment Workflow ​

text
Select model version → Select target → Pre-deployment checks → Submit → Instance running → Safety confirmation → Start inference → On-site validation

1. Select a Model Version and Target ​

  1. Click New Deployment and select a model version. You can choose a model from completed training or a checkpoint model.
  2. Choose a deployment type:
    • Cloud inference: select the inference device and compute tier.
    • Edge deployment: select the edge node and the robot connected to it.
  3. Once you select a device, the page shows its name, model, SN, online status, and last heartbeat.

2. Pre-Deployment Checks (Edge) ​

CheckRequirement
Hardware compatibilityThe node's GPU memory and compute meet the model's requirements
Approved clientThe NexCore edge client is installed and signed in on the node
Matching inference codeThe client has synced the inference code for the model family
Disk space and network authorizationThere is enough disk space for the model, and the node is authorized for network access

If any check is missing, Submit Deployment is disabled and the missing items are listed.

Station Lite requires an external inference node

Station Lite doesn't include inference compute by default. Edge deployment requires an external inference node that meets the requirements.

3. Submit the Deployment ​

  1. Turn on NexBench evaluation if needed (off by default). The setting is saved in the deployment record.
  2. Click Submit Deployment. The platform validates the configuration, then sends the model and inference code to the node.
  3. Once the client finishes pulling them, the instance status changes to Running.

4. Safety Confirmation and Starting Inference ​

Before every inference run, the platform requires a safety confirmation:

  • [ ] No one is within 1 m of the robot
  • [ ] The robot's motion path is clear of obstacles
  • [ ] The emergency stop is within reach and working
  • [ ] The current model version matches the task

Inference starts only after every item is confirmed; cancelling means nothing starts. You can interrupt execution at any time with the emergency stop.

Deployment Instances ​

StateMeaningAvailable actions
CheckingThe four edge checks are runningWait
DeployingThe client is pulling the model and inference codeWait, view logs
RunningReady for inferenceInference test, view logs, restart, stop
StoppedStopped manuallyStart again
FailedThe deployment failed; the existing stable version is unaffectedView logs, roll back, redeploy

View logs: Instance details show recent log output, 200 lines by default and up to 5,000.

Restart: If the inference code or client misbehaves, click Restart to relaunch the instance with the same model version.

Failures and Rollback ​

  • A failed deployment never replaces the current stable version, which keeps running.
  • You can roll back to any earlier version in the deployment history.
  • Failure records are kept for troubleshooting.

Inference Tests ​

Click Inference Test on a deployment instance or in Model Management, then select a model, device, test scenario, and number of trials (for example, 10) to evaluate the model quickly. Tests also require a safety confirmation.

Deployment Records ​

FieldDescription
Deployment IDGenerated by the platform
Model versionThe deployed model, linked to its training job and dataset versions
Deployment typeCloud / edge
Target deviceDevice name · SN, or edge node ID
NexBenchOn / off
Status and timestamps—

In This Section ​

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