Model Management
Models are the core of the asset layer. The real measure of whether a project has delivered is the number of model versions and deployment records, not the number of training runs.
Where: Model Deployment → Model Management in the left navigation.
Where Models Come From
| Source | Description |
|---|---|
| Completed training | A training job that runs its full step budget registers a model version automatically |
| Checkpoint export | A checkpoint exported during training is registered as a model, also labeled with the training progress at export |
End-to-End Lineage
Every model version records:
| Field | Description |
|---|---|
| Model name and version | Versions are generated from the date and a sequence number, for example v20261020-01 |
| Training job | The ID of the training job that produced the model |
| Dataset versions | The one or more dataset versions used for training |
| Training strategy | The unified long-horizon strategy identifier |
| Hyperparameter snapshot | The parameters locked when training was submitted |
| Step | The training step the model corresponds to |
| Progress | For checkpoint models only, the training progress at export |
| Model family | ACT / EasyDP / Pi 0.5 / R0, and so on |
Deployment records link back to model versions, so you can trace lineage in either direction:
Dataset version → Training job → Model version → Deployment record → Inference record → Bad caseHow Models Are Counted
Models are counted by model version, not by training job. A single training job can produce several model versions: the final model plus any number of checkpoints.
Model Actions
| Action | Description |
|---|---|
| Details | View lineage, snapshots, and all versions |
| Inference test | Select a device and start an inference test directly. See Deployment Overview. |
| Deploy | Start a cloud or edge deployment |
| Edit | Change the model name and notes; lineage fields can't be edited |
| Delete | Before deleting, confirm that no running deployment uses this version |
Multiple Model Families
The same data can be used to train models from different families, which makes it easier to move across robot embodiments and frameworks. The supported artifact formats and devices are determined by R&D validation; model and device combinations that haven't been validated can't be deployed.

