Workflow Overview

The platform home page presents a clickable flow diagram that serves as the navigation reference for the overall functionality. This chapter explains the architecture represented by that diagram, and may be read alongside the walkthrough video.

Two data routes and a single training line

• Physical route (green, upper): the operator guides the Leader Arm through the demonstration, recording one episode at a time.
• Simulation route (blue, lower): following a single demonstration, SDG automatically generates a full batch of varied data.

The two routes are not mutually exclusive. Both produce datasets in the LeRobot standard format and feed the same training line. In practice, most deployments use both: the simulation route to establish data volume, and physical data to supply the physical detail that simulation cannot reproduce.

The diagram also marks two cross-lane paths, dataset transfer and model deployment. Both refer to moving assets between hosts using short codes: datasets are transferred to the training host, and models to the production host.

The five interface entry points

The sidebar is ordered according to the workflow:

Entry point	Function	Corresponding chapter
Home	The flow diagram; nodes may be clicked to navigate to the corresponding function	This chapter
Record	Prompts for a choice between physical and virtual recording on entry	Demonstration Recording
Training	Dataset and policy selection, parameter configuration, and training status monitoring	Model Training
Inference	Inference verification and model conversion	Inference & Model Conversion
Data Tools	Hardware settings, dataset management, and model management	Hardware Integration & Calibration, Data Tools

It should be noted that hardware settings are located within Data Tools. Robot model selection, camera management, and arm calibration are all consolidated under the "Data Tools → Hardware Settings" tab, which is the function area least readily located on first use.

The complete workflow

1. Hardware connection: connect the robotic arms and cameras, and confirm the correct model on the hardware settings page (Hardware Integration & Calibration).
2. Calibration: SO-101 requires serial port detection and six-axis range calibration.
3. Policy determination: select ACT or GR00T N1.5. This decision simultaneously establishes the deployment hardware (Policy Selection).
4. Demonstration recording: by physical recording or simulation-based SDG generation (Demonstration Recording).
5. Data review: preview episode by episode in Data Tools and remove those of insufficient quality.
6. Training: select the dataset, policy, and compute device, configure the parameters, and start (Model Training).
7. Training monitoring: observe loss and GPU utilization, and terminate early in the event of anomalies.
8. Inference verification: execute inference and evaluate the arm's performance (Inference & Model Conversion).
9. Conversion and deployment: perform model conversion for the target hardware and transfer to the production host using a short code (Data Tools).

The hardware readiness check

The platform performs a hardware readiness check before entry to the Record, Simulate, and Inference pages, and issues a notification when components are missing. The items checked vary by page:

Page	Camera	Leader Arm	Follower Arm
Record	Required	Required	Required
Simulate	Not checked (imagery is rendered by the simulator)	Required	Not checked (the Follower is virtual)
Inference	Required	Not checked (the model assumes the operator's role)	Required

This table also describes the operating logic of the platform. See the hardware readiness check.