Real to Real
Record demonstrations with a physical Leader arm, Follower arm, and cameras; train a policy from those demonstrations; then validate the policy on the physical task.
In this tutorial, Real to Real describes the physical demonstration-to-inference workflow in Robotic Suite/Physical AI. The operator supplies actions through the Leader during recording. During inference, the trained model supplies actions to the Follower.
What you will build
You will produce a reviewed LeRobot dataset, a trained policy checkpoint, and a record of task performance on the physical setup. If another host will execute the task, you will also transfer the selected model and repeat validation on that host.
Hardware at each stage
| Stage | Physical Leader | Physical Follower | Physical cameras |
|---|---|---|---|
| Demonstration recording | Required | Required | Required |
| Training | Not used | Not used | Not used; training reads the saved dataset |
| Inference and validation | Not used | Required | Required |
The installation guide covers the required platform setup. Its optional simulation environment build can be skipped for this physical recording workflow.
Follow the tutorial
| Step | Start here | Completion check |
|---|---|---|
| 1 | System requirements and installation | The selected host can run the platform and the intended policy |
| 2 | Hardware preparation | Both arms are identified and prepared; the required cameras are connected and enabled |
| 3 | Policy selection | ACT or GR00T N1.5 matches the task and target device |
| 4 | Physical demonstration recording | Complete task demonstrations are saved with the required camera views |
| 5 | Dataset management | Episodes are reviewed and the intended dataset is selected for training |
| 6 | Model training | A usable checkpoint is saved and training behavior has been reviewed |
| 7 | Real-world validation | Repeated task results and failure cases are recorded |
| 8 | Model deployment | The selected model passes validation on its execution host |
Start with a repeatable task
Choose a task with an observable completion condition, such as moving an object from a start area to a target area. Fix the camera mounts, check that the entire motion is visible, and define the variation in object position that the demonstrations should cover.
Review each recording before adding it to the training dataset. The recording screen's default episode count is a starting setting; choose the amount of data according to the task and selected policy. See demonstration quality.
Choosing between the two routes
| Route | Where demonstration data comes from | Start here |
|---|---|---|
| Sim to Real | Demonstrations in Isaac Sim and SDG-generated episodes | Sim to Real |
| Real to Real | A physical Leader, Follower, and cameras | Physical recording |
Both routes use the shared training and inference tools. If you later add simulation data, review the data before merging compatible datasets. Moving a model to another computer also requires checking its device support and physical setup; it does not establish compatibility with a different robot.
Walkthrough
The shared product walkthrough covers recording, training, and inference. Follow the physical recording steps in this tutorial when using it.
For the product's full workflow, see the platform workflow overview. For troubleshooting, see the FAQ.