Real-World Validation
Validate the trained policy on the physical task before choosing a checkpoint for deployment. A completed training run does not by itself establish that the real setup will perform the task successfully.
Prepare the real setup
- Complete the physical Follower setup using arm setup and the calibration procedure applicable to the robot.
- Connect and enable the physical cameras. Check the Wrist and Top viewpoints against those used in simulation.
- Confirm the task layout, intended object variation, and the model checkpoint to evaluate.
- Keep the robot's working area clear and have the stop control available during the first trials.
The physical Leader is not required for inference because the trained model provides the actions.
Run the inference procedure
Follow Inference & Model Conversion to select the checkpoint, instruction, device, FPS, and action chunk settings. Where a CUDA host is available, the shared guide recommends verifying the original model before evaluating the converted model on the target device.
Repeat the task under comparable conditions. Record at least the following information so that subsequent models can be compared:
| Record | Why it matters |
|---|---|
| Dataset, policy, and checkpoint | Identifies what was evaluated |
| Host and model format | Separates original-model and converted-model results |
| Camera placement and object position | Describes the input conditions |
| Successful attempts / total attempts | Provides an observed task success rate |
| Failure description | Shows which part of the workflow needs attention |
Set the acceptance criteria for your task before testing. The inference guide describes task completion, motion path, alignment, and behavior at new object positions as evaluation dimensions.
Address the Sim-to-Real gap
Read the Sim-to-Real gap discussion and use the inference diagnosis table to investigate observed failures.
- Check arm configuration and camera viewpoints before changing training parameters.
- Review the simulation demonstration and generated episodes. Confirm that scene variation covers the actual task conditions.
- If real-world examples are needed, follow physical data supplementation, review the data, and merge compatible datasets using Dataset Management.
- After changing the dataset, create and evaluate a new training run following the training guide.
Next: Once a checkpoint meets the task's acceptance criteria, continue to Model Deployment.