Learning Path

Back to Star Arm 102

Choose a learning goal

To run an existing task first, open the pretrained ACT example; data collection is not a prerequisite. To teach the arm your own task, follow the sequence below.

This page organizes the existing LeRobot guide into a learning path. Each stage links to the relevant section of the original guide. Complete first use and version checks before starting, and keep different plugin paths separate.

1. Define a clear task

Begin with one object, one target area, and a defined range of starting positions. Write down what counts as success, such as the object entering the target area with the gripper released. Fix the cameras and workspace, and record permitted changes in object position and lighting.

Success check: the goal, starting conditions, and success criterion are repeatable, and the operator can demonstrate the task consistently.

2. Install, calibrate, and teleoperate

Installation → Leader and follower calibration → Teleoperation without cameras

Verify direction and stopping with small movements before expanding the range. Record device types, ports, and calibration IDs.

Success check: leader and follower movements match without persistent communication errors, and you can reconnect without additional verbal instructions.

3. Set up cameras and workspace

Read Add Cameras and check device numbers, views, resolution, and frame rate. Keep the cameras fixed afterward and save photographs of the workspace layout.

Success check: each camera shows the task clearly, input names remain stable, and overhead and front views are not swapped.

4. Collect and inspect demonstrations

Read Data Collection. Record a few episodes first, review images, task labels, and action records, then increase collection volume.

• Check for occlusion, frozen images, or missing frames.
• Check task consistency and distinguish failed demonstrations from useful ones.
• Check output paths and episode counts, and retain original data.
• Before action replay, check the real workspace and device state. Replay moves the arm.

Success check: you can inspect at least one complete episode and explain which data is suitable for training.

5. Train an ACT policy

Read Training and Resuming Training and use your own dataset and output directory. Record code, dependencies, data versions, and training configuration. Keep checkpoints with their preprocessing configuration.

Success check: you can locate the training output and configuration, and confirm that they match your task, joints, and image inputs. A completed training command does not establish real-robot success.

6. Evaluate and iterate on the real arm

Read Run Inference and Evaluate. Use separate evaluation records, begin with one controlled trial, then test different starting positions. Record successes, total trials, and failure categories.

Success check: you can assess whether failures are more likely related to data, scene changes, calibration, or execution, and use that assessment to add demonstrations or adjust configuration.

Continue developing

Control & Development provides Python, ROS 2, and simulation entry points. Hardware & Open Resources provides structural resources for workspace setup. For help, gather versions and logs as described in Maintenance & Support.