Examples & Models

ACT block placement: no training required

Use the released ACT policy to place a block in the center of a workspace with 102-FL. It learns a task in a specific scene. It is not a general grasping model and does not automatically recognize arbitrary new objects.

Watch the Demonstration Download the Model and Video

Equipment checklist

Item	Requirement
Robot arm	102-FL with a secured base, communication checks, and calibration completed
Cameras	Overhead `up` and frontal `front` RGB views, both 640 × 480 at 30 FPS
Scene	Match the workspace, block, target, and cameras to the release guide photographs
Software	Ubuntu 22.04, Python 3.10, LeRobot 0.4.1, and StarArm leader and follower plugins at 0.0.1
Computer	Check GPU, driver, and PyTorch requirements in the release guide; the recorded setup uses an RTX 5070 Ti and CUDA 12.8 builds
Leader	Not required for autonomous inference; connect and calibrate separately if used for teleoperation resets between trials

Complete Getting Started first, then prepare your cameras and environment as described below.

Cameras and scene

Open the version-pinned scene photographs and run guide and compare the side, oblique, and front views with your setup.

Fix the relative positions of the cameras, arm base, and work mat.
Check the live views: up must be overhead and front frontal.
Ensure the block and target are visible without obstruction from the base or cables.
Check lighting and the starting pose. Begin with the objects and workspace covered by the demonstrations.

The photographs illustrate viewpoints, not precise dimensions. Set the serial port and camera device numbers for your own computer while keeping the model input names unchanged.

Download and run sequence

Follow the complete published run guide in this order:

Step	Action	Success check
1. Install the environment	Create a separate ACT environment with the specified dependencies	`lerobot-record --help` works and the dependency check reports no conflicts
2. Download the model	Download `stararm102_pick_act_torch271.tar.gz` and retain the entire extracted directory	`pretrained_model` contains configuration, weights, and processor files
3. Calibrate the follower	Confirm the actual serial port and calibrate with the matching plugin	Use the same `robot.id` for calibration and execution
4. Check images	Check both camera views and device paths	Names, viewpoints, resolution, and frame rate match the model
5. Run one trial	Set the ports, model path, and a new evaluation directory; run one episode first	Movement is appropriate for the scene and a local evaluation record is created
6. Repeat evaluation	Increase trials only after the first works; record resets and results	Each trial has a reviewable record and a clear outcome

Important

Running the policy moves the real arm. Clear its workspace and prepare a stopping method first. Stop immediately on unexpected motion. Do not rename model joints or reuse another plugin's calibration files merely to suppress an error.

Full installation and execution commands are in the linked Chinese guide. Use the versions specified in that guide.

Check results and troubleshoot

Symptom	Check first
Camera does not open or views are swapped	Actual device paths, camera access by other programs, and `up` / `front` mapping
Calibration is missing	Device type and `robot.id` match the calibration setup
Model fails to load	`policy.path` points to the complete `pretrained_model` directory
Evaluation directory already exists	Use a new name and directory for the new trial; keep earlier records
Unstable movement or grasping	Stop, then check calibration, camera placement, lighting, objects, and starting pose

Record trials, successes, and failure reasons. Distinguish missed grasps, placement errors, and occlusion. One successful trial does not establish reliability across all scenes. See Maintenance & Support for more help.

Model information and further learning

Item	Published information
Release	`act-pick-v1.0.0`
Policy	ACT, trained for 100,000 steps
State and action	7 dimensions
Image inputs	Two RGB views: `up` and `front`
Task	Place the block in the center
Documentation reference	Example guide in Star-Arm-102 commit `307a86d9`

To change the task or scene, use the Learning Path and begin with task design, demonstration quality, and evaluation. Do not treat the current policy as a finished model for another task.