Install

Quickstart

Start a Session

1. Power on the arms, leaders, and cameras.

2. Launch the GUI:

Collect Tab

• Station config
• robot type, gripper type
• controller type
• camera type, serial
after editing the config, click Preview to preview three view videos

• Can bus setup
CAN buses are brought up for you when you press Start Teleop. If an arm is unresponsive, click Reset CAN and try again.

• Calibrate the leader arms (when needed)
Do this the first time, or whenever a leader was moved/re-seated, or an arm tracks the wrong way. 1. Set the home zero
hold the leader arm in the pose you want to map to the follower's home, then either:
• In the GUI: click Zero L (left leader) / Zero R (right leader), and confirm; or
• From a terminal:
2. Check a joint's direction (optional)
to see live readings while you wiggle each joint:
If a follower joint moves the opposite way to the leader during teleop, flip that joint's sign in configs/station_yam.yaml(leader_joint_signs) and restart teleop.

Collect Data

1. In the Station panel (left), set the Task name (e.g. put the bottle in the bin).

2. Click Start Teleop. The followers ease to the leaders' pose, then track them.

3. Use the teaching-handle buttons or the on-screen buttons:
1. Top button = sync on/off (follower mirrors the leader).
2. Second button = start/stop recording an episode.

4. Perform the task, stop recording. Repeat for as many episodes as you need. The Episodes counter shows how many are saved for the current task.

5. Press E-STOP at any time to stop all motion immediately. To resume after an E-STOP, click Start Teleop again.

Review the data

1. Open the Review tab and refresh.

2. Pick an episode to see its camera videos and the joint/gripper curves.

3. Delete any bad episodes so they aren't used for training.

Train a policy

1. Environment Setup & Prepare pre-train checkpoints
details refer to GitHubGitHub - Hatty-z/openpi at c144147d64b7727a519d429f28799056eecbc3d3

GitHub - Hatty-z/openpi at c144147d64b7727a519d429f28799056eecbc3d3
Contribute to Hatty-z/openpi development by creating an account on GitHub.
GitHub
, GitHubGitHub - Hatty-z/molmoact2 at f266b758cbe64ef8cec74b7befb42c22749c8b84

GitHub - Hatty-z/molmoact2 at f266b758cbe64ef8cec74b7befb42c22749c8b84
Contribute to Hatty-z/molmoact2 development by creating an account on GitHub.
GitHub
, GitHubGitHub - Hatty-z/abc at 8619cdf42b59cb8ec726ea2af0b929789f7bfa5c

GitHub - Hatty-z/abc at 8619cdf42b59cb8ec726ea2af0b929789f7bfa5c
Contribute to Hatty-z/abc development by creating an account on GitHub.
GitHub

2. Prepare Dataset

• For pi0/pi0.5/molmoact2:

• For ABC:

3. Start training in the Train tab:
1. Choose the backend (`pi0`,`pi05`, `molmoact2`, or `abc`).
2. Fill in the fields:
• run name — a label for this run.
• dataset / task — the converted dataset-name
• steps, batch / GPU, GPUs — change as you want.
• base checkpoint / pretrained ckpt (optional) — path to a checkpoint to fine-tune from. Leave blank to train from scratch.

2. Click Start. Watch the live loss curve; logs stream in the panel.

Deploy / run a rollout

1. On the GPU box, start the server for your backend (see docs/deploy.md for the exact per-backend command), e.g. ABC:

2. On the robot host, in the Deploy tab:
1. Enter the server host, port, and the task prompt.
2. Click Start. The arms ease to a start pose, then the policy drives them.
3. Keep a hand near E-STOP for the first rollouts of any new checkpoint.
You can also run a rollout from the terminal: