Tutorial Overview¶

RIO supports the full robot learning workflow — from collecting demonstrations on real hardware to deploying a fine-tuned policy. The diagram below shows how the pieces fit together.

flowchart LR
    A("<b>Setup</b><br/>Configure your station<br/>and connect hardware")
    B("<b>Collect Data</b><br/>Teleoperate the robot<br/>to record demonstrations")
    C("<b>Convert Data</b><br/>Export trajectories to<br/>LeRobot / DROID format")
    D("<b>Fine-tune</b><br/>Train a VLA policy<br/>on your data")
    E("<b>Deploy</b><br/>Run closed-loop<br/>policy inference")

    A --> B --> C --> D --> E

Steps¶

1. Setup¶

Configure Station

All examples are driven by a station config: a Python dataclass that declares every hardware and software node in your setup (robot arm, gripper, cameras, recorder, policy server).

Each config lives in examples/cfg/ as a self-contained file (e.g. xarm_eef.py, xarm_gello.py, so100.py). Edit the one that matches your hardware.

2. Collect Data¶

Teleoperate the Robot and Visualize Data

Use a teleoperation script to control the robot and record demonstrations. RIO supports several input devices:

Device	Config	Teleop script	Guide
Spacemouse / Gamepad / Keyboard	`examples/cfg/xarm_eef.py`	`examples.teleop_eef`	Gamepad
Gello leader arm	`examples/cfg/xarm_gello.py`	`examples.teleop_leader_follower`	Gello
SO100 leader-follower	`examples/cfg/so100.py`	`examples.teleop_leader_follower`	—
SO100 bimanual	`examples/cfg/bimanual_so100.py`	`examples.teleop_leader_follower`	—

Run uv run -m examples.cfg to list all registered stations. Set STATION=<ClassName> before running any script.

3. Convert Data¶

Exporting Data

Convert the recorded .vla trajectories to LeRobot / DROID format for use with the openpi training pipeline.

uv run examples/data/convert_to_lerobot_droid.py --input /data/rollouts/my_task/

4. Fine-tune¶

PI0.5 DROID Fine-tuning

Fine-tune a pi0 policy on your collected dataset using the openpi training pipeline.

5. Deploy¶

Run Policy Inference

Run the fine-tuned policy on the robot in a closed-loop at a fixed frequency.

STATION=<YourStation> POLICY=Pi05Cfg uv run -m examples.policy_inference --instruction "Place the cup on the shelf."