Collect Demonstrations¶

Use a teleoperation example to collect demonstrations. The recorder saves trajectories automatically when configured.

Verify Your Camera Setup¶

Before running teleoperation, stream your cameras to confirm they're working:

uv run -m examples.stream_cameras

Parameter	Default	Description
`--freq`	`50`	Streaming frequency (Hz)
`--mw`	`Thread`	Middleware backend

Configure the Recorder¶

The recorder is already set up in each config file. Open examples/cfg/xarm_eef.py (or whichever config you're using) and update the path:

# examples/cfg/xarm_eef.py

TASK = "pick_and_place"  # ← trajectories save to data/{TASK}/

recorder: str | None = "Recorder"
recorder_cfg: RecorderCfg = field(
    default_factory=lambda: RecorderCfg(path=f"data/{TASK}/")
)

Set recorder = None to disable recording during testing:

recorder: str | None = None

Run Teleoperation for Data Collection¶

Select your station by setting STATION=<ClassName> before running. Run uv run -m examples.cfg to list all registered stations. Override any config field at the command line via tyro.

XArm7 + Spacemouse / Gamepad / KeyboardXArm7 + GelloSO100 Leader-Follower

Control the robot in Cartesian space (EEF delta):

STATION=Xarm7EEFStation uv run -m examples.teleop_eef

# override device:
STATION=Xarm7EEFStation uv run -m examples.teleop_eef --teleop "Gamepad"
STATION=Xarm7EEFStation uv run -m examples.teleop_eef --teleop "Keyboard"

# language instructions
STATION=Xarm7EEFStation uv run -m examples.teleop_eef --instruction "pick up can"

# data visualization while collecting
STATION=Xarm7EEFStation uv run -m examples.teleop_eef --visualizer "Rerun"

Parameter	Default	Description
`--teleop`	`Spacemouse`	Teleop device (`Spacemouse`, `Gamepad`, `Keyboard`)
`--freq`	`50`	Control frequency (Hz)
`--instruction`	`""`	Language instruction recorded with the trajectory
`--visualizer`	`None`	Visualizer to use (e.g. `Rerun`)

See Gamepad for setup instructions.

Mirror joint positions using a Gello leader arm. On startup the script checks that the Gello joints are aligned with the robot — fix any misaligned joints before confirming, then wait ~3.5 s for the arm to settle before control begins.

STATION=Xarm7GelloStation uv run -m examples.teleop_leader_follower

# load Gello config from YAML:
STATION=Xarm7GelloStation uv run -m examples.teleop_leader_follower --teleop_cfg_yaml path/to/gello.yaml

# bimanual setup:
STATION=Xarm7GelloStation uv run -m examples.teleop_leader_follower --teleop2 "Gello" --teleop2_cfg_yaml path/to/gello2.yaml

# language instructions
STATION=Xarm7GelloStation uv run -m examples.teleop_leader_follower --instruction "pick up can"

Parameter	Default	Description
`--teleop_cfg_yaml`	`None`	Path to Gello YAML config for the primary arm
`--teleop2`	`None`	Second Gello device class (bimanual)
`--teleop2_cfg_yaml`	`None`	Path to Gello YAML config for the second arm
`--freq`	`15`	Control frequency (Hz)
`--instruction`	`""`	Language instruction recorded with the trajectory

See Gello for setup instructions.

Leader-follower joint mirroring over serial. Before running, update the calibration paths and serial ports in examples/cfg/so100.py:

# follower arm
calibration_file="/path/to/follower.json"
port="/dev/ttyACM0"

# leader arm (teleop)
calibration_file="/path/to/leader.json"
port="/dev/ttyACM1"

STATION=SO100Station uv run -m examples.teleop_leader_follower