alexhegit/so101-simstudio-lab01-pnp-agent193-lingbot-eef

SO-101 Lab01 — agent193 LingBot-VA EEF labels (derived)

LeRobot v3.0 derived view of the 193-episode agent-yaw follow2 joint dataset. Videos and proprioception match the source; only action (and its stats / feature meta) are rewritten to LIBERO-style 7-D Cartesian EEF Δ + gripper for fair LingBot-VA fine-tuning.

Hub id	alexhegit/so101-simstudio-lab01-pnp-agent193-lingbot-eef
Source (joints)	alexhegit/so101-simstudio-lab01-pnp-agent-yaw-v4-actonly-follow2
Robot	so101_mujoco
Action	7-D: eef.delta_{x,y,z}, eef.delta_axis_angle_{x,y,z}, gripper.command
State on disk	15-D (unchanged from source)
Cameras	camera_front, camera_top, camera_wrist (640×480, 20 Hz)
Size	193 episodes / 130835 frames

Action contract

Matches lerobot/lingbot_va_libero_long / LIBERO OSC scaling:

• translation channels clipped to [-1, 1], scale 0.05 m
• rotation channels clipped to [-1, 1], scale 0.5 rad (world-frame axis-angle)
• gripper: +1 close, -1 open (held per frame)

Labels are produced offline with MuJoCo FK (SO101EEFAdapter / labs/lab01_pnp/build_lingbot_eef_dataset.py). Closed-loop eval on SO-101 still needs the project IK postprocessor (LAB01_LINGBOT_EEF_ADAPTER=true).

Do not train joint-space policies (ACT, MolmoAct2 6-D, …) on this repo — use the source joint dataset instead.

Rebuild

.venv-rocm/bin/python labs/lab01_pnp/build_lingbot_eef_dataset.py \
  --source datasets/so101-simstudio-lab01-pnp-agent-yaw-v4-actonly-follow2 \
  --output .tmp/lab01_lingbot_eef_agent193

LAB01_LINGBOT_EEF_DATASET_ROOT=./.tmp/lab01_lingbot_eef_agent193 \
  ./labs/lab01_pnp/push_lingbot_eef_dataset.cmd

Pipeline and reuse notes: lingbot_va_eef_adapt.md.

Not a Home20 ranking dataset

This set enables fair EEF-policy transfer experiments. Published Home20 scores for joint policies stay on the source joint Hub id and home20.md. Do not quote Home20 for a LingBot-VA checkpoint until fixed-cube motion is clearly task-directed.