SO101 pick_place_v2 — two-camera (front + wrist) demonstrations

Teleoperated pick-and-place demonstrations on an SO101 follower arm (leader→follower teleop), LeRobot v3.0 format. Successor to Dillonjohnson/pick_place: the v1 policy trained on a single front camera executed the full task sequence but missed the grasp by ±1–2 cm (1/8 grasps) — an error the front camera cannot see. v2 re-records the task with an added wrist camera for grasp-phase observability.

• Task string: "Pick up the cube and place it in the bin" — NOTE: the object is actually a ~7 cm blue puck (outside-wrap grasp); the string is inherited from v1.
• Episodes: 45 · Frames: 31,063 · FPS: 30
• Robot: so_follower (6-DOF: shoulder_pan, shoulder_lift, elbow_flex, wrist_flex, wrist_roll, gripper)
• Cameras: observation.images.front (fixed scene view) + observation.images.wrist (jaws + table) — both 1280×720, stored as video (AV1)
• Action / state: 6-D joint positions (degrees)

Composition (recorded 2026-07-21, one session, fixed scene)

episodes	content
0–35	clean picks, puck start position deliberately varied (~18×18 cm region)
36–40	rim-recovery: grasp initially catches the rim, operator recovers
41–42	mid-descent displacement: puck moved during the reach, operator re-targets
43–44	low-light variants (scene brightness 113 vs the session's 170)

Start pose: episode-mean [-40.1, -100.6, 96.5, 72.4, -4.8, 1.4] deg (pan/lift/elbow/wrist_flex/roll/grip), pan σ 7.6° across episodes. Demos hold the puck at gripper position 26–36 while carrying; a close ending below ~20 means the jaws missed the puck body (useful as a rollout verdict signal).

Known warts (verified harmless for training)

• ep28: the video window carries ~2.5 s of orphan frames from a redo at its tail — the recorder's discard drops parquet rows but not already-encoded frames. Kept takes sit at the window head; training decode is unaffected.
• eps 0–2: videos re-encoded near-all-intra AV1 (~5× bulkier, decode fine) as a side effect of deleting a junk episode via lerobot-edit-dataset.

Train (ACT, the v2 config actually used)

lerobot-train \
  --policy.type=act --policy.device=cuda --policy.push_to_hub=false \
  --dataset.repo_id=Dillonjohnson/pick_place_v2 --dataset.video_backend=pyav \
  --dataset.image_transforms.enable=true \
  --batch_size=8 --steps=150000 --save_freq=10000 --num_workers=8 \
  --output_dir=outputs/train/act_pick_place_v2 --job_name=act_pick_place_v2 \
  --wandb.enable=false

Trained models live in Dillonjohnson/act_pick_place (ckpt-v2-last). Two 720p streams cost ~2.7× v1's per-step time; CPU inference is ~1.05 s/chunk vs 0.34 s single-cam.