playground-stage2-v1 — RoboTwin sim aimless-grasp conditioning data

Stage-2 of a 3-stage curriculum (DROID CoT pretrain → playground @grip conditioning → task finetune). 35 episodes (~47k frames @15fps) of aimless cube pick-and-place teleop in RoboTwin sim (Franka), stored as a LeRobot v2.1 video/h264 dataset, schema-identical to Numbnut/droid-keeps-v6:

• observation.images.ext = bird_eye camera (the <loc>/@grip grounding view), 320×240
• observation.images.wrist = right_camera
• observation.state (8,) = joint qpos (7) + gripper open (1=open)
• action (7,) = EE delta-euler + gripper, via compute_ee_delta_actions
• task = neutral prompt pool ("rearrange the cubes on the tabletop", …) — deliberately intent-free; intent is carried by the @grip condition tokens.

Repo-root artifacts:

• cot_table_playground.pkl — per-keyframe tokenized S (scene + @grip_now) and A (action steering + future @grip) spans, 3,258 keyframes, PaliGemma <loc>/<seg> grounded tokens projected to bird_eye. Deterministically generated (FK + HSV), no VLM. Frame indices already divided by the conversion stride (2).
• norm_stats.json — state/action normalization statistics over the train split.
• droid_provenance.json (inside the dataset) — lerobot episode → source episode map + the 3 held-out episodes (--split all repo: exclude held-out at load time).

Sources: teleop_stack_scatter (30 eps), teleop_playground (3 eps), teleop_stack_scatter_raster (2 eps, +front-camera renders, slight visual shift — kept in-distribution as mild domain randomization).

Built by policy/pi05-pure/scripts/convert_playground_to_lerobot.py in the RoboTwin fork.