pick_panda_v2

Co-training set for the cathlab pick task ("Hand me the {bottle | vial | syringe}"), Franka Panda, panda_libero I/O: the three single-engine v1 datasets merged with lerobot-edit-dataset --operation.type merge (from their v1 tags).

Episodes	Source	Physics
0–383 (69,191 frames)	Rakeshdharan56/pick_robosuite@v1	robosuite 1.5.2 / MuJoCo 3.3.7
384–752 (45,387 frames)	Rakeshdharan56/pick_maniskill@v1	ManiSkill3 3.0.1 / PhysX 5
753–1135 (66,447 frames)	Rakeshdharan56/pick_genesis@v1	Genesis 1.4.5

1,136 episodes, 181,025 frames, 20 fps, robot_type panda: action float32[7], observation.state float32[8], observation.images.{agentview,eye_in_hand} 256×256. Same cathlab_v1 train grid in every engine (paired configs). Why: single-engine policies score 0–2 % in the other engines (cross-physics table in the repo); this set tests co-training. Code: https://github.com/Rakeshdharan/vla-cathlab-companion