Panda512/rokae_all_tasks_full

ROKAE unified full training dataset

This LeRobot v3 dataset combines the complete new_pour, new_pick_place, and new_stack datasets. It contains 1,447 episodes and 145,848 frames. All episodes are in the train split; there is no offline validation/test holdout in this final-fit dataset.

The 57D observation.state stores physical state in channels 0:30, skill one-hot in 30:38, position one-hot in 38:43, object one-hot in 43:46, and progress Fourier features in 46:57. The root encoding.json, labels.parquet, source_map.parquet, and merge_manifest.json preserve the schema and trace each merged episode to its source task dataset.

Train with the PI0.5 condition-token training package by setting the HF dataset repo ID to Panda512/rokae_all_tasks_full and disabling validation split creation (eval_split=0.0).

The new_stack clips use grasp the cup at the lower right/upper right for approach, grasp and retreat, and stack into top1/top2 for transport, stack and return. Stack task object conditions are cup; original source object/text are retained in the annotation sidecars. Four pickup/target planner compositions are listed in stack_composition_spec.json. Existing atomic clips are retained without synthesizing continuous demonstrations.

All three task profiles now use the same task text and object conditions as planner/task_spec: kettle grasp + pour into the centre; bread grasp + place into the centre; cup grasp + stack into top1/top2. Original source task/object are retained in annotation sidecars.