Online 2-Stage 파이프라인 — 모델 I/O 명세 & 실측 예시

You are the task planner of a household robot in a simulator. Each turn you receive: the current scene image, an object inventory, and the outcome context of the previous subgoal. Decide exactly ONE next subgoal. Skills: - pick(object_id): grasp and lift the object - pick_and_place(object_id, receptacle_id): one atomic unit — you cannot split pick and place into separate subgoals - open/close: DISABLED (이 로봇의 학습 데이터에 없음 — 선택 금지) - done: the user instruction is fully satisfied - abort: impossible to proceed (explain in reasoning) Rules: - Refer to objects ONLY by the exact id in [brackets] from the inventory. - Output a single JSON object, nothing else: {"reasoning": "...", "skill": "...", "object_id": "...", "receptacle_id": "..."}

k=0 {"reasoning": "The user wants to place a mug on the dining table. I will pick up mug_3ebc...4a99_1_0_6 and place it on diningtable_f113...1c05_1_0_2.", "skill": "pick_and_place", "object_id": "mug_3ebc45568ed53a18c8797978b3744a99_1_0_6", "receptacle_id": "diningtable_f113cf7f8367e89f709b53cbee1a1c05_1_0_2"} → 실행: 900스텝 timeout k=1 {"reasoning": "The previous attempt timed out while trying to reach the dining table in room_2. I will try placing the mug on the dining table in room_3 instead.", ...} → 같은 머그, 다른 테이블로 전환. 900스텝 timeout k=2 {"reasoning": "Both dining tables in room_2 and room_3 timed out. I will try placing the mug on the dining table in room_4 instead.", ...} → 세 번째 테이블. 900스텝 timeout k=3 {"reasoning": "All attempts to place the mug onto the available dining tables timed out due to navigation or execution failure. Aborting the task.", "skill": "abort", ...} → 에피소드 종료 (final="abort")

actions: float (15, 8) # horizon 15 [:, 0:7] → 절대 관절 타깃 (rad) # 변환 없이 그대로 arm 명령 [:, 7] → 그리퍼 (이진화: >0.5 → 닫힘 255, else 0) 소비 규칙: 앞 8스텝만 실행 후 재추론 (chunk_size=8), 15Hz (policy_dt 66ms)