# ABot-Style Caption Report v3 Total samples: 1500 Dense caption eligible: 1082 ## Task Family Counts - pick_place: 180 (12.0%) - object_to_container: 170 (11.3%) - button_press_click: 150 (10.0%) - articulated_open_close: 130 (8.7%) - tool_use: 120 (8.0%) - stacking: 120 (8.0%) - lifting: 100 (6.7%) - coverage_unknown: 100 (6.7%) - scanning_qrcode: 80 (5.3%) - shaking: 70 (4.7%) - rotation_orientation: 70 (4.7%) - dumping_pouring: 60 (4.0%) - ranking_arrangement: 60 (4.0%) - handover: 50 (3.3%) - hanging: 40 (2.7%) ## Parser / Quality Counts - role_parse_failed: 3 - manipulated_object_missing: 29 - target_missing: 217 - low_confidence_dense_caption: 418 - unresolved_pronoun: 11 - compound_instruction: 83 - secondary_action_parsed: 83 - target_pollution_fixed: 43 - dense_caption_too_generic: 418 - generic_object: 29 - unknown_task_family: 100 - needs_vlm_review: 103 ## Stage 1 JSON Example ```json { "high_level_task": { "task_family": "pick_place", "robotwin_task_name": "place_a2b_left", "intent": "place the pink toycar shaped like real car to the left of the palm-sized phone", "source_instruction": "Place the pink toycar shaped like real car directly to the left of the palm-sized phone." }, "role_parse": { "active_arm": "unknown", "main_verb": "place", "manipulated_object": "the pink toycar shaped like real car", "target_object": "the palm-sized phone", "target_receptacle": "", "spatial_relation": "left of", "secondary_action": "", "primary_action": "place_relative", "pre_action": "", "compound_instruction": "false", "unresolved_pronoun": "false", "target_pollution_fixed": "false", "parse_confidence": 0.95 }, "robot_initial_state": { "morphology": "Aloha-AgileX-style dual-arm gripper robot", "visible_assumption": "robot arms may enter the fixed camera view from outside the frame", "initial_contact": "no contact is assumed before approach unless visible in the video" }, "scene_context": { "location_type": "fixed tabletop robotic workspace", "camera_view": "static head-camera view, 640x480 video rendered from a fixed workspace camera", "objects": [ "the pink toycar shaped like real car" ], "static_objects": [ "the palm-sized phone" ] }, "skill_sequence": [ { "skill_index": 1, "verb": "approach", "active_arm": "unknown", "object": "the pink toycar shaped like real car", "target": "workspace contact region", "description": "The active gripper approaches the pink toycar shaped like real car while staying in the fixed tabletop workspace." }, { "skill_index": 2, "verb": "contact_or_grasp", "active_arm": "unknown", "object": "the pink toycar shaped like real car", "target": "the pink toycar shaped like real car", "description": "The gripper or tool should make visible or physically plausible contact with the pink toycar shaped like real car before the main motion." }, { "skill_index": 3, "verb": "place_relative", "active_arm": "unknown", "object": "the pink toycar shaped like real car", "target": "the palm-sized phone", "description": "The robot performs the action: place the pink toycar shaped like real car to the left of the palm-sized phone. Parsed relation: left of." }, { "skill_index": 4, "verb": "finish", "active_arm": "unknown", "object": "the pink toycar shaped like real car", "target": "the palm-sized phone", "description": "The motion ends with the pink toycar shaped like real car in a stable state relative to the palm-sized phone." } ], "key_state_changes": [ "the pink toycar shaped like real car changes pose or state according to the task intent", "the gripper state may open or close near the contact moment", "the final state should preserve the relation between the pink toycar shaped like real car and the palm-sized phone" ], "physical_relations": { "contact": "motion of the pink toycar shaped like real car should be explained by gripper, end-effector, or tool contact", "support": "the pink toycar shaped like real car should remain supported by the gripper, tabletop, receptacle, or target structure", "containment": "not required unless visible", "violations_to_avoid": [ "object teleportation", "unexplained floating motion", "severe penetration through rigid objects", "non-rigid deformation of the robot or target object" ] }, "task_result_note": "Use visible evidence from the clip; if an object or contact is not visible, keep the description generic.", "caption_quality_flags": [ "template_exact" ] } ``` ## Stage 2 Dense Caption Example [Phase 1] Scene Setup A fixed head-camera view observes an Aloha-AgileX-style dual-arm gripper robot in a tabletop workspace. The parsed manipulated object is the pink toycar shaped like real car. The parsed target or reference is the palm-sized phone, and the intended spatial relation is left of. The task family is pick_place, so the description focuses on visible rigid-body manipulation rather than appearance embellishment. [Phase 2] Action Detail The robot uses an active gripper to perform the action: place the pink toycar shaped like real car to the left of the palm-sized phone. The motion should begin with an approach toward the pink toycar shaped like real car, followed by contact, grasping, pressing, or tool interaction as appropriate for the task. During the main action, the pink toycar shaped like real car should move with the gripper or tool rather than independently, and the trajectory should remain continuous in the fixed camera view. [Phase 3] State Transition The expected state transition is that the pink toycar shaped like real car moves from its initial tabletop pose to left of the palm-sized phone. The final arrangement should be stable and physically supported by the tabletop, the gripper, the receptacle, or the referenced object. The clip should not show object teleportation, unexplained floating, severe penetration, or deformation of rigid objects. [Phase 4] Camera Summary The camera remains static and documents the manipulation in an observation-style view. White tabletop background, mild simulator grain, and partial robot visibility are normal for this domain. The caption avoids claiming contact or object identity beyond what is parsed from the instruction and task metadata. ## Recommended Caption Mix - SFT: 50% caption_worldarena_style, 30% caption_short, 20% abot_stage2_dense_caption, but dense captions only for rows without low-confidence flags. - A2V: 70% caption_short, 30% caption_worldarena_style, 0% dense caption by default. - Rows marked role_parse_failed, target_missing, manipulated_object_missing, or low_confidence_dense_caption should use short/worldarena prompts until VLM captioning is available.