| { | |
| "skill_list_Q1": "1. Functions for the actions of a gripper:\n 1.1 move_to(object, target_object): This function represents the movement of the gripper, with the first parameter representing the object held in the hand during movement and the second parameter representing the target object. If there is nothing in hand, the first parameter is' none '. For example, move_to(none, towel) means that the gripper is moving towards a towel with nothing in hand. And move_to(panda_toy, bowl) means that the gripper with a panda toy is moving towards a bowl.\n 1.2 hold(object): This function represents the static state of the gripper with an object. Note that if there is nothing in gripper, this function is not applicable. For example, hold(cup) means that the gripper is holding a cup and keeping static.\n2. Functions for grabbing and releasing:\n 2.1 pick_up(object): This function represents that the gripper picks up an object. Note that the object must be graspable. For example, pick_up(apple) means that the gripper picks up an apple.\n 2.2 grasp(object): This function represents that the gripper touches and lightly grabs the object. Note that objects can be either pick-upable or non-pick-upable. The difference between this function and pick_up is that grasp means to hold lightly, while pick_up usually means to lift up after contact. For example, grasp(door_handle) means that the gripper grabs the door handle (not lift up). In planning, before interacting with any new object, a grasp action must be executed first. If the object has a clearly defined affordance region (e.g., manipulating a door requires grasping the door handle), the grasp action should be performed on the affordance (e.g., grasp(door_handle)). If no explicit affordance is defined for the object, the grasp action should be applied directly to the object itself (e.g., grasp(orange)). \n 2.3 place(object, target_object): This function represents that the gripper place the object at the location of the target_object. Note that target_object can be a specific object or a position relative to a reference object. For example, place(apple, table) means that placing an apple on the table. And place(apple, right_of_banana) means that placing an apple on the right of a banana.\n3. Functions for using a tool to operate objects:\n 3.1 scoop(tool, contents, container): This function represents that the gripper holding the tool is scooping something in a container with the tool. If \"contents\" is uncertain, use \"unknown\" as the second parameter. For example, scoop(spoon, water, bowl) means that the gripper holds the spoon and uses it to scoop water from the bowl. If \"contents\" is uncertain, use \"unknown\" instead.\n 3.2 pour(container, contents, target_container): This function represents that the gripper holding the container is pouring something into the target_container. If \"contents\" is uncertain, use \"unknown\" as the second parameter. For example, pour(bowl, water, pot) means that the gripper holds a bowl with water and pours water into a pot. If \"contents\" is uncertain, use \"unknown\" instead.\n 3.3 wipe(tool, object, target_object): This function represents that the gripper is using a tool to wipe \"object\" on the \"target_object\". For example, wipe(towel, water, table) means that the gripper is wiping water on the table using a towel. If \"object\" is uncertain, use \"unknown\" instead.\n 3.4 stir(tool, contents, target_container): This function represents that the gripper is using a tool to stir \"contents\" in the \"target_container\". For example, stir(spoon, soup, pot) means that using a spoon to stir soup in the pot. If \"contents\" is uncertain, use \"unknown\" instead.\n 3.5 draw(tool, character, target_object): This function represents that the gripper is drawing a character using a tool on the target_object. For example, draw(marker, 'A', whiteboard) means that drawing an 'A' with a marker on the whiteboard.\n 3.6 cut(tool, object, target_object): This function represents that the gripper is cutting object with a tool on the target_object. For example, cut(knife, tomato, chopping_board) means that the gripper is cutting a tomato with a knife on the chopping board.\n4. Functions for interacting with objects directly:\n 4.1 fold(object, target_position): This function represents that the gripper is folding object towards the target_position. For example, fold(left_side_of_towel, right_side_of_towel) means that the gripper holds the left side of the towel and folds it to the right side.\n 4.2 unfold(object, target_position): This function represents that the gripper is unfolding object towards the target_position. For example, unfold(left_side_of_towel, right_side_of_towel) means that the gripper holds the left side of the towel and unfolds it to the right side.\n 4.3 turn(object, direction, state_of_target_object): This function represents that the gripper rotates the object in a certain direction to the target position or state. The 'direction' can be chosen in {clockwise, anticlockwise, up, down, forward, backward, left, right}. If it is an articulated object (it originally has axis rotation, such as a faucet), use the following {up, down, forward, backward, left, right}. If it is originally a rigid body (it originally has no axis rotation, such as a bottle adjusting its direction in the air), use clockwise and counterclockwise. For example, turn(faucet, clockwise, middle_of_sink) means that turn faucet right until it faces the middle of sink.\n 4.4 press(tool, object): This function represents that the gripper press object using a tool. If there is no tool, use 'none' instead. For example, press(none, red_button) means that pressing a red button using the gripper directly.\n 4.5 push(object, target_location): This function represents that the gripper pushes the object to the target_location. For example, push(chair, under_of_table) means that pushing a chair under to a table.\n 4.6 pull(object, target_location): This function represents that the gripper pulls the object to the target_location. For example, pull(towel, right_side_of_table) means that pulling a towel to the right side of a table.\n 4.7 insert(object, target_object): This function represents that the gripper inserts the object into the target object. For example, insert(plug, socket) means inserting a plug into the socket.\n 4.8 pullout(object, target_object): This function represents that the gripper pullouts the object from the target object. For example, pullout(plug, socket) means pullout a plug from the socket.\n5. Function for only dual-arm tasks:\n 5.1 transfer(left/right, right/left, object): This function represents the gripper transfers an object from one hand to another hand. For example, transfer(left, right, bottle) means that transfering bottle from left hand to right hand.\n6. Function for only mobile-manipulation tasks:\n 6.1 observation(object): This function represents that the target object is not in the field of view and needs to be found. For example, observation(chair) means that finding the chair.\n 6.2 mobile(target_object): This function represents that the target object is in the field of view but too far from the robot to operate, so the robot needs to move the chassis to approach the target position. For example, mobile(table) means moving the chassis to approach the table.\n\n**Examples**\nSingle-arm Task Example: Put tea bag in a cup\n1-move_to(none, tea_bag),2-grasp(tea_bag),3-pick_up(tea_bag),4-move_to(tea_bag, cup),5-place(tea_bag, cup)\n\nDual-arm Task Example: Placing a lid on a spray bottle\n1-left:move_to(none, spray_bottle), right:move_to(none, lid_of_the_spray_bottle),2-left:grasp(spray_bottle), right:grasp(lid_of_spray_bottle), 3-left:no_ops, right:pick_up(lid_of_spray_bottle), 4-left:no_ops, right:place(lid_of_spray_bottle, spray_bottle)\n\nMobile-manipulation Task Examples\nMobile-manipulation-with-single-arm Example: Finding a book and placing it on a shelf\n1-observation(book), 2-mobile(table), 3-move_to(none, book), 4-pick_up(book), 5-observation(bookshelf), 6-mobile(bookshelf), 7-move_to(book, empty_spot_on_shelf), 8-place(book, empty_spot_on_shelf)\n\nMobile-manipulation-with-dual-arm Example: Moving a large box to a table\n1-observation(large_box), 2-mobile(large_box), 3-left: move_to(none, left_side_of_box), right: move_to(none, right_side_of_box), 4-left: pick_up(large_box), right: pick_up(large_box), 5-left: hold(large_box), right: hold(large_box), 6-observation(table), 7-mobile(table), 8-left: place(large_box, table), right: place(large_box, table)\n\nOutput skills can only be selected from the predefined skill library.\n\n", | |
| "skill_list_Q2": "1. Functions for the actions of a gripper:\n 1.1 move_to(object, target_object): This function represents the movement of the gripper, with the first parameter representing the object held in the hand during movement and the second parameter representing the target object. If there is nothing in hand, the first parameter is' none '. For example, move_to(none, towel) means that the gripper is moving towards a towel with nothing in hand. And move_to(panda_toy, bowl) means that the gripper with a panda toy is moving towards a bowl.\n 1.2 hold(object): This function represents the static state of the gripper with an object. Note that if there is nothing in gripper, this function is not applicable. For example, hold(cup) means that the gripper is holding a cup and keeping static.\n2. Functions for grabbing and releasing:\n 2.1 pick_up(object): This function represents that the gripper picks up an object. Note that the object must be graspable. For example, pick_up(apple) means that the gripper picks up an apple.\n 2.2 grasp(object): This function represents that the gripper touches and lightly grabs the object. Note that objects can be either pick-upable or non-pick-upable. The difference between this function and pick_up is that grasp means to hold lightly, while pick_up usually means to lift up after contact. For example, grasp(door_handle) means that the gripper grabs the door handle (not lift up). In planning, before interacting with any new object, a grasp action must be executed first. If the object has a clearly defined affordance region (e.g., manipulating a door requires grasping the door handle), the grasp action should be performed on the affordance (e.g., grasp(door_handle)). If no explicit affordance is defined for the object, the grasp action should be applied directly to the object itself (e.g., grasp(orange)). \n 2.3 place(object, target_object): This function represents that the gripper place the object at the location of the target_object. Note that target_object can be a specific object or a position relative to a reference object. For example, place(apple, table) means that placing an apple on the table. And place(apple, right_of_banana) means that placing an apple on the right of a banana.\n3. Functions for using a tool to operate objects:\n 3.1 scoop(tool, contents, container): This function represents that the gripper holding the tool is scooping something in a container with the tool. If \"contents\" is uncertain, use \"unknown\" as the second parameter. For example, scoop(spoon, water, bowl) means that the gripper holds the spoon and uses it to scoop water from the bowl. If \"contents\" is uncertain, use \"unknown\" instead.\n 3.2 pour(container, contents, target_container): This function represents that the gripper holding the container is pouring something into the target_container. If \"contents\" is uncertain, use \"unknown\" as the second parameter. For example, pour(bowl, water, pot) means that the gripper holds a bowl with water and pours water into a pot. If \"contents\" is uncertain, use \"unknown\" instead.\n 3.3 wipe(tool, object, target_object): This function represents that the gripper is using a tool to wipe \"object\" on the \"target_object\". For example, wipe(towel, water, table) means that the gripper is wiping water on the table using a towel. If \"object\" is uncertain, use \"unknown\" instead.\n 3.4 stir(tool, contents, target_container): This function represents that the gripper is using a tool to stir \"contents\" in the \"target_container\". For example, stir(spoon, soup, pot) means that using a spoon to stir soup in the pot. If \"contents\" is uncertain, use \"unknown\" instead.\n 3.5 draw(tool, character, target_object): This function represents that the gripper is drawing a character using a tool on the target_object. For example, draw(marker, 'A', whiteboard) means that drawing an 'A' with a marker on the whiteboard.\n 3.6 cut(tool, object, target_object): This function represents that the gripper is cutting object with a tool on the target_object. For example, cut(knife, tomato, chopping_board) means that the gripper is cutting a tomato with a knife on the chopping board.\n4. Functions for interacting with objects directly:\n 4.1 fold(object, target_position): This function represents that the gripper is folding object towards the target_position. For example, fold(left_side_of_towel, right_side_of_towel) means that the gripper holds the left side of the towel and folds it to the right side.\n 4.2 unfold(object, target_position): This function represents that the gripper is unfolding object towards the target_position. For example, unfold(left_side_of_towel, right_side_of_towel) means that the gripper holds the left side of the towel and unfolds it to the right side.\n 4.3 turn(object, direction, state_of_target_object): This function represents that the gripper rotates the object in a certain direction to the target position or state. The 'direction' can be chosen in {clockwise, anticlockwise, up, down, forward, backward, left, right}. If it is an articulated object (it originally has axis rotation, such as a faucet), use the following {up, down, forward, backward, left, right}. If it is originally a rigid body (it originally has no axis rotation, such as a bottle adjusting its direction in the air), use clockwise and counterclockwise. For example, turn(faucet, clockwise, middle_of_sink) means that turn faucet right until it faces the middle of sink.\n 4.4 press(tool, object): This function represents that the gripper press object using a tool. If there is no tool, use 'none' instead. For example, press(none, red_button) means that pressing a red button using the gripper directly.\n 4.5 push(object, target_location): This function represents that the gripper pushes the object to the target_location. For example, push(chair, under_of_table) means that pushing a chair under to a table.\n 4.6 pull(object, target_location): This function represents that the gripper pulls the object to the target_location. For example, pull(towel, right_side_of_table) means that pulling a towel to the right side of a table.\n 4.7 insert(object, target_object): This function represents that the gripper inserts the object into the target object. For example, insert(plug, socket) means inserting a plug into the socket.\n 4.8 pullout(object, target_object): This function represents that the gripper pullouts the object from the target object. For example, pullout(plug, socket) means pullout a plug from the socket.\n5. Function for only dual-arm tasks:\n 5.1 transfer(left/right, right/left, object): This function represents the gripper transfers an object from one hand to another hand. For example, transfer(left, right, bottle) means that transfering bottle from left hand to right hand.\n6. Function for only mobile-manipulation tasks:\n 6.1 observation(object): This function represents that the target object is not in the field of view and needs to be found. For example, observation(chair) means that finding the chair.\n 6.2 mobile(target_object): This function represents that the target object is in the field of view but too far from the robot to operate, so the robot needs to move the chassis to approach the target position. For example, mobile(table) means moving the chassis to approach the table.\n\n**Examples**\nSingle-arm Task Next Step Example: Q: With <Put tea bag in a cup> as the goal, steps so far are shown in the video, what is the next step to do?\nA: 2-grasp(tea_bag)\n\nDual-arm Task Next Step Example: Q: With <Placing a lid on a spray bottle> as the goal, steps so far are shown in the video, what is the next step to do?\nA: 3-left:no_ops, right:pick_up(lid_of_spray_bottle)\n\nMobile-manipulation Task Next Step Examples\nMobile-manipulation-with-single-arm Next Step Example: Q: With <Finding a book and placing it on a shelf> as the goal, steps so far are shown in the video, what is the next step to do?\nA: 2-mobile(table)\n\nMobile-manipulation-with-dual-arm Next Step Example: Q: With <Moving a large box to a table> as the goal, steps so far are shown in the video, what is the next step to do?\nA: 4-left: pick_up(large_box), right: pick_up(large_box)\n\nOutput skills can only be selected from the predefined skill library.\n\n", | |
| "skill_list_Q3": "Please only answer one word, yes or no.\n\n**Examples**\nSingle-arm Task Example: Q: With <Put tea bag in a cup> as the goal, steps so far are shown in the video, is step 1-move_to(none, tea_bag),2-grasp(tea_bag),3-pick_up(tea_bag) finished?\nA: No\n\nDual-arm Task Example: Q: With <Placing a lid on a spray bottle> as the goal, steps so far are shown in the video, is step 1-left:move_to(none, spray_bottle), right:move_to(none, lid_of_the_spray_bottle) finished?\nA: Yes\n\nMobile-manipulation Task Examples\nMobile-manipulation-with-single-arm Example: Q: With <Finding a book and placing it on a shelf> as the goal, steps so far are shown in the video, is step 1-observation(book), 2-mobile(table), 3-move_to(none, book), 4-pick_up(book), 5-observation(bookshelf), 6-mobile(bookshelf), 7-move_to(book, empty_spot_on_shelf), 8-place(book, empty_spot_on_shelf) finished?\nA: No\n\nMobile-manipulation-with-dual-arm Example: Q:With <Moving a large box to a table> as the goal, steps so far are shown in the video, is step 1-observation(large_box), 2-mobile(large_box), 3-left: move_to(none, left_side_of_box), right: move_to(none, right_side_of_box) finished?\nA: Yes\n\nOutput skills can only be selected from the predefined skill library.\n\n", | |
| "navigation_skill_list": "Functions for the navigation task: 1. turn_left(): This function represents the robot rotating 90 degrees to the left in place.\n 2. turn_right(): This function represents the robot rotating 90 degrees to the right in place.\n 3. turn_back(): This function represents the robot rotating 180 degrees in place (turning around).\n 4. go_forward(target_location): This function represents the robot moving forward in its current facing direction until it reaches the target_location. The parameter should be the name or description of the target location. For example: go_forward(dining_table) means the robot moves forward until it reaches the dining table. go_forward(kitchen_door) means the robot moves forward until it arrives at the kitchen door.\n\n**Examples:**\nNavigate to the stove\n1-go_forward(fireplace), 2-turn_left(), 3-go_forward(stove)\n\n", | |
| "robot_type": "You are given a multiple-choice question about the type of robotic arm shown in a video. The possible answers are defined as follows:\n\n(A) Single arm with gripper: A single robotic arm with a simple gripper (e.g., two-finger claw) at the end. It is capable of basic grasping but does not have fine manipulation abilities.\n\n(B) Bi-manual with gripper: A robot with two arms, each equipped with a simple gripper. It can perform tasks that require coordinated use of both arms.\n\n(C) Single arm with dexterous hand: A single robotic arm equipped with a dexterous (multi-fingered) hand, capable of fine and complex manipulation, similar to a human hand.\n\n(D) Mobile Manipulation: A robot that combines robotic arm with a mobile base, allowing it to navigate its environment and manipulate objects across locations.\n\n", | |
| "robot_view": "You are given a multiple-choice question about the type of camera view shown in a video. The possible answers are defined as follows:\n\n(A) Exo: An external (third-person) view of the robot or scene, typically captured from a fixed camera in the environment. The robot is fully visible from the outside.\n\n(B) Front Ego: A first-person view captured from the front of the robot, such as a chest-mounted or torso-mounted camera. The view moves with the robot's body and shows the scene ahead.\n\n(C) Hand Ego: A first-person view from a camera mounted near the robot's hand or end-effector. Often used to capture manipulation tasks closely from the hand's perspective.\n\n(D) Head Ego: A first-person view from a camera mounted on or near the robot's head or eye-level, providing a perspective similar to human vision.\n\n", | |
| "spatial_relation": "You are given a 2D image and asked to identify the type of visual perspective. The possible camera reference frames are defined as follows:\n\nEgocentric (ego-centric): The camera is mounted on the agent (e.g., a robot or person) who are sterring the image. The view reflects how the agent \"sees\" the image from its own perspective, often first-person.\n\nWorld-centric: The camera is placed in a fixed position in the environment, independent of any agent or object. It provides a stable, third-person view of the scene.\n\nObject-centric: The camera view is centered on or moves with a specific object of interest. The object remains relatively central or stationary in the frame, even if the background or other entities move.\n\n", | |
| "spatial_temporal_causality": "Given a question and an image with a set of points on it, the points format is (x, y). In this format, x and y typically represent the coordinates of the point. Note that the horizontal length of the image is the x-axis, and the vertical height is the y-axis. The coordinates have been normalized (between 0 and 1). Please answer the question based on the points.\n\n", | |
| "high_level_planning_error": "You are given a video showing a robot performing a task. Your goal is to identify what kind of planning error has occurred, if any, based on the following definitions:\n\nA: Object mistaken error — The robot selects or manipulates the wrong object (e.g., picking up the wrong item instead of the intended one).\n\nB: Incomplete error — The robot starts the task but fails to finish it (e.g., picks up an object but does not place it, or drops it midway).\n\nC: Wrong order error — The robot performs the correct subtasks but in the wrong sequence (e.g., placing an object before picking it up, or doing step 2 before step 1).\n\nD: Correct — The task is executed correctly, with appropriate object selection, full completion, and correct order.\n\n", | |
| "low_level_execution_error": "You are given a video showing a robot attempting to execute a task. Your goal is to identify the type of execution error that has occurred, if any, based on the following definitions:\n\nA: Position error — The robot attempts the right action but at the wrong spatial position (e.g., trying to grasp an object but missing its actual location).\n\nB: Trajectory error — The robot follows an incorrect or unstable movement path (e.g., a shaky, inefficient, or unintended trajectory between start and end points).\n\nC: End-effector open-close error — The robot opens or closes its gripper at the wrong time or fails to open/close it as needed (e.g., drops the object prematurely or fails to grasp it).\n\nD: Incomplete error — The robot begins the action but does not complete it (e.g., starts moving toward the object but stops before grasping or placing).\n\n", | |
| "Transparency": "Transparency refers to how much can be seen through an object. A transparent object can be clearly seen through, almost as if it was not there. A translucent object can be seen through with some details, but not as clearly as if it was transparent. An opaque object cannot be seen through at all. The transparency of an object does not refer to the transparency of its contents if it has anything inside it.\n\n", | |
| "Deformability": "Deformability refers to how easily an object can change shape without breaking.\n\n", | |
| "Fragility": "Fragility refers to how easily an object can be broken or damaged.\n\n", | |
| "Mass": "The heaviness of an object refers to its mass. It includes the contents of the object if it has something inside it.\n\n", | |
| "Material": "The material of an object refers to what material makes up the largest portion of the object that is visible. It does not refer to the contents of a container.\n\n", | |
| "Content": "Only respond with contents that are clearly visible and identifiable.\n\n", | |
| "Liquid": "A container can contain liquid if it can be used to transport a liquid across a room without a person needing to be particularly careful about not spilling it.\n\n", | |
| "Sealed": "A container is sealed if it can be rotated by any amount in any direction without spilling its contents if it has anything inside.\n\n", | |
| "multi-view": "When you planning the steps to encounter the goal, please combine images from different perspectives, in order to recognize occluded objects or select arms for operation based on the distance between each arm and the target object.\n\n", | |
| "afforadance_point": "Given a question and an image with a set of points on it, the points format is (x, y). In this format, x and y typically represent the coordinates of the point, which has been normalized to the range [0, 1]. Please answer the question based on the points.\n\n", | |
| "MCQ_post_prompt": "\n Please answer only one word, e.g., A, B, C, or D.\n\n", | |
| "color": "When you planning the steps to encounter the goal, replace color-described objects with the number in your responses according to the visual prompt in the image. For example, when given a goal like 'put a white object into a yellow object' (and combined with the visual prompt in the image), your response should be specific and practical, such as '..., place(2, 6), ...'.\n\n ", | |
| "number": "When you planning the steps to encounter the goal, please align with the number of objects specified in the image. For example, when given a goal like 'put apple in the cabinet' (and combined with relevant visual context), your response should be specific and practical, such as '..., place(first_apple, cabinet), place(second_apple, cabinet), ...'.\n\n ", | |
| "shape": "When you planning the steps to encounter the goal, replace shape-described objects with the number in your responses according to the visual prompt in the image. For example, when given a goal like 'pick up a spherical object' (and combined with the visual prompt in the image), your response should be specific and practical, such as '..., pick_up(5), ...'.\n\n ", | |
| "size": "When you planning the steps to encounter the goal, consider the inherent size characteristics of the number in the image to ensure they fit the goal. For example, when given a goal like 'sort fruits by size' (and combined with the visual prompt in the image), your response should be specific and practical, such as '..., pick_up(2), pick_up(3), pick_up(4), ...'.\n\n " | |
| } |