tyb197/RoboAfford

Dataset Description

RoboAfford is a large-scale dataset with dense, affordance-aware annotations for instruction grounded manipulation. This dataset contains 819,987 images and 1.9 million QA pairs, unifying object affordances and spatial affordances to support interaction-centric learning in robotics.

Dataset Composition

RoboAfford aggregates images from multiple datasets and generates QA pairs to provide a comprehensive dataset for affordance understanding. It consists of the following components:

• LVIS_absxy_513K.json: 513K object detection QA pairs for 152,152 images sourced from LVIS

• pointing_absxy_190K.json: 190K object pointing QA pairs for 63,907 images selected from PixMo-Points

• object_affordance_prediction_absxy_561K.json: 561K object affordance prediction QA pairs for 45,790 images sourced from PACO-LVIS

• object_ref_max_points_10_absxy_347K.json: 347K object reference QA pairs for 287,956 images sourced from RoboPoint-data

• region_ref_max_points_10_absxy_320K.json: 320K region reference QA pairs for 270,182 images sourced from RoboPoint-data

Dataset Format

Each json file contains a list of structured conversations with image references. An example QA pair and its associated image are shown as follows:

{
    "id": "paco_403013",
    "image": "train2017/000000403013.jpg",
    "conversations": [
      {
        "from": "human",
        "value": "<image>\nWhat appliance can be used to heat food quickly? Your answer should be formatted as a list of tuples, i.e. [(x1, y1, x2, y2), ...], where each tuple contains the x, y coordinates of the top-left corner and bottom-right corner of the bounding box. The coordinates should be rounded to two decimal places, indicating the absolute pixel locations of the points in the image."
      },
      {
        "from": "gpt",
        "value": "[(258.14, 174.87, 283.79, 222.17)]"
      },
      {
        "from": "human",
        "value": "What appliance can be used to heat food quickly? Your answer should be formatted as a list of tuples, i.e. [(x1, y1), (x2, y2), ...], where each tuple contains the x and y coordinates of a point on the object. The coordinates should be rounded to two decimal places, indicating the absolute pixel locations of the points in the image."
      },
      {
        "from": "gpt",
        "value": "[(258.61, 213.0)]"
      }
    ]
  }

Evaluation

For benchmarking protocols and evaluation metrics, please refer to RoboAfford-Eval and https://github.com/tyb197/RoboAfford for more detailed information.