ENACT: Evaluating Embodied Cognition with World Modeling of Egocentric Interaction

ENACT is a benchmark dataset for evaluating embodied cognition in vision–language models via egocentric world modeling. It probes whether models can reason about how the world changes under sequences of actions, using long-horizon household activities in a mobile manipulation setting.

• Project page: https://enact-embodied-cognition.github.io/
• Code & evaluation: https://github.com/mll-lab-nu/ENACT

Dataset Summary

Each ENACT example is a multi-image, multi-step reasoning problem built from robot trajectories:

• Forward world modeling

  • Input: one current state image, several future state images (shuffled), and a list of actions in correct order.
  • Task: output a Python list of integers giving the correct chronological order of future images (e.g., [1, 3, 2]).

• Inverse world modeling

  • Input: an ordered sequence of images showing state changes, plus actions in shuffled order.
  • Task: output a Python list of integers giving the correct chronological order of actions (e.g., [2, 3, 1]).

All images are egocentric RGB observations rendered from long-horizon household tasks (e.g., assembling gift baskets, bringing water, preparing lunch boxes, cleaning up a desk).

File Structure

After unpacking, the dataset has the following structure:

.
├── enact_ordering.jsonl        # All QA examples (one JSON per line)
└── images/
    ├── forward_world_modeling_3_steps/
    ├── forward_world_modeling_4_steps/
    ├── ...
    ├── forward_world_modeling_10_steps/
    ├── inverse_world_modeling_3_steps/
    ├── ...
    └── inverse_world_modeling_10_steps/

Each task folder (e.g., forward_world_modeling_3_steps/) contains one subfolder per activity, such as:

images/forward_world_modeling_3_steps/
  ├── assembling_gift_baskets_1749468508582193/
  ├── bringing_water_1750844141719178/
  ├── ...

Inside each activity folder are the PNGs for that trajectory (current state and future states, or ordered states in the inverse setting).

JSONL Format

Each line in enact_ordering.jsonl is a JSON object:

{
  "id": "assembling_gift_baskets_1749468508582193_forward_world_modeling_3_steps_cfbcc15c",
  "type": "forward_world_modeling_3_steps",
  "task_name": "assembling_gift_baskets_1749468508582193",
  "key_frame_ids": ["4150", "11360", "11834"],
  "images": [
    "QA/images/forward_world_modeling_3_steps/..._cur_state.png",
    "QA/images/forward_world_modeling_3_steps/..._next_state_1.png",
    "QA/images/forward_world_modeling_3_steps/..._next_state_2.png"
  ],
  "question": "...natural language instructions and actions...",
  "options": [],
  "gt_answer": [1, 2]
}

• id – unique identifier for this QA instance.

• type – question type and horizon, e.g. forward_world_modeling_3_steps or inverse_world_modeling_4_steps.

• task_name – underlying household task instance.

• key_frame_ids – frame indices of selected key frames in the trajectory.

• images – relative paths to PNG images:

  • index 0 is the current state;
  • subsequent entries are future states (forward) or later states (inverse).

• question – natural language prompt specifying the task setup, actions, and the required output as a Python list of integers.

• gt_answer – ground-truth ordering of image or action labels (list of integers, e.g. [1, 3, 2]).

Usage

To evaluate, follow the scripts in the code repository: https://github.com/mll-lab-nu/ENACT

Citation

If you use ENACT, please cite the paper:

@article{wang2025enact,
  title={ENACT: Evaluating Embodied Cognition with World Modeling of Egocentric Interaction},
  author={Wang, Qineng and Huang, Wenlong and Zhou, Yu and Yin, Hang
          and Bao, Tianwei and Lyu, Jianwen and Liu, Weiyu and Zhang, Ruohan
          and Wu, Jiajun and Li, Fei-Fei and Li, Manling}
}