sudidaren/lightwm-data

LightWM Data (AI2-THOR)

Training data for the LightWM world model: RGB / depth / instance-segmentation frames with ground-truth object bounding boxes, agent poses, actions and action outcomes from the AI2-THOR simulator.

Contents

The dataset repo has three top-level folders:

Folder	Frames	Description
lightwm_data/	34,326	1,067 episodes / 117 scenes: exploration sweeps, task golden replays
lightwm_data_cov/	5,316	Exhaustive coverage of 12 scenes: every reachable position × 4 yaws × 3 horizons
lightwm_data_objviews/	25,719	Multi-angle object views: 56 small-object types × 15-30 views each (89 scenes)
fd_benchmark_full_20260811_224644/	3,618	Failure-detection benchmark trajectories with RGB + actions + errors

Directory layout

<root>/
  episodes/<scene>__<episode_id>/episode.json
  frames/step_00000_rgb.png     # first-person RGB, 800x600
  frames/step_00000_depth.png   # uint16 PNG, meters * 1000
  frames/step_00000_seg.png     # instance segmentation (colored)
  scene_gt/<scene>.json         # per-scene static object world positions (GT)

Each episode.json contains per-frame: action + success + error message, agent pose (position/rotation/horizon), and visible_objects with 2D bboxes and 3D positions.

Usage

export LIGHTWM_DATA_ROOT=<this_repo>/lightwm_data
export LIGHTWM_COV_ROOT=<this_repo>/lightwm_data_cov
export LIGHTWM_OBJVIEW_ROOT=<this_repo>/lightwm_data_objviews
export LIGHTWM_FD_ROOT=<this_repo>/fd_benchmark_full_20260811_224644
python shared/data_index.py

See the code repository README for full training/evaluation instructions.

Notes

• Data is synthetic (AI2-THOR 5.0.0, Unity). No real-world privacy concerns.
• Depth/segmentation are training-time supervision only; the runtime world model uses RGB alone.