Datasets:

ACT3D
/

Act3D-data

pretty_name: Act3D-data
license: other
task_categories:
  - visual-question-answering
tags:
  - 3d
  - embodied
  - spatial-reasoning
  - reinforcement-learning
configs:
  - config_name: rl
    data_files:
      - split: train
        path: annotations/rl_train.parquet
  - config_name: sft
    data_files:
      - split: train
        path: annotations/sft_train.parquet
      - split: validation
        path: annotations/sft_val.parquet

Act3D-data

Training data for Act3D: Learning to Think and Explore in 3D World (paper, code).

This repository contains annotations only: questions, answers, answer boxes, camera poses and exploration trajectories. It contains no scene data and no images.

Every scene comes from ScanNet, ARKitScenes or Matterport3D. Their licenses do not allow us to redistribute meshes or anything rendered from them.
You download the meshes yourself from the original providers.
build_images.py then renders every image locally. The trajectory images come out exactly as they were when we trained.

file	rows	what it is
`annotations/rl_train.parquet`	51,543	RL (GRPO) questions: one question + one start pose per row
`annotations/sft_train.parquet`	4,949	SFT trajectories (think / move / observe / answer)
`annotations/sft_val.parquet`	213	SFT validation trajectories
`annotations/arkit_eval.jsonl`	1,000	ARKit evaluation questions (ours; ARKitScenes 3DOD `Training` scenes, room-disjoint from RL training, see below), read by the code's `act3d/eval/prepare_eval.py`
`annotations/meshes.sha256`		checksums of every mesh file the training annotations use
`build_images.py`		renders the images and writes the training parquets
`prep_mp3d_regions.py`		crops Matterport3D houses into the per-room meshes used here

RL questions by source:

`data_source`	rows	scenes	questions from
`scanqa`	25,563	ScanNet	ScanQA (Azuma et al., CVPR 2022)
`genqa`	8,591	ScanNet	ours: generated from the scene geometry, verified by an LLM
`mp3d`	9,855	Matterport3D regions	MMScan (Lyu et al., NeurIPS 2024), MP3D subset: ids only, see below
`genqa_ark`	7,534	ARKitScenes 3DOD	ours: generated from the scene geometry, verified by an LLM

Fields

RL (rl_train.parquet): index, data_source, scene_id, mesh_path, question, answers (JSON list), answer_boxes (JSON list of [cx, cy, cz, sx, sy, sz] in metres, the boxes of the objects the answer refers to; used by the exploration reward), qtype, ground_truth, and start_eye / start_at (the start camera: position and look-at point, z up).

SFT (sft_*.parquet): qid, scene_id, src, mesh_path, messages (system / user / assistant / tool turns; every <image> is one rendered view), start_eye / start_at, start_pose_index, actions (the action applied by each tool turn that returned an image), bounds_min / bounds_max (the scene box used to clip camera moves) and n_images.

mesh_path is always relative to $ACT3D_ROOT.

MMScan questions are not included.

The Matterport3D questions come from MMScan, which is distributed under its own access terms.
Those rows carry only mmscan_id, the id of the entry in MMScan's LLaVA-3D_MMScan_QA.json:
- in RL rows, question, answers and ground_truth are empty;
- in SFT rows, the question in the user turn is {MMSCAN_QUESTION}, and the MMScan answer is {MMSCAN_ANSWER} (in the final <answer>, and in the reasoning where it is quoted verbatim).
build_images.py --mmscan fills them back in from your own copy of MMScan.

Setup

1. Get the scene meshes

Request access from each provider and accept its terms of use. Then place the files below under $ACT3D_ROOT. Only these files are read; you don't need the RGB-D streams.

$ACT3D_ROOT/
├── annotations/                     # this repository
├── scannet/scans/<scene_id>/
│   ├── <scene_id>_vh_clean.ply      # dense mesh: required, it is what we rendered
│   └── <scene_id>_vh_clean_2.ply    # decimated mesh: the name mesh_path points to
├── arkitscenes/3dod/Training/<video_id>/
│   └── <video_id>_3dod_mesh.ply
└── mp3d/regions/
    └── <house_id>_region<N>.ply     # made by prep_mp3d_regions.py

ScanNet v2 (https://github.com/ScanNet/ScanNet): sign the ScanNet Terms of Use. With the official download-scannet.py, fetch --type _vh_clean.ply and --type _vh_clean_2.ply for the scenes listed in annotations/meshes.sha256, into scannet/scans/. The renderer automatically uses _vh_clean.ply next to the _vh_clean_2.ply that mesh_path names. Without the dense mesh, renders will not match the training data.
ARKitScenes (https://github.com/apple/ARKitScenes): download the 3DOD Training split with the official download_data.py 3dod --split Training. Keep <video_id>/<video_id>_3dod_mesh.ply under arkitscenes/3dod/Training/.
Matterport3D (https://niessner.github.io/Matterport/): sign the Matterport3D Terms of Use and download house_segmentations for the houses in annotations/meshes.sha256. Then crop the rooms with:
```
python prep_mp3d_regions.py --scans /path/to/v1/scans --root $ACT3D_ROOT
```
This writes each <house>_region<N>.ply: the house mesh cropped to that region's box from <house>.house, plus 0.15 m.

Check your files (from $ACT3D_ROOT):

cd $ACT3D_ROOT && sha256sum -c annotations/meshes.sha256 --quiet

ScanNet and ARKitScenes files should match exactly.
The MP3D crops can differ byte-wise across Open3D versions (we used Open3D 0.19); renders are unaffected.

2. Get the MMScan questions and ScanQA val

Request access to MMScan (https://github.com/OpenRobotLab/EmbodiedScan, Hugging Face 3dllm/MMScan-beta) and download LLaVA-3D_MMScan_QA.json.

RL validation is the full official ScanQA v1.0 val split (4,675 questions; not redistributed here). Download ScanQA_v1.0_val.json from https://github.com/ATR-DBI/ScanQA (dataset link in docs/dataset.md) to $ACT3D_ROOT/benchmarks/scanqa/ScanQA_v1.0_val.json, and get its 71 ScanNet scenes the same way as the training scenes.

3. Render the images

With the Act3D code checked out (it provides the renderer, act3d/act3d/env.py):

export ACT3D_ROOT=/path/to/act3d_data
PYTHONPATH=/path/to/Act3D/act3d python build_images.py --mmscan /path/to/LLaVA-3D_MMScan_QA.json --workers 32

This runs on CPU and writes the files the training scripts read:

$ACT3D_ROOT/sft/train.parquet, sft/val.parquet   # messages + images (verl multi-turn SFT format)
$ACT3D_ROOT/rl/train.parquet,  rl/val.parquet    # prompt + start view + extra_info + reward_model
                                                 # (rl/val.parquet = the full ScanQA val split)

SFT images: each trajectory is replayed from its start pose with the recorded actions. On our machine this reproduced all training images pixel-exactly (see "Verification").
RL start views: these are rendered with the same renderer at the start pose. The views baked into our original RL parquets came from an earlier pipeline that used the raw ScanNet camera (.sens intrinsics and 6-DoF pose). The release renders the same position and direction with the Act3D renderer instead, so the start view matches the tool's own renders.

Then follow the training and evaluation steps in the code README.

License and attribution

Our annotations (GenQA questions, answer boxes, trajectories, poses): CC BY-NC-SA 4.0.
ScanQA questions follow ScanQA's license.
MMScan questions are not redistributed: only their ids are, and you get the text from MMScan under its terms.
Scenes: you must obtain ScanNet, ARKitScenes and Matterport3D yourself and follow their terms. Nothing derived from their imagery is included here.

Please cite Act3D and the source datasets:

@article{li2026act3d,
  title   = {Learning to Think and Explore in 3D World},
  author  = {Li, Yijiang and Wang, Bingyang and Vasconcelos, Nuno},
  journal = {arXiv preprint arXiv:2610.21783},
  year    = {2026}
}

ScanNet (Dai et al., CVPR 2017), ARKitScenes (Baruch et al., NeurIPS 2021 Datasets and Benchmarks), Matterport3D (Chang et al., 3DV 2017), ScanQA (Azuma et al., CVPR 2022), MMScan (Lyu et al., NeurIPS 2024).