HM3D-M3 / README.md
skyhyding's picture
Fix code/release links to NJU hm3d-m3-code; refresh README checksum
1559e3d verified
|
Raw History Blame Contribute Delete
3.51 kB
---
license: cc-by-nc-4.0
pretty_name: HM3D-M3 (train split and teacher samples)
task_categories:
- robotics
tags:
- embodied-ai
- multi-agent
- object-navigation
- instance-image-navigation
- habitat
- hm3d
size_categories:
- 100K<n<1M
extra_gated_prompt: >-
The episodes refer to Habitat-Matterport 3D (HM3D) scenes and the IIN goal images
are renderings of HM3D. By requesting access you confirm that you have accepted the
Matterport End User License Agreement for Academic Use of HM3D and will use this
data for non-commercial research only.
extra_gated_fields:
I have accepted the HM3D / Matterport academic EULA: checkbox
---
# HM3D-M3: train split and teacher samples
HM3D-M3 is a multi-agent, multi-goal, multi-modal navigation benchmark on HM3D
(K = 2-4 agents, M = 3-5 goals per episode, goals given as object category (ON),
text description (TN) or instance image (IIN)).
- Code: https://github.com/NJU-R-L-Group-Embodied-Lab/hm3d-m3-code
- Val-1000, Mini-100 and K / M scaling episodes + IIN images: https://github.com/NJU-R-L-Group-Embodied-Lab/hm3d-m3-code/releases/tag/v1.0.0
- Project page: https://robo-lavira.github.io/hm3d-m3/
## Contents
| path | content |
| --- | --- |
| `episodes/ma_train_16000_samefloor_mixed3_seed2026_2027.json.gz` | 16,000 train episodes, 145 HM3D train scenes (disjoint from val) |
| `iin/iin_images_train16000_mixed3_stretch512.partNN-of-MM.tar.gz` | 21,326 IIN goal images (512x512) + one `manifest.json` |
| `teacher_100k/` | 100,000 observation-assignment teacher samples (safetensors + jsonl), see `SCHEMA.md` |
| `id_remap.json` | 12 renamed train14000 episode ids |
| `SHA256SUMS` | checksums of every file in this repository |
The train split is the union of two generator runs: train2000 (seed 2026) and
train14000 (seed 2027). 12 ids occurred in both runs with different content; the
train14000 copies are renamed `<scene>_<n+90000>` (`id_remap.json`), consistently
in the episodes, IIN file names/manifest and the 100K samples. Each episode has a
`source_split` field.
## Usage
```bash
huggingface-cli download skyhyding/HM3D-M3 --repo-type dataset --local-dir hm3d_m3_hf
cd hm3d_m3_hf && sha256sum -c SHA256SUMS
# place under sim-code/data/datasets/ma_objectnav_ovon/train/ of the code repository
mkdir -p $SIM/data/datasets/ma_objectnav_ovon/train
gunzip -c episodes/ma_train_16000_samefloor_mixed3_seed2026_2027.json.gz > $SIM/data/datasets/ma_objectnav_ovon/train/ma_train_16000_samefloor_mixed3_seed2026_2027.json
for f in iin/*.tar.gz; do tar -xzf "$f" -C $SIM/data/datasets/ma_objectnav_ovon/train; done
```
Duplicate (byte-identical) images are stored once and extracted as hard links.
Every archive part carries the same complete `manifest.json`.
## IIN goal images
Rendered with the Stretch camera: 512x512, 42 deg HFOV, camera height 1.31 m, 1.8 m
stand-off, looking at the target object, square crop (`stretch512`). Each IIN goal in the
episode file carries this in `iin_render_camera`. The older `iin_viewpoint` field is the
source IIN viewpoint metadata from episode generation (79 deg, 640x480, 0.88 m); it is not
the camera the released images were rendered with.
## License
Annotations, episodes and samples: CC BY-NC 4.0 (c) 2026 HM3D-M3 authors. The data refers
to HM3D scenes, which you must obtain yourself under the Matterport academic EULA; the IIN
images and map tensors are derived from HM3D and are subject to those terms as well
(non-commercial use only). Goal categories follow the OVON vocabulary.