Pre-built datasets for hm3denv, a physically valid indoor robot-navigation simulator with a Gymnasium API. Each dataset holds 2D maps of building storeys and navigation tasks (start, goal, geodesic distance, difficulty labels) for one or more robots. Every start and goal was checked for collisions with the robot's real footprint and for connectivity; an oracle agent solves 100 % of the tasks.

The 3D source scenes (HM3D .glb files, Isaac Sim scenes) are not included.

Datasets

Folder	Environment	Source	Maps	Scenes train / val / test	Robots	Tasks	Size
`isb-svg-v1`	`HM3D/Svg-v0` (continuous)	Isaac-Scene-Builder	204	143 / 31 / 30	9	36 720	121 MB
`isb-grid-v1`	`HM3D/Grid-v0` (grid)	Isaac-Scene-Builder	202	141 / 30 / 31	`jetauto_pro`	4 040	5 MB
`svg-v1`	`HM3D/Svg-v0` (continuous)	HM3D	177	67 / 14 / 15	9	30 835	40 MB
`grid-jetauto-v1`	`HM3D/Grid-v0` (grid)	HM3D	157	67 / 14 / 15	`jetauto_pro`	3 140	4 MB
`grid-s15-v1`	`HM3D/Grid-v0` (grid, 15 cm cells)	HM3D	166	67 / 14 / 15	`s15_h63`	3 320	5 MB

The 9 robots of the continuous datasets: agilex_limo, clearpath_jackal, jetauto_pro, kobuki, pal_tiago, turtlebot3_burger, turtlebot3_waffle_pi, turtlebot4, turtlebot4_lite. Splits are by scene: a building never appears in two splits.

Usage

With a (free) Hugging Face account, request access with the form on this page and wait for the approval.
Install the simulator, log in once, and download datasets by name:

git clone https://github.com/TrKimHieu/2D_Navigation_Sim && cd 2D_Navigation_Sim
python -m venv .venv
source .venv/bin/activate        # Windows (PowerShell): .venv\Scripts\Activate.ps1
pip install ".[fast,hub]"        # the simulator + huggingface_hub (the `hf` command)
hf auth login                    # once per machine
hm3d download                    # list the datasets in this repository
hm3d download isb-svg-v1         # -> ~/.hm3denv/datasets/isb-svg-v1 (or $HM3D_HOME/datasets)

import gymnasium as gym
import hm3denv

env = gym.make("HM3D/Svg-v0", dataset="isb-svg-v1", robot="turtlebot4", split="train")
obs, info = env.reset(seed=0)

Downloaded datasets are found by name and checked against the sha256 of every file in their manifest.json. You can also download a folder with any Hugging Face tool and point HM3D_DATASETS to its parent directory.

Files

Each folder is one dataset (schema 2.0):

manifest.json            config, splits, per-robot statistics, sha256 of every file
maps/<class>/<map>.svg   continuous maps: polygons in metres (svg datasets)
grids/<map>.npz|.json    occupancy grids and their metadata (grid datasets)
robots/<robot>.json      the robot presets the tasks were built for
tasks/<robot>/<map>.json start / goal poses, geodesic and straight-line distance, labels
preview/                 PNG previews of the maps (not used by the simulator)

The full format is described in the simulator README.

License

Isaac-Scene-Builder datasets (isb-svg-v1, isb-grid-v1): built from scenes created by the author; released under CC BY 4.0.
HM3D-derived datasets (svg-v1, grid-jetauto-v1, grid-s15-v1): derived from the Habitat-Matterport 3D dataset. They are shared only with users who have been granted access to HM3D by Matterport, and their use is subject to the HM3D Terms of Use (non-commercial research).

The simulator code is MIT-licensed.

Citation

@software{hm3denv,
  author  = {Tran, Kim Hieu},
  title   = {hm3denv: physically valid indoor robot-navigation environments from HM3D},
  year    = {2026},
  url     = {https://github.com/TrKimHieu/2D_Navigation_Sim},
  version = {0.8.0}
}

If you use the HM3D-derived datasets, please also cite HM3D:

@inproceedings{ramakrishnan2021hm3d,
  title     = {Habitat-Matterport 3D Dataset ({HM3D}): 1000 Large-scale 3D Environments for Embodied {AI}},
  author    = {Santhosh Kumar Ramakrishnan and Aaron Gokaslan and Erik Wijmans and Oleksandr Maksymets
               and Alexander Clegg and John M Turner and Eric Undersander and Wojciech Galuba
               and Andrew Westbury and Angel X Chang and Manolis Savva and Yili Zhao and Dhruv Batra},
  booktitle = {Thirty-fifth Conference on Neural Information Processing Systems Datasets and Benchmarks Track},
  year      = {2021}
}

Downloads last month: 36

Total file size:

176 MB