license: cc-by-4.0
task_categories:
- robotics
language:
- en
size_categories:
- 10B<n<100B
EmbodiedSplat Dataset & Checkpoints
High-fidelity 3D scene reconstructions, navigation episodes, and policy checkpoints
released with the EmbodiedSplat: Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device (ICCV 2025) paper.
This dataset supports research on sim-to-real robot navigation using 3D Gaussian Splatting (DN-Splatter) and low-effort iPhone LiDAR captures.
Repository Structure
embodied-splat/
├── ckpts/
├── datasets/
│ └── pointnav/
│ ├── pointnav_dn_splatter/
│ ├── pointnav_dn_splatter_2/
│ ├── pointnav_hm3d_stretch/
│ ├── pointnav_hssd/
│ ├── pointnav_mushroom_dn_splatter/
│ ├── pointnav_polycam_mesh/
│ └── pointnav_polycam_mesh_2/
├── fine_tune_ckpts/
│ ├── hm3d_fine_tuned_ckpts/
│ └── hssd_fine_tuned_ckpts/
├── grad_lounge/
└── scene_datasets/
├── mushroom/
│ └── dn_splatter/
└── polycam_data/
├── dn_splatter/
├── dn_splatter_2/
├── polycam_mesh/
└── polycam_mesh_2/
ckpts/
Pre-trained PointNav policy checkpoints:
- hm3d_ckpt_204.pth – trained on the HM3D dataset
- hssd_ckpt_332.pth – trained on the HSSD dataset
Use these as starting points for fine-tuning on custom scenes.
datasets/pointnav/
Habitat-Sim pointNav episode datasets used in the paper:
- HM3D episodes
- HSSD episodes
- MuSHRoom dataset episodes
- Episodes for our custom DN-Splatter and Polycam meshes
castleberry→ conf_b in the paperclough_classroom→ classroom in the papergrad_lounge→ lounge in the paperpiedmont→ conf_a in the paperpolycam_mesh_2/anddn_splatter_2/→ coda_conference_room (conf_c in the supplementary material)
These provide navigation tasks (start/goal poses, goals, etc.) for training and evaluation.
fine_tune_ckpts/
Fine-tuned policy checkpoints corresponding exactly to the evaluation results reported in the paper:
hm3d_fine_tuned_ckpts/
- Policies fine-tuned from the HM3D pre-trained model on each of the 5 scenes released in the paper + supplementary material.
hssd_fine_tuned_ckpts/
- Policies fine-tuned from the HSSD pre-trained model on each of the 4 scenes released in the paper.
These checkpoints directly reproduce the quantitative results shown in the main and supplementary text.
grad_lounge/
Checkpoints used for real-world robot navigation in the “lounge” scene
described in the paper’s real-robot experiments.
scene_datasets/
3D scene reconstructions and meshes.
- mushroom/ – DN-Splatter reconstructions of the MuSHRoom dataset.
- polycam_data/ – Our own Polycam captures and corresponding DN-Splatter reconstructions:
castleberry→ conf_b in the paperclough_classroom→ classroom in the papergrad_lounge→ lounge in the paperpiedmont→ conf_a in the paperpolycam_mesh_2/anddn_splatter_2/→ coda_conference_room (conf_c in the supplementary material)
Each subfolder contains:
- Polycam exported meshes (
.glb) - DN-Splatter reconstructed meshes (
.glb)
Usage
These resources can be used to:
- Train navigation agents in Habitat-Sim or similar simulators.
- Reproduce all sim-to-real experiments from the paper.
- Fine-tune your own navigation policies on high-fidelity reconstructions.
Please cite our paper if you use this dataset or checkpoints:
@inproceedings{chhablani2025embodiedsplat,
title={EmbodiedSplat: Personalized Real-to-Sim-to-Real Navigation with Gaussian Splats from a Mobile Device},
author={Gunjan Chhablani and Xiaomeng Ye and Muhammad Zubair Irshad and Zsolt Kira},
booktitle={Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV)},
year={2025}
}
License
All files are released under the CC-BY-4.0 license.
You are free to use and modify the data with appropriate attribution.
Contact
For questions or collaboration inquiries, please reach out through the project page:
➡️ https://gchhablani.github.io/embodied-splat/