ICCV 2025
Xiao Chen  Tai Wang  Quanyi Li  Tao Huang  Jiangmiao Pang  Tianfan Xue 
The Chinese University of Hong Kong Shanghai AI Laboratory

1. About
2. Dataset
3. Installation
4. Training & Evaluation
5. Citation
6. License

Generalizable active mapping in complex unknown environments remains a critical challenge for mobile robots. Existing methods, constrained by insufficient training data and conservative exploration strategies, exhibit limited generalizability across scenes with diverse layouts and complex connectivity. To enable scalable training and reliable evaluation, we introduce GLEAM-Bench, the first large-scale benchmark designed for generalizable active mapping with 1,152 diverse 3D scenes from synthetic and real-scan datasets. Building upon this foundation, we propose GLEAM, a unified generalizable exploration policy for active mapping. Its superior generalizability comes mainly from our semantic representations, long-term navigable goals, and randomized strategies. It significantly outperforms state-of-the-art methods, achieving 66.50% coverage (+9.49%) with efficient trajectories and improved mapping accuracy on 128 unseen complex scenes.

GLEAM-Bench includes 1,152 diverse 3D scenes from synthetic and real-scan datasets for benchmarking generalizable active mapping policies. These curated scene meshes are characterized by near-watertight geometry, diverse floorplan (≥10 types), and complex interconnectivity. We unify these multi-source datasets through filtering, geometric repair, and task-oriented preprocessing. Please refer to the guide for more details and scrips.

We provide all the preprocessed data used in our work, including mesh files (in obj folder), ground-truth surface points (in gt folder) and asset indexing files (in urdf folder). We recommend users fill out the form to access the download link [HERE]. The directory structure should be as follows.

GLEAM
├── README.md
├── gleam
│   ├── train
│   ├── test
│   ├── ...
├── data_gleam
│   ├── README.md
│   ├── train_stage1_512
│   │   ├── gt
│   │   ├── obj
│   │   ├── urdf
│   ├── train_stage2_512
│   │   ├── gt
│   │   ├── obj
│   │   ├── urdf
│   ├── eval_128
│   │   ├── gt
│   │   ├── obj
│   │   ├── urdf
├── ...

We test our code under the following environment:

• NVIDIA RTX 3090/4090 (24GB VRAM)
• NVIDIA Driver: 545.29.02
• Ubuntu 20.04
• CUDA 11.8
• Python 3.8.12
• PyTorch 2.0.0+cu118

1. Clone this repository.

git clone https://github.com/zjwzcx/GLEAM
cd GLEAM

2. Create an environment and install PyTorch.

conda create -n gleam python=3.8 -y
conda activate gleam
pip install torch==2.0.0+cu118 torchvision==0.15.1+cu118 torchaudio==2.0.1+cu118 --extra-index-url https://download.pytorch.org/whl/cu118

3. NVIDIA Isaac Gym Installation: https://developer.nvidia.com/isaac-gym/download

cd isaacgym/python
pip install -e .

4. Install GLEAM.

pip install -r requirements.txt
pip install -e .

Weights & Bias (wandb) is highly recommended for analyzing the training logs. If you want to use wandb in our codebase, please paste your wandb API key into wandb_utils/wandb_api_key_file.txt. If you don't want to use wandb, please add --stop_wandb into the following command.

We provide the standard checkpoints of GLEAM [HERE]. Please use the 40k-step checkpoint as the standard. We also provide the Stage 2 checkpoints excluding the 96 Gibson scenes, as this exclusion made the model more robust and stable overall.

Please run the following command to reproduce the standard two-stage training of GLEAM.

Stage 1 with 512 scenes:

python gleam/train/train_gleam_stage1.py --sim_device=cuda:0 --num_envs=32 --headless

Stage 2 with additional 512 scenes, continually trained based on the pretrained checkpoint (specified by --ckpt_path) from stage 1. Take our released checkpoint as example, ckpt_path should be runs/train_gleam_stage1/models/rl_model_40000000_steps.zip.

python gleam/train/train_gleam_stage2.py --sim_device=cuda:0 --num_envs=32 --headless --ckpt_path=${YOUR_CKPT_PATH}$

If you want to customize a novel training environment, you need to create your environment and configuration files in gleam/env and then define the task in gleam/__init__.py.

Please run the following command to evaluate the generalization performance of GLEAM on 128 unseen scenes from the test set of GLEAM-Bench. The users should specify the checkpoint via --ckpt_path.

python gleam/test/test_gleam_gleambench.py --sim_device=cuda:0 --num_envs=32 --headless --stop_wandb --ckpt_path=${YOUR_CKPT_PATH}$

• Release GLEAM-Bench (dataset) and the arXiv paper in May 2025.
• Release the training code in May 2025.
• Release the evaluation code in June 2025.
• Release the key scripts in June 2025.
• Release the pretrained checkpoint in June 2025.

Q: Is it normal that the program gets stuck for about 5-60 minutes during training and testing?
A: This is normal because the simulation environment needs to load 1024 complex 3D scenes (for training) or 128 complex 3D scenes (for evaluation) at once, which is very time-consuming. For initial use, it is recommended to modify the hardcoded parameters (number of training scenes for stage1 and number of evaluation scenes) to reduce the number of loaded scenes for a quick run-through.

Q: Is it normal that the 3D scenes in the visualization UI have no textures?
A: This is normal. Textures have been removed from the preprocessed data to speed up simulation and rendering, as RGB/texture information is not required for geometry-level exploration. It's a trade-off to accelerate training, allowing focus on 3D spatial exploration. If you need scenes with textures, we recommend downloading the raw version of GLEAM-Bench. Please refer to the guide for more details.

If you find our work helpful, please cite it:

@article{chen2025gleam,
  title={GLEAM: Learning Generalizable Exploration Policy for Active Mapping in Complex 3D Indoor Scenes},
  author={Chen, Xiao and Wang, Tai and Li, Quanyi and Huang, Tao and Pang, Jiangmiao and Xue, Tianfan},
  journal={arXiv preprint arXiv:2505.20294},
  year={2025}
}

If you use our codebase, dataset, and benchmark, please kindly cite the original datasets involved in our work. BibTex entries are provided below.

@article{ai2thor,
  author={Eric Kolve and Roozbeh Mottaghi and Winson Han and
          Eli VanderBilt and Luca Weihs and Alvaro Herrasti and
          Daniel Gordon and Yuke Zhu and Abhinav Gupta and
          Ali Farhadi},
  title={{AI2-THOR: An Interactive 3D Environment for Visual AI}},
  journal={arXiv},
  year={2017}
}

@inproceedings{chen2024gennbv,
  title={GenNBV: Generalizable Next-Best-View Policy for Active 3D Reconstruction},
  author={Chen, Xiao and Li, Quanyi and Wang, Tai and Xue, Tianfan and Pang, Jiangmiao},
  year={2024}
  booktitle={IEEE Conference on Computer Vision and Pattern Recognition (CVPR)},
}

@inproceedings{rudin2022learning,
  title={Learning to walk in minutes using massively parallel deep reinforcement learning},
  author={Rudin, Nikita and Hoeller, David and Reist, Philipp and Hutter, Marco},
  booktitle={Conference on Robot Learning},
  pages={91--100},
  year={2022},
  organization={PMLR}
}

@inproceedings{procthor,
  author={Matt Deitke and Eli VanderBilt and Alvaro Herrasti and
          Luca Weihs and Jordi Salvador and Kiana Ehsani and
          Winson Han and Eric Kolve and Ali Farhadi and
          Aniruddha Kembhavi and Roozbeh Mottaghi},
  title={{ProcTHOR: Large-Scale Embodied AI Using Procedural Generation}},
  booktitle={NeurIPS},
  year={2022},
  note={Outstanding Paper Award}
}

@inproceedings{xiazamirhe2018gibsonenv,
  title={Gibson Env: real-world perception for embodied agents},
  author={Xia, Fei and R. Zamir, Amir and He, Zhi-Yang and Sax, Alexander and Malik, Jitendra and Savarese, Silvio},
  booktitle={Computer Vision and Pattern Recognition (CVPR), 2018 IEEE Conference on},
  year={2018},
  organization={IEEE}
}

@article{khanna2023hssd,
    author={Khanna*, Mukul and Mao*, Yongsen and Jiang, Hanxiao and Haresh, Sanjay and Shacklett, Brennan and Batra, Dhruv and Clegg, Alexander and Undersander, Eric and Chang, Angel X. and Savva, Manolis},
    title={{Habitat Synthetic Scenes Dataset (HSSD-200): An Analysis of 3D Scene Scale and Realism Tradeoffs for ObjectGoal Navigation}},
    journal={arXiv preprint},
    year={2023},
    eprint={2306.11290},
    archivePrefix={arXiv},
    primaryClass={cs.CV}
}

@article{Matterport3D,
  title={Matterport3D: Learning from RGB-D Data in Indoor Environments},
  author={Chang, Angel and Dai, Angela and Funkhouser, Thomas and Halber, Maciej and Niessner, Matthias and Savva, Manolis and Song, Shuran and Zeng, Andy and Zhang, Yinda},
  journal={International Conference on 3D Vision (3DV)},
  year={2017}
}

This work is under the Creative Commons Attribution-NonCommercial-ShareAlike 4.0 International License.