• January 2026: Accepted at ICRA 2026! 🎉
• October 2025: Training and evaluation code published. Weights and dataset available in releases.

1. Create a virtual environment with conda and python 3.9.

conda create python=3.9 cmake=3.14.0 -n metricnet
conda activate metricnet

2. Install the dependencies using pip or pdm

conda install habitat-sim headless -c conda-forge -c aihabitat
conda env update -n metricnet --file environment.yaml
pip install git+ssh://[email protected]/debOliveira/diffusion_policy.git@db1434cc256b53deb0ad7228c129c0ce7c733822
pip install git+ssh://[email protected]/debOliveira/depth-anything-V2.git@7885bbc0647bc64d55ff5803561ea2c7dea1af72

1. Download and process the datasets according to NoMaD's instructions.
2. Generate the benchmark data using the provided script python generate_benchmark_data.py. For more information on the generation, please use the --help flag.
3. Generate the training data using the provided script python generate_training_data.py . For more information on the generation, please use the --help flag.
4. Download DepthAnything-V2 ViT-s weights for training MetricNet, and DepthAnything-V2 Metric ViT-B weights for MetricNav.
5. If you want to use the pretrained model, download the weights from our latest release.

To train the model, you need to adjust the in the configuration YAML metricnet.yaml as:

1. train to True
2. depth_encoder_weights to the DepthAnythingV2 checkpoint path
3. datasets/<DATASET>/data_folder, datasets/<DATASET>/train and datasets/<DATASET>/test to the folders generated during the (data processing step)[#-download-data-and-weights]

Then, run the following command:

python train.py -c <YOUR_CONFIG>.yaml

If you want to use wandb to log the training, you can set the use_wandb flag in the configuration YAML to True and the project and entity to your desired project and entity (usually your username). Don't forget to login first:

wandb login

To test the model, you need to have the model trained. Weights are available in the latest release. Adjust in the configuration YAML metricnet.yaml as:

1. train to False.
2. depth_encoder_weights to the DepthAnythingV2 checkpoint path.
3. load_run to the path of the desired weights.

Then, run the following command:

python train.py -c <YOUR_CONFIG>.yaml

To run the benchmark, you need to have the model trained and a local copy of Matterport3D. Weights are available in the latest release. Adjust in the configuration YAML benchmark.yaml as:

1. benchmark_dir to the path of the Matterport3D dataset.
2. result_dir to the desired output folder.
3. model_cfg to the path of your model configuration YAML.
4. model_ckpt to the path of the desired weights.

Then, run the following command:

python benchmark.py -c <YOUR_CONFIG>.yaml

We deployed to a TurtleBot 4 using ROS2 Humble.

• NoMaD: Goal Masking Diffusion Policies for Navigation and Exploration
• NaviDiffusor: Cost-Guided Diffusion Model for Visual Navigation
• Imperative Path Planner (iPlanner)

@misc{nayak2025metricnetrecoveringmetricscale,
      title={MetricNet: Recovering Metric Scale in Generative Navigation Policies}, 
      author={Abhijeet Nayak and Débora N. P. Oliveira and Samiran Gode and 
              Cordelia Schmid and Wolfram Burgard},
      year={2025},
      eprint={2509.13965},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2509.13965} 
  }