---
license: other
license_name: prl-1.0
license_link: LICENSE
language:
- en
tags:
- robotics
- navigation
- supernav
- arxiv:2610.12126
---

# SuperNav Learned Executor

Learned local-navigation checkpoint for [SuperNav: An Agentic Navigation System for Any Task in Any Scene](https://arxiv.org/abs/2610.12126).

- **Code:** [zju3dv/SuperNav](https://github.com/zju3dv/SuperNav).
- **Checkpoint:** `ckpt_latest125.pt` (249,157,807 bytes).
- **Role:** the learned local executor, served through SuperNav's `nomad` backend. The high-level agent, simulator, scene assets, and task manifests are separate requirements.
- **Format:** PyTorch checkpoint containing `state_dict` and `train_config`. The SuperNav loader reads the checkpoint's architecture settings.

## Architecture

The checkpoint metadata selects a fine-tuned DINOv2 ViT-S/14 encoder (`dino-s-finetune`) with 224-pixel input images, GCA context (`context_size=4`, `memory_budget=12`), and a flow-matching action head with four inference steps. It predicts an eight-waypoint action horizon with waypoint scale 0.25 meters. The SuperNav loader reads these settings from the checkpoint.

## Download

Install the Hugging Face CLI with `python -m pip install -U huggingface_hub`, then run:

```bash
hf download the0xka1/SuperNav-Learned-Executor \
  ckpt_latest125.pt CHECKSUMS.sha256 --local-dir ./checkpoints
(cd checkpoints && sha256sum -c CHECKSUMS.sha256)
```

The checkpoint's SHA256 is:

```text
efd3f50def5f65d0aa57da1d2bbe1b159cf7b14e870d82495d4b4167c0aa62dd
```

## Start the policy service

Use Python 3.12 and install compatible PyTorch and torchvision builds for your device. For CUDA 12.1, the following example uses [official PyTorch wheels](https://pytorch.org/get-started/previous-versions/#v251):

```bash
python -m pip install torch==2.5.1 torchvision==0.20.1 \
  --index-url https://download.pytorch.org/whl/cu121
```

Install the published SuperNav code and its runtime dependencies in the policy-service environment:

```bash
git clone https://github.com/zju3dv/SuperNav.git
cd SuperNav
python -m pip install -e '.[agents,evaluation]'

export TORCH_HOME=/absolute/path/to/torch-cache
export NAV_LOCALNAV_CKPT=/absolute/path/to/checkpoints/ckpt_latest125.pt
python -m supernav.methods.localnav.server \
  --backend nomad --checkpoint "$NAV_LOCALNAV_CKPT" \
  --host 127.0.0.1 --port 18914 --device cuda
```

Set `TORCH_HOME` to a writable cache directory. The DINOv2 loader uses Torch Hub to obtain the [upstream repository](https://github.com/facebookresearch/dinov2) and [initial ViT-S/14 backbone weights](https://dl.fbaipublicfiles.com/dinov2/dinov2_vits14/dinov2_vits14_pretrain.pth) (about 84.2 MiB) before loading this checkpoint. The first startup requires access to GitHub and `dl.fbaipublicfiles.com`, or an already populated Torch Hub cache. xFormers is optional.

For CPU inference, use `--device cpu`. In another terminal, check:

```bash
curl --fail http://127.0.0.1:18914/healthz
```

The response should report `backend: nomad` and the downloaded checkpoint path. Keep the service running for the experiment. If using an HTTP proxy, include `127.0.0.1,localhost` in `NO_PROXY` and `no_proxy` so local policy and bridge requests bypass it.

## Run a navigation task

Follow the published [Habitat setup guide](https://github.com/zju3dv/SuperNav/blob/main/docs/release/en/habitat.md) to prepare Habitat-GS, scene assets/NavMeshes, the agent, and an external task manifest. In the agent environment:

```bash
export SUPERNAV_HABITAT_PYTHON=/absolute/path/to/habitat-env/bin/python
export SUPERNAV_SCENE_DATASET_CONFIG=/absolute/path/to/scene_dataset_config.json
export NAV_LOCALNAV_URL=http://127.0.0.1:18914

supernav run --experiment habitat-learned-executor \
  --instructions /absolute/path/to/tasks.json \
  --task-ids example-task --arms default \
  --model gpt-6-astra --output-dir data/runs/learned-executor
```

Replace the paths and task ID with your own assets and manifest. If the Habitat bridge runs on a different host, set `NAV_LOCALNAV_URL` to an address it can reach.

## License

This checkpoint is distributed under the [Project Registration License (PRL) v1.0](LICENSE). The DINOv2 backbone retains its Apache-2.0 terms. External dependencies, simulator assets, and datasets retain their respective licenses; see [third-party notices](THIRD_PARTY_NOTICES.md).