SimpleNav

Make Navigation VLA Simple.
A simple, unified, reproducible, and extensible framework for navigation VLA research.

中文 · Project Page · Data Pipeline · Documentation · Results · Data, Environments & Models

SimpleNav is a simple, unified, reproducible, and extensible framework for navigation VLA research, jointly developed and open-sourced by THUNLP at Tsinghua University, AI9Stars, OpenBMB, and HITDIP. It provides a unified research pipeline that connects heterogeneous navigation data, long-horizon VLA models, training, and benchmark evaluation through well-defined interfaces. SimpleNav supports both aerial and ground navigation, preserves dataset-specific coordinate systems and simulator semantics through dedicated adapters, and standardizes model, action, artifact, and evaluation interfaces to enable efficient reuse, comparison, and extension across datasets, tasks, and platforms.

Vision

Navigation research should not require a separate data-model-evaluation stack for every dataset. SimpleNav provides one research loop in which:

• source datasets enter through explicit conversion adapters;
• model components remain replaceable and composable;
• training runs are defined by portable configs;
• benchmark-specific behavior stays inside evaluation plugins;
• results remain traceable to code, data, config, checkpoint, and simulator versions.

Why SimpleNav

Framework

Data Protocol

The primary LeRobot dataloader keeps storage, model input, and prediction target separate:

Benchmark adapters may declare a different action Protocol when required by the benchmark. The config and adapter Protocol are authoritative. See Data Structure and State/Action Protocol.

Trajectory augmentation

Each animation aligns one raw trajectory with its enhanced version. Click an animation to open the MP4.

AerialVLN · Example 1

OpenFly · Episode 000008

Model

SimpleNav combines a vision-language backbone, selected long history, temporal-view context, and a continuous action head. The model consumes the protocol above and keeps dataset-specific coordinate semantics in the adapter.

Results

We adopt Qwen3.5-VL 4B as the unified vision-language backbone, and complete model training and closed-loop evaluation on 6 benchmarks respectively. Except for the necessary adaptation of data and task interfaces, we do not perform task-specific performance optimization for any individual benchmark. The results are summarized as follows. Full comparison tables and protocol notes are in Release 01 Benchmarks.

Demos

Selected rollout trajectory previews are shown below. See the project-page video gallery for full videos.

OpenFly · Env 16	TravelUAV · Modern City
AerialVLN · Env 8	RxR-CE · Episode 10129
EVT-Bench · Scene 2	EVT-Bench · Scene 30

Risks and Limitations

• SimpleNav is a research framework, not a safety-certified flight-control system. Do not use model outputs as the sole control authority.
• Validate in simulation and controlled environments with qualified supervision, manual override, emergency stop, geofencing, and independent safety monitors.
• Performance can degrade under distribution shift, perception or communication latency, actuator/simulator mismatch, and coordinate or action-protocol errors.
• No guarantee is made for collision avoidance, fail-safe behavior, or regulatory compliance; operators remain responsible for deployment decisions.

Quick Start

Public resources:

• Data, environments, and models

Place downloaded packages in the repository-relative local/ layout below.

1. Clone and install the model environment

Requirements: Linux, Python 3.10, and a model-compatible NVIDIA driver. Dataset conversion also requires ffmpeg; closed-loop evaluation requires the corresponding simulator and scene assets.

git clone -b SimpleNav https://github.com/OpenBMB/SimpleNav.git SimpleNav
cd SimpleNav
curl -LsSf https://astral.sh/uv/install.sh | sh
uv python install 3.10
uv sync --frozen --no-dev
uv run --no-sync python -c "import torch, transformers; print(torch.__version__, transformers.__version__)"

For the Qwen3.5 reference recipe, install the optional CUDA extension after the base environment succeeds:

uv sync --frozen --no-dev --extra flash-attention

See Installation for the Conda data-tool environments and system packages.

2. Prepare data

Install only the data component you need. For raw-dataset conversion:

cd data_pipeline/dataset_conversion
conda env create -f environment.yml
conda activate vln-dataset-conversion
vln-convert --help

The other component entry points are:

data_pipeline/trajectory_augmentation  -> vln-augment
data_pipeline/image_collection         -> vln-collect

Follow Data Preparation, then place local resources under:

local/
├── models/                         # base VLMs and auxiliary models
├── data/                           # converted datasets and benchmark inputs
├── checkpoints/                    # SimpleNav checkpoints + dataset_statistics.json
├── simulators/                     # AirSim/Habitat runtimes and scene assets
├── eval_results/
└── results/

Validate a converted split:

uv run --no-sync python -m tool.navvla.cli.validate_dataset \
  local/data/<dataset>/<split> --visual-token-mode online_images --smoke-load 8

3. Train

The public reference recipe is OpenFly Qwen3.5-VL training:

bash examples/NavVLA/train_files/qwen35/run_train.sh \
  examples/NavVLA/train_files/qwen35/navvla_qwen35_cpm_openfly_portable.yaml \
  --dry-run

bash examples/NavVLA/train_files/qwen35/run_train.sh \
  examples/NavVLA/train_files/qwen35/navvla_qwen35_cpm_openfly_portable.yaml

Copy the portable config before changing data mixtures, GPU counts, or attention implementations. See Training.

4. Evaluate

Each public config resolves paths relative to its own directory.

Inspect a two-episode plan before starting a simulator:

bash NavVLAeval/openfly/run_eval.sh --dry-run \
  --override benchmark.max_samples=2 \
  --override parallel.gpu_ids='[0]' \
  --override output.run_name=openfly_dry_run

See Evaluation for resource layout, execution, resume, and artifacts.

For the OpenFly, AerialVLN, and TravelUAV data-to-training-to-evaluation workflow, use Aerial Training and Evaluation.

For the released R2R-CE and RxR-CE Qwen3.5 workflow, use VLN-CE Training and Evaluation.

Documentation

Roadmap

• Maintain released checkpoints, model cards, converted-data manifests, dataset cards, and simulator packages.
• Add portable multi-domain training recipes and extend the released single-dataset workflows.
• Expand model backbones, history and memory modules, action heads, and platform adapters.
• Publish reproducible result bundles with resolved configs and episode-level artifacts.
• Connect evaluation failures to data generation and the next training iteration.

Citation

If SimpleNav is useful in your work, please cite the repository.

@software{simplenav,
  title = {SimpleNav: Make Navigation VLA Simple},
  author = {{SimpleNav Contributors}},
  year = {2026},
  url = {https://github.com/OpenBMB/SimpleNav},
}

License

Repository source code is released under the MIT License. Datasets, pretrained models, simulators, scene assets, and third-party components retain their own licenses.

Acknowledgements

We thank pioneering navigation VLA studies, including NavFoM, Qwen-RobotNav, ABot-N0, starVLA and InternVLA-N1, whose valuable explorations have helped shape and advance this field.

SimpleNav builds upon starVLA, Qwen-VL, LeRobot, PyTorch, Transformers, DeepSpeed, AirSim, Habitat, as well as the datasets and benchmarks described above. We gratefully acknowledge these open-source contributions and encourage users to cite the original projects and datasets employed in their experiments.