ScaleNav

Scalable Internet Video Pretraining for Language-Conditioned Local Navigation

Paper · Code · Data · Project page · Collection

ScaleNav learns navigation priors through action-free pretraining on Internet videos, then turns an RGB observation and a language instruction into robot actions. It navigates to a target visible in the initial view and stops at the instructed location.

5B parameters · 10.02 GB checkpoint · CC BY-NC 4.0

Watch the full demo in the video preview, or visit the project page.

Download

Complete the installation, then run from the ScaleNav code repository root:

python scripts/download_checkpoint.py --output checkpoints/scalenav

The downloader resumes interrupted transfers and verifies the pinned checkpoint checksums. No HF login is required for this checkpoint.

Also download the Wan VAE, text encoder and tokenizer (14.20 GB) following the checkpoint setup. For inference, use a BF16-capable NVIDIA GPU; 24 GB GPU memory and 64 GB system RAM are recommended.

Download with the Hugging Face CLI

hf download Yangyihui/ScaleNav --revision 8808b3589142b52e3bf0719aa0ecff8f86715a4a --local-dir checkpoints/scalenav \
  --include 'transformer/*' ar_model.pt train_meta.pt SHA256SUMS.json

Inference

Use ScaleNav's awm CLI with your own image and an instruction describing a target visible in it:

awm check --checkpoint checkpoints/scalenav --sha256
awm infer --checkpoint checkpoints/scalenav \
  --base-model checkpoints/Wan2.2-TI2V-5B \
  --image /path/to/observation.png \
  --prompt "Walk to the chair visible in the initial view." \
  --output outputs/example

Replace the image path and instruction with a matching pair. This saves nine per-frame action deltas in observation_trajectory.npy and a visualization in observation_trajectory.png. Actions are forward/right/clockwise-yaw displacements in metres/metres/radians. Add --decode-video to save the nine predicted RGB frames.

For continuous navigation, use awm serve. See the inference guide for paired examples, outputs and the robot adapter.

Model settings

Setting	Value
Observation size	288 × 512
Observation/action rate	4 Hz
Chunk length	9 frames
Sampling steps	4
Video / action guidance	5 / 1
Video / action scheduler shifts	5 / 1
History	Initial-frame anchor and two recent chunks
Action normalization	Symlog, c = 0.01

Checkpoint contents

File	Contents
transformer/diffusion_pytorch_model.safetensors	ScaleNav model weights
transformer/config.json	Model architecture
train_meta.pt, ar_model.pt	Deployment metadata
SHA256SUMS.json	File checksums

Total checkpoint size: 10.02 GB. The separate Wan components occupy about 14.2 GB. The code repository includes training, evaluation and trajectory visualization.

The released checkpoint is the final joint-fine-tuning checkpoint at step 6500. It contains model weights and deployment metadata. Optimizer, scheduler and RNG state for exact training resumption are not included. The Wan VAE, text encoder and tokenizer are downloaded separately.

License

The ScaleNav fine-tuned checkpoint is released under CC BY-NC 4.0 for noncommercial use. When sharing it or adaptations, provide attribution, link to the license, and indicate any changes.

ScaleNav modifies Wan2.2-TI2V-5B through navigation pretraining, action alignment and joint fine-tuning. The CC BY-NC 4.0 terms apply to the ScaleNav contributions; upstream Wan materials retain their Apache-2.0 license. The code is separately licensed under Apache-2.0, and dataset components retain their respective terms.