Add ScaleNav model card and checkpoint configuration

README.md

@@ -0,0 +1,81 @@
+---
+language:
+- en
+base_model: Wan-AI/Wan2.2-TI2V-5B
+base_model_relation: finetune
+tags:
+- robotics
+- navigation
+- video-diffusion
+- image-text-to-action
+- safetensors
+---
+
+# ScaleNav
+
+**Scalable Internet Video Pretraining for Language-Conditioned Local Navigation**
+
+[Code](https://github.com/yangyihui0502/ScaleNav) · [Project page](https://yangyihui0502.github.io/scalenav-review/)
+
+ScaleNav learns navigation priors from Internet videos through action-free
+pretraining, then predicts local actions and future observations conditioned on
+RGB observations and a navigation instruction.
+
+## Download
+
+Download the checkpoint from the [GitHub release](https://github.com/yangyihui0502/ScaleNav/releases/tag/scalenav-v1).
+From the ScaleNav code repository:
+
+```bash
+python scripts/download_checkpoint.py --output checkpoints/scalenav
+```
+
+This Hub repository provides the model card, architecture configuration and
+file checksums.
+
+Download the Wan VAE, text encoder and tokenizer using the
+[installation and checkpoint guide](https://github.com/yangyihui0502/ScaleNav#checkpoints).
+
+## Inference
+
+After installing the ScaleNav code and preparing an observation:
+
+```bash
+awm check --checkpoint checkpoints/scalenav --sha256
+awm infer --checkpoint checkpoints/scalenav \
+  --base-model checkpoints/Wan2.2-TI2V-5B \
+  --image data/example/observation.png \
+  --prompt 'Approach the chair beside the table.' --output outputs/example
+```
+
+This produces a nine-step action trajectory and a generated video. Actions are
+per-frame forward/right/clockwise-yaw displacements in metres/metres/radians.
+Continuous navigation uses incoming camera frames through
+[`awm serve`](https://github.com/yangyihui0502/ScaleNav/blob/main/docs/INFERENCE.md).
+
+## 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](https://github.com/yangyihui0502/ScaleNav/blob/main/docs/TRAINING.md),
+[evaluation](https://github.com/yangyihui0502/ScaleNav/blob/main/docs/EVALUATION.md)
+and [trajectory visualization](https://github.com/yangyihui0502/ScaleNav#inference).

SHA256SUMS.json

@@ -0,0 +1,6 @@
+{
+  "ar_model.pt": "7e8d953987c3d6c24eb90da5b576e7fd8cc014ebb61819d980ffbd75524d057f",
+  "train_meta.pt": "35528281fb59deaa6ede1b18ab6c7b30aa7751e515337f14a77a7871931ffaf0",
+  "transformer/config.json": "c8061133ae831595a98104f7f4fb1d4c85d83f2745842301eca34b89967b45c6",
+  "transformer/diffusion_pytorch_model.safetensors": "d66cc8e698ac201646d832e190dac65deaa616bd1a8d3f1c844248aef155035d"
+}

transformer/config.json

@@ -0,0 +1,39 @@
+{
+  "_class_name": "CausalChunkDiTModel",
+  "_diffusers_version": "0.37.1",
+  "action_dim": 3,
+  "action_temporal_scale": 1.0,
+  "chunk_overlap_frames": 0,
+  "cross_attn_norm": true,
+  "cross_attn_type": "cross_attn",
+  "dim": 3072,
+  "eps": 1e-06,
+  "ffn_dim": 14336,
+  "freq_dim": 256,
+  "hidden_size": 3072,
+  "in_channels": 48,
+  "in_dim": 48,
+  "max_attn_chunks": 4,
+  "model_type": "ti2v",
+  "num_heads": 24,
+  "num_layers": 30,
+  "num_special_tokens": 8,
+  "out_dim": 48,
+  "patch_size": [
+    1,
+    2,
+    2
+  ],
+  "qk_norm": true,
+  "special_memory_recent_chunks": 2,
+  "text_dim": 4096,
+  "text_len": 512,
+  "traj_head_hidden_dim": 768,
+  "traj_head_type": "deep",
+  "use_continuous_history": true,
+  "use_special_memory": false,
+  "window_size": [
+    -1,
+    -1
+  ]
+}