---
license: apache-2.0
library_name: starvla
tags:
  - robotics
  - vision-language-action
  - world-model
  - robodojo
base_model:
  - Wan-AI/Wan2.2-TI2V-5B-Diffusers
  - google/umt5-xxl
---

# Ego4WAM-Joint

Released weights for the Joint variant of Ego4WAM.

## Contents

| File | Size | Description |
| --- | ---: | --- |
| `model.safetensors` | 25.07 GiB | 2,092 BF16 tensors containing the complete framework state dict |
| `inference_config.yaml` | <1 KiB | Minimal model selection: `Ego4WAM` with `interaction_mode: joint` |
| `model_assets/` | 20.46 MiB | VAE/text-encoder configs and the UMT5 tokenizer required for offline construction |
| `LICENSE` | | Apache-2.0 license for the released weights |

The checkpoint contains the Video DiT, VAE, UMT5 text encoder, Action DiT, and
proprioceptive encoder weights. No additional backbone weights are downloaded
when loading this release.


## Loading

```python
from huggingface_hub import snapshot_download

from starVLA.model.framework.base_framework import baseframework

root = snapshot_download("HorizonRobotics/Ego4WAM")
model = baseframework.from_pretrained(f"{root}/model.safetensors")
```

This constructs the framework from `inference_config.yaml`, resolves the local
files in `model_assets/`, and loads the state dict with `strict=True`.

## Model geometry

| | |
| --- | --- |
| Tensors / dtype | 2,092 / BF16 |
| Parameters | 13,456,608,092 |
| State-dict prefixes | `backbone.` 1,266; `action_model.` 824; `proprio_encoder.` 2 |
| Interaction mode | Joint synchronous video/action attention |
| State / action dimension | 32 / 32 |
| Action horizon | 32 |
| RoboDojo execution output | First 14 dimensions after inverse normalization |
| MoT | 30 layers; video hidden size 3,072; action hidden size 1,024 |
| Sampling | 20-step Euler flow matching |
| Camera input | Head, left wrist, and right wrist RGB views |

The SHA-256 digest of `model.safetensors` is:

```text
ae1b4595e1909e4573ad3214e1c5112f1920a34675340f6adfd4f2e9ec6c12c4
```

## License and attribution

The released weights are provided under Apache-2.0. Ego4WAM source code is
provided under MIT. Ego4WAM builds on StarVLA and uses Wan2.2 TI2V-5B and
UMT5-XXL components; retain the corresponding upstream attribution when
redistributing derivative work.