README.md

---
license: apache-2.0
library_name: pytorch
pipeline_tag: robotics
tags:
  - robotics
  - world-action-model
  - world-model
  - video-generation
  - dexbotic
  - dw05
  - base-checkpoint
---

# DW05-Base

DW05-Base is a released DW05 base world-action model checkpoint for the Dexbotic
DW05 runtime.  It contains a trained DW05 video/action/MoT checkpoint with a
32-dimensional action and proprioception interface, packaged together with local
runtime components for offline loading.

This repository is a **DW05 runtime bundle**.  Users should point Dexbotic to
the repository root with `DW05_MODEL_BASE_PATH`; they do not need to reproduce
upstream model cache directory names.

## What Is Included

```text
DW05-Base/
  model.pt
  vae/
    model.pth
  text_encoder/
    model.pth
  tokenizer/
    tokenizer_config.json
    tokenizer.json
    spiece.model
    special_tokens_map.json
```

The base checkpoint is intended as a general DW05 model release and as a starting
point for method development, fine-tuning, and downstream policy/runtime
integration.  It is not packaged with RobotWin policy normalization statistics.
For RobotWin policy inference, use a downstream RobotWin-specific DW05 checkpoint
and its matching `norm_stats.json`.

## Intended Runtime

Use this checkpoint with the Dexbotic DW05 runtime:

```bash
git clone https://gitlab.dexmal.com/robotics/dexbotic-open.git dexbotic
cd dexbotic
pip install -e .
```

Set the bundle root:

```bash
export DW05_MODEL_BASE_PATH=/path/to/DW05-Base
export TOKENIZERS_PARALLELISM=false
```

## Checkpoint Notes

- `model.pt`: DW05 base checkpoint.  The checkpoint contains trained
  video/action/MoT weights and a proprio encoder.
- Action dimension: 32.
- Proprioception dimension: 32.
- Stored dtype metadata: `torch.bfloat16`.
- Training step recorded in the checkpoint: 140000.

## Runtime Components

- `vae/`: local VAE runtime component for image/video latent encoding and
  decoding.
- `text_encoder/`: local text encoder runtime component for prompt encoding.
- `tokenizer/`: local tokenizer files for prompt tokenization.

These are DW05-facing bundle directories.  They contain upstream-compatible
runtime components, but the user-facing package layout remains DW05-owned.

## Example Loading

```python
import torch

from dexbotic.model.dw05 import DW05ModelConfig

model_cfg = DW05ModelConfig(
    load_text_encoder=True,
    skip_dit_load_from_pretrain=True,
    action_dim=32,
    proprio_dim=32,
)

model = model_cfg.build_model(model_dtype=torch.bfloat16, device="cuda:0")
model.load_checkpoint("/path/to/DW05-Base/model.pt")
model.eval()
```

For downstream policy inference, make sure the policy preprocessing,
action/proprio dimensions, and normalization statistics match the checkpoint you
load.

## Relationship To DW05-Robotwin

DW05-Base is a general 32D base checkpoint.  DW05-Robotwin is a downstream
RobotWin-oriented release with RobotWin policy normalization assets and a
RobotWin-specific runtime configuration.  Use DW05-Robotwin for the packaged
RobotWin online demo and policy evaluation path.

## License And Attribution

This DW05-Base release is distributed under the Apache License 2.0.  See
[`LICENSE`](./LICENSE) for the full license text and [`NOTICE`](./NOTICE) for
third-party attribution.

This release is trained from and used with open third-party components,
including Wan2.2 and uMT5-compatible tokenizer/text components.  Those
components remain subject to their own upstream licenses and attribution
requirements.

In particular:

- Wan2.2 components are licensed upstream under Apache License 2.0.
- uMT5 tokenizer/text components are licensed upstream under Apache License 2.0.

Users who redistribute a modified bundle or include additional third-party files
should preserve the corresponding upstream license and attribution notices.

## Limitations

- The checkpoint is released for research and development of DW05-style
  world-action models and downstream fine-tuning.
- It is not a drop-in RobotWin policy checkpoint unless paired with compatible
  policy preprocessing and normalization statistics.
- Real-robot deployment requires independent safety validation, robustness
  evaluation, and environment-specific testing.
- Performance outside the training and downstream fine-tuning distributions has
  not been guaranteed.

## Troubleshooting

**Runtime components are not found.**

Set `DW05_MODEL_BASE_PATH` or pass `--model_base_path` to the DW05 runtime.  The
path should be the root of this DW05 bundle.

**Action dimension mismatch.**

Build the DW05 model with `action_dim=32` and `proprio_dim=32` before loading
this checkpoint.

**RobotWin policy output is misaligned.**

Use a RobotWin-specific DW05 checkpoint and matching `norm_stats.json`, or
fine-tune this base checkpoint with the intended RobotWin action/state
preprocessing.

## Citation

If you use DW05-Base, please cite or acknowledge DW05/Dexbotic and the upstream
projects listed in [`NOTICE`](./NOTICE), including Wan2.2 and uMT5 where
applicable.