README.md

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

# DW05-Robotwin

DW05-Robotwin is a released DW05 world-action model checkpoint for
RobotWin-style robot policy inference and video rollout.  It predicts future
robot actions and can generate future robot-view video through the Dexbotic DW05
runtime.

The repository is packaged as a **DW05 runtime bundle**.  Users should point the
runtime to this repository root; they do not need to reproduce upstream model
cache directory names.

## What Is Included

The core checkpoint is:

```text
model.pt
```

For RobotWin policy inference, the runtime also needs normalization statistics.
The recommended release layout is:

```text
DW05-Robotwin/
  model.pt
  norm_stats.json
```

If `norm_stats.json` is not included in a particular release snapshot, provide
the matching RobotWin normalization statistics explicitly through the runtime
`--stats` / `--norm-stats-path` argument.

This release is organized as a full offline runtime bundle:

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

`vae/`, `text_encoder/`, and `tokenizer/` are DW05-facing bundle directories.
They contain upstream-compatible runtime components, but the user-facing package
layout remains DW05-owned.

## 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-Robotwin
export TOKENIZERS_PARALLELISM=false
```

If `norm_stats.json` is not placed at the bundle root, pass its path explicitly
with `--stats`, `--norm-stats-path`, or the corresponding Dexbotic config field.

## Online Demo

Run the RobotWin online demo from the Dexbotic repository:

```bash
python playground/online_demos/robotwin_online_demo.py --web \
  --ckpt /path/to/DW05-Robotwin/model.pt \
  --stats /path/to/DW05-Robotwin/norm_stats.json \
  --model_base_path /path/to/DW05-Robotwin \
  --device cuda:0 \
  --num_inference_steps 5
```

The demo exposes the original interactive RobotWin joint-condition UI and uses
the shared `DW05RobotWinPolicy` runtime.

## Programmatic Policy Loading

```python
from dexbotic.policy.dw05_policy import DW05RobotWinPolicy, DW05RobotWinPolicyConfig

policy = DW05RobotWinPolicy(
    DW05RobotWinPolicyConfig(
        checkpoint_path="/path/to/DW05-Robotwin/model.pt",
        norm_stats_path="/path/to/DW05-Robotwin/norm_stats.json",
        model_base_path="/path/to/DW05-Robotwin",
        device="cuda:0",
        mixed_precision="bf16",
        num_inference_steps=5,
    )
)
```

The policy expects RobotWin-style observations with RGB camera images, robot
state, and a natural-language instruction.  See the Dexbotic DW05 README for the
complete runtime, evaluation, and deployment examples.

## File Notes

- `model.pt`: DW05 trained checkpoint.  It contains the DW05 world-action model
  parameters used by the released policy, including trained video/action/MoT
  weights and the proprio encoder.
- `norm_stats.json`: action/state normalization statistics used by RobotWin
  policy inference.  This is required for action normalization and
  denormalization.
- `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.

## License And Attribution

This DW05-Robotwin 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, uMT5-compatible tokenizer/text components, and
RoboTwin/RobotWin-style data and evaluation.  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.
- RoboTwin code and public dataset metadata were observed under MIT License.

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 world-action
  models, robot policy inference, and video rollout experiments.
- Real-robot deployment requires independent safety validation, robustness
  evaluation, and environment-specific testing.
- The model expects preprocessing compatible with the Dexbotic DW05 RobotWin
  runtime, including image composition, state/action normalization, and prompt
  formatting.
- Performance outside RobotWin-style observations and task distributions has not
  been guaranteed.

## Troubleshooting

**Model 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.

**Norm stats are missing.**

Place `norm_stats.json` at the bundle root or pass its path explicitly with
`--stats` / `--norm-stats-path`.

**The online demo starts but generation looks misaligned.**

Check that the runtime uses the Dexbotic DW05 preprocessing path: RobotWin image
composition, DW05 normalization statistics, and the matching checkpoint should be
used together.

## Citation

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