---
tags:
- vision-language-action
- world-model
- robodojo
---

# HWM-0-Preview-RoboDojo

HWM-0-Preview combines a vision-language subtask planner, visual and semantic
memory, and joint video/action flow matching for dual-arm manipulation.
This repository contains the complete RoboDojo inference bundle exported from
the **60,000-step EMA checkpoint**.

Run this model through the `HWM_0_preview` adapter in
[XPolicyLab](https://github.com/XPolicyLab/XPolicyLab), which provides policy
serving and RoboDojo evaluation. This repository supplies the weights and
runtime assets used by that adapter.

## Contents

| File | Size | Description |
| --- | ---: | --- |
| [model.pt](model.pt) | 31.97 GiB | Complete BF16 model state dictionary |
| [config.yaml](config.yaml) | 5.06 KiB | Architecture, conditioning, history and inference settings |
| [dataset_statistics.json](dataset_statistics.json) | 5.14 KiB | Required state/action normalization statistics |
| [processor/](processor/) | 19.08 MiB | VLM processor, tokenizer and chat template |
| [tokenizer/](tokenizer/) | 16.06 MiB | T5 tokenizer assets |
| [manifest.json](manifest.json) | 1.11 KiB | Export checksums and configuration fingerprint |

```text
HWM-0-Preview-RoboDojo/
├── README.md
├── .gitattributes
├── model.pt
├── config.yaml
├── dataset_statistics.json
├── manifest.json
├── processor/
└── tokenizer/
```

The checkpoint includes the VLM, T5 encoder, VAE, Video DiT, Action DiT,
state projection and learned queries. The matching adapter constructs these
components from the bundled configuration without downloading additional
backbone weights. Both tokenizers load from local JSON assets.

## Usage with XPolicyLab

Follow the official [XPolicyLab guide](https://github.com/XPolicyLab/XPolicyLab)
for setup, checkpoint conventions and deployment, and the
`policy/HWM_0_preview/README.md` in the matching source release for this model's
installation requirements. Use a checkout that includes the `HWM_0_preview`
adapter. Keep the policy environment separate from the RoboDojo simulator/client
environment, as recommended by XPolicyLab.

### Download the bundle

From the XPolicyLab checkout, enter the adapter directory. With the Hugging Face
CLI available, download the complete bundle into `HWM-0-Preview-RoboDojo/`:

```bash
cd policy/HWM_0_preview
hf download "HorizonRobotics/HWM-0-Preview-RoboDojo" \
  --local-dir HWM-0-Preview-RoboDojo
```

The adapter also provides `download_checkpoint.sh`, which defaults to
current repo:

```bash
bash download_checkpoint.sh
```

Both commands use `HWM-0-Preview-RoboDojo/` inside the adapter directory.
Keep the entire bundle together, including `config.yaml`, normalization
statistics, `processor/` and `tokenizer/`.

### Evaluate through the standard entrypoint

After installing the policy and evaluation environments according to their
respective guides, run from the same adapter directory:

```bash
# Replace these with your installed environment names and available GPU IDs.
POLICY_ENV=hwm_policy
EVAL_ENV=robodojo
POLICY_GPU=0
ENV_GPU=1

EVAL_ENV_TYPE=sim bash eval.sh \
  RoboDojo stack_bowls ./HWM-0-Preview-RoboDojo arx_x5 joint 0 \
  "$POLICY_GPU" "$ENV_GPU" "$POLICY_ENV" "$EVAL_ENV"
```

`eval.sh` starts the policy server, waits for it to become ready, runs the
environment client and cleans up the server. The adapter resolves the checkpoint
and loads the model and its assets. Change `stack_bowls` to the desired task.
The `ckpt_name` argument is the explicit path `./HWM-0-Preview-RoboDojo`,
so different evaluation seeds use the same downloaded bundle.

Use `EVAL_ENV_TYPE=debug` for an offline interface check; it does not measure task
success. For simulation, follow the official RoboDojo setup and XPolicyLab
workspace layout. For split-machine deployment, follow XPolicyLab's
**Deployment Flow** using `setup_eval_policy_server.sh` on the policy machine and
`setup_eval_env_client.sh` on the simulator machine. Refer to those official
guides for argument details and environment setup.

The model requires approximately 32 GiB for weights alone, plus memory for
activations and caches. Select a policy GPU with sufficient headroom; prior
integration validation used H20 hardware. Model-specific inference settings are
configured through the adapter's `deploy.yml`: `num_inference_steps`, `flow_shift`
and `replan_steps` inherit bundle defaults of 10, 5.0 and 30 when set to `null`.
`inference_batch_size` defaults to 4; reduce it for lower GPU memory use.

## Model geometry

| Item | Value |
| --- | --- |
| Checkpoint | 60,000-step EMA |
| Tensors / dtype | 3,078 / BF16 |
| Stored tensor elements | 17,162,844,810 |
| Robot | ARX X5 dual arm |
| State / action dimension | 14 / 14 |
| Action horizon | 48 absolute joint-position targets |
| Execution horizon | 30 control steps per replan |
| Sampling | 10 flow-matching steps; flow shift 5.0 |
| Camera input | Head, left-wrist and right-wrist RGB |
| Visual history | Up to 20 frames, sampled every 30 control steps |
| Normalization | Bundled mean/std statistics, key `new_embodiment` |

## Integrity

`model.pt` is **34,326,694,240 bytes**. Its SHA256 digest is:

```text
40d998b67d27824fab5c6b64b4f20c2926d299dac9393dadb07a8c831a226988
```

The remaining original bundle checksums are recorded in
[manifest.json](manifest.json). Exported weights, configuration, normalization
statistics and tokenizer assets are preserved byte for byte.

This release contains inference weights and assets. Training state and
benchmark results are not included; no complete benchmark score is claimed here.