README.md

---
license: gemma
library_name: transformers
base_model:
- Dexmal/DM05
datasets:
- RoboChallenge/Table30v2
tags:
- robotics
- vision-language-action
- dm05
- opendm
- robochallenge
- arx5
---

# DM05-ARX5 (RoboChallenge)

![DM0.5](https://raw.githubusercontent.com/dexmal/opendm/main/docs/image/header.png)

<p align="center">
  <a href="https://www.dexmal.com/blog/dm0.5/index_en.html"><img src="https://img.shields.io/badge/📖-Tech_Blog-blue" alt="Tech Blog"></a>
  <a href="https://github.com/dexmal/opendm"><img src="https://img.shields.io/badge/GitHub-OpenDM-181717?logo=github" alt="GitHub"></a>
  <a href="https://huggingface.co/datasets/RoboChallenge/Table30v2"><img src="https://img.shields.io/badge/Dataset-Table30v2-0EA5E9?logo=huggingface" alt="Table30v2 Dataset"></a>
  <a href="https://maas.dexmal.com/"><img src="https://img.shields.io/badge/MaaS-Online-brightgreen.svg" alt="MaaS"></a>
</p>

OpenDM-format BF16 checkpoint for **ARX5** on RoboChallenge Table 30 v2.
Use with [OpenDM](https://github.com/dexmal/opendm) `third_party/robochallenge_inference`
(`configs/generalist/arx5.yaml`).
See the [DM05 RoboChallenge Table 30 v2 Inference Guide](https://github.com/dexmal/opendm/blob/main/docs/en/dm05_robochallenge.md).

Weights: BF16 `model.safetensors`.

## Model Card

| Field | Value |
| --- | --- |
| Config | `generalist/arx5` |
| Env vars | `ARX5_CHECKPOINT` / `ARX5_NORM_STATS` |
| OpenDM `robot_type` | `ARX5` |
| Control | EEF relative (`single_arm_target=eef`) |
| Cameras | Head / Left wrist / Right wrist |
| Platform cams | `cam_global` → image_0, `cam_side` → image_1, `cam_arm` → image_2 |
| Native state / action stats | 7 / 7 (action-only quantile `norm_stats.json`) |
| Defaults | `action_horizon=50`, `action_playback_target_steps=25`, `history_mode=logical_step`, `is_history=true` |

## Table30 Tasks

`arrange_flowers`, `hang_the_cup`, `pick_out_the_green_blocks`, `press_the_button`,
`turn_on_the_light_switch`, `water_the_flowers`, `wipe_the_table`

Per-task horizon overrides: `third_party/robochallenge_inference/configs/generalist/arx5.yaml` → `task_overrides`.

## Use with OpenDM RoboChallenge Inference

The RoboChallenge client now lives in OpenDM at
`third_party/robochallenge_inference` (`configs/generalist/arx5.yaml`).
See the [DM05 RoboChallenge Table 30 v2 Inference Guide](https://github.com/dexmal/opendm/blob/main/docs/en/dm05_robochallenge.md).

```bash
# From the OpenDM repository root.
export OPENDM_ROOT=/path/to/opendm
pip install -e ".[fast-infer]"

cd third_party/robochallenge_inference
export ARX5_CHECKPOINT=/path/to/DM05-ARX5
export ARX5_NORM_STATS=${ARX5_CHECKPOINT}/norm_stats.json
pip install -r requirements.txt

python execute.py --config-name generalist/arx5 \
  user_id=YOUR_USER_ID \
  submission_id=YOUR_SUBMISSION_ID
```

Override without env vars:

```bash
python execute.py --config-name generalist/arx5 \
  checkpoint=/path/to/DM05-ARX5 \
  norm_stats=/path/to/DM05-ARX5/norm_stats.json \
  user_id=YOUR_USER_ID \
  submission_id=YOUR_SUBMISSION_ID
```

If `ARX5_NORM_STATS` is unset, the client falls back to `${ARX5_CHECKPOINT}/norm_stats.json`.


## Runtime Profile

From `third_party/robochallenge_inference/configs/default.yaml` → `robot_profiles.arx5`:

- `action_type=leftpos`, `action_mode=relative`, `single_arm_target=eef`
- `history_mode=logical_step`, `history_max_frames=5`
- `speed=0.5`, `add_state=true`
- attn: llm/action `sdpa`, vision `flash_attention_2`

## Files

```text
.
├── config.json
├── model.safetensors
├── norm_stats.json
├── tokenizer.json
├── tokenizer_config.json
├── processor_config.json
├── chat_template.jinja
├── generation_config.json
└── README.md
```

## Citation

```bibtex
@misc{dm05,
    title  = {{DM0.5}: An Open-World Foundation Model for General-Purpose Embodied Intelligence},
    author = {{Dexmal Team}},
    month  = {July},
    year   = {2026},
    url    = {https://www.dexmal.com/blog/dm0.5/index_en.html}
}
```