Modelpublic

Dexmal/DM05

sourceHugging Facegemmaupdated 2mo agoView on Hugging Face

9likes568downloads

Model Card

DM05

DM0.5

News

—[2026-08-03] Published the physical robot modification guide for AgileX COBOT Magic and DOS-W1, documenting camera changes and the robot-name mapping used by the algorithm.
—[2026-07-24] DM0.5 has added the SO101 pick cube fine-tuned checkpoint and the LoRA SFT workflow. See the DM05 SO101 LoRA Training Guide.
—[2026-07-17] DM0.5 has open-sourced the RoboTwin2.0 generalist model checkpoint, along with the supervised fine-tuning (SFT) code built upon the DM0.5 pretrained model. See the DM05 RoboTwin2.0 Training and Evaluation Guide.
—[2026-07-09] DM0.5 is officially released. Read the technical blog for more details.

Introduction

DM0.5 is Dexmal's next-generation Vision-Language-Action model (VLA) for open-world robot control. It builds on the native embodied modeling approach introduced by DM0, with systematic upgrades for open-ended instructions, long-horizon tasks, dynamic disturbances, and multi-embodiment robot control.

OpenDM provides DM0.5 model weights, training and inference scripts, dataset registration examples, and evaluation workflows for researchers and developers to train, fine-tune, evaluate, and deploy the model.

Fine-Tuned Models

Model	Description	Checkpoint	Documentation
DM05-libero	LIBERO fine-tuned DM0.5 model for evaluation	🤗 DM05-libero	Training & Evaluation
DM05-robotwin2	RoboTwin2.0 fine-tuned DM0.5 model for evaluation	🤗 DM05-robotwin2	Training & Evaluation
DM05-SO101-Pick-Cube	SO101 Pick Cube fine-tuned DM0.5 model for real-robot evaluation	🤗 DM05-SO101-Pick-Cube	LoRA SFT Training Guide
DM05-VLA-Arena	VLA-Arena fine-tuned DM0.5 model for evaluation	🤗 DM05-Vla-Arena	Training & Evaluation
DM05-Table30v2	RoboChallenge Table 30 v2 DM0.5 model collection for evaluation	🤗 DM05-Table30v2	RoboChallenge Table 30 v2 Inference Guide

Quick Start

We recommend using Docker to set up the runtime environment first, which helps avoid version mismatches across CUDA, PyTorch, flash-attn, and other dependencies on the host machine.

Requirements

text

System requirements:
Ubuntu 20.04 / 22.04
NVIDIA GPU
NVIDIA Driver
Docker
NVIDIA Container Toolkit
Conda (optional, only required for local pip installation)

Recommended GPUs:
RTX 4090, A100, H100, H20
8 GPUs are recommended for training, and 1 GPU is sufficient for deployment inference.

Docker Installation

bash

git clone https://github.com/dexmal/opendm.git
cd opendm

docker run -it --rm --gpus all --network host \
  --name opendm \
  --shm-size=16g \
  -v "$PWD":/app/opendm \
  -w /app/opendm \
  dexmal/opendm:latest /bin/bash

# Run from the OpenDM repository root inside the container.
conda activate opendm
pip install -e .

Local Installation

bash

conda create -n opendm python=3.10 -y
conda activate opendm

pip install torch torchvision \
  --index-url https://download.pytorch.org/whl/cu128

pip install ninja packaging
MAX_JOBS=2 pip install flash-attn --no-build-isolation

# Enter the OpenDM repository root.
cd opendm
pip install -e .

Inference

After installing the environment and initializing the source code, you can start the model inference service. The service loads the specified checkpoint and exposes an HTTP endpoint for benchmark clients or other applications to request action predictions. Use a checkpoint that contains norm_stats.json, or make sure the matching stats already exist under ./norm_stats/.

bash

script/dm05_launcher.sh \
  --task inference \
  --nproc_per_node 1 \
  --model-config.model-name-or-path ./checkpoints/DM05 \
  --model-config.chunk-size 50 \
  --inference-config.port 7891

Arguments:

—--task: task type. Use inference for inference.
—--nproc_per_node: number of GPUs on a single node. 1 GPU is sufficient for inference.
—--model-config.model-name-or-path: model checkpoint path.
—--model-config.chunk-size: action chunk length.
—--inference-config.port: inference service port.

During inference, the service first looks for norm_stats.json in the checkpoint directory. If it is not found, it falls back to the matching file under ./norm_stats/, which is normally generated during training for the same dataset, action mode, and chunk size.

After the service starts, send a test request to verify that the endpoint returns a valid response:

bash

bash tests/curl_demo.sh http://SERVER_IP:7891/process_frame

/process_frame accepts a multipart/form-data request:

—text: task instruction.
—states: JSON array of the current robot state. The dimension and order must match the model's training and normalization statistics.
—image: image files, one field per configured image key. The order must match --inference-config.image-keys.
—robot_type: optional built-in robot type. Currently only DOS W1 is supported. It provides the robot state description when relative actions need to be converted back to absolute actions.
—control_mode and speed: text conditioning fields required when directly serving the pretrained Dexmal/DM05 model. They are normally not required for SFT checkpoints unless your SFT data was trained with the same fields.

A successful response has the following shape.

text

{
  "response": [
    [0.012, -0.034, 0.18, "..."],
    [0.015, -0.031, 0.17, "..."],
    ...
  ]
}

Community and Support

—Learn more about Dexmal products and model updates on the Dexmal website.
—If you encounter issues, please report them through GitHub Issues.
—For further discussion, scan the WeChat QR code to contact us.

We will continue to release more model weights, technical documentation, and examples. If this project is helpful to you, please consider giving us a star on GitHub ![GitHub](https://github.com/dexmal/opendm). Your support helps us move forward.

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}
}