DM05-vla-arena

Introduction

DM05-vla-arena is the VLA-Arena fine-tuned checkpoint of DM0.5, Dexmal's open-world Vision-Language-Action foundation model for embodied intelligence. DM0.5 uses a Gemma3 4B vision-language backbone with a 680M Action Expert to generate continuous robot actions, and is designed for natural-language manipulation, zero-shot generalization, efficient downstream fine-tuning, long-horizon historical context, robust policy behavior, and transfer across robot embodiments.

VLA-Arena Results

Method	Metric	L0	L1	L2
DM0.5	SR	89.0%	53.6%	44.1%
	CC	1.9	18.75	22.52

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

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

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

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 .

VLA-Arena Testing

Use the VLA-Arena-specific experiment configuration when evaluating this checkpoint. The inference service and evaluation client run separately and communicate over HTTP.

For the complete training and evaluation workflow, see the DM05 VLA-Arena Training and Evaluation Guide.

Start the VLA-Arena Inference Service

Run this command from the OpenDM repository root. The checkpoint must contain the matching norm_stats.json.

script/dm05_launcher.sh \
  --exp playground/dm05_vla_arena.py \
  --task inference \
  --nproc_per_node 1 \
  --model-config.model-name-or-path ./checkpoints/DM05-vla-arena-bf16 \
  --inference-config.output-action-dim 7

Keep the service running while executing the benchmark.

Prepare the VLA-Arena Evaluation Client

git clone https://github.com/PKU-Alignment/VLA-Arena
cd VLA-Arena

Install the VLA-Arena dependencies:

sudo apt-get install -y libosmesa6-dev libglfw3 libgl1-mesa-glx libglib2.0-0

pip install robosuite==1.5.1 bddl numpy==1.26.4 requests tqdm pyyaml "imageio[ffmpeg]" pillow

pip install -e .

export MUJOCO_GL=osmesa

Copy the evaluation files from the OpenDM repository:

mkdir -p vla_arena/models/DM05

# Copy evaluation files from the OpenDM repository
cp <path-to-opendm>/third_party/vla_arena/eval.py vla_arena/models/DM05/eval.py
cp <path-to-opendm>/third_party/vla_arena/eval_config.yaml vla_arena/models/DM05/eval_config.yaml

Edit vla_arena/models/DM05/eval_config.yaml so that server_url points to the running inference service:

server_url: "http://<SERVER_IP>:7891/process_frame"
request_timeout: 30

action_horizon: 20
replan_steps: 10
robot_type: "Franka"
batch_size: 1
speed: "0.5"

# "all" evaluates all 170 tasks (11 suites × 3 levels)
task_suite_name: "all"
task_level: 0

num_trials_per_task: 10
seeds: [7, 42, 1000]

model_name: "DM05"
local_log_dir: "./experiments/eval_results"
save_video_mode: "first_success_failure"

Run the VLA-Arena Benchmark

From the VLA-Arena repository root, run:

python -m vla_arena.models.DM05.eval \
  --config vla_arena/models/DM05/eval_config.yaml \
  --output-dir ./experiments/eval_results

Evaluation artifacts are written under <output-dir>/seed_<N>/, including results_<timestamp>.json, tasks_<timestamp>.csv, and rollout videos under videos/.

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 . Your support helps us move forward.

Citation

@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}
}

@misc{zhang2025vlaarena,
    title         = {VLA-Arena: An Open-Source Framework for Benchmarking Vision-Language-Action Models},
    author        = {Borong Zhang and Jiahao Li and Jiachen Shen and Yishuai Cai and Yuhao Zhang and Yuanpei Chen and Juntao Dai and Jiaming Ji and Yaodong Yang},
    year          = {2025},
    eprint        = {2512.22539},
    archivePrefix = {arXiv},
    primaryClass  = {cs.RO},
    url           = {https://arxiv.org/abs/2512.22539},
}