DM05-MEM-Robodojo-Sim

DM05-MEM-Robodojo-Sim

<p align="center"> <a href="https://www.dexmal.com/blog/dm0.5/index_en.html"><img src="https://img.shields.io/badge/%F0%9F%93%96-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://robodojo-benchmark.com/leaderboard"><img src="https://img.shields.io/badge/Benchmark-RoboDojo-orange" alt="RoboDojo Leaderboard"></a> <a href="https://maas.dexmal.com/"><img src="https://img.shields.io/badge/MaaS-Online-brightgreen.svg" alt="MaaS"></a> </p>

Introduction

DM05-MEM-Robodojo-Sim is the RoboDojo simulation fine-tuned generalist checkpoint of DM0.5, Dexmal's open-world Vision-Language-Action foundation model for embodied intelligence. DM0.5 uses a Gemma 3 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.

This memory-enabled checkpoint targets the ARX X5 bimanual embodiment used by RoboDojo-Sim. It consumes current head, left-wrist, and right-wrist RGB views together with up to 20 head-camera history frames sampled at 1 FPS. At the beginning of an episode, unavailable history slots are left-padded until enough observations have been collected. The model generates 14-dimensional absolute joint-position action chunks of length 50, and the RoboDojo evaluation adapter executes the first 25 actions from each predicted chunk.

RoboDojo-Sim Results

The following results are a snapshot of the official RoboDojo leaderboard on August 24, 2026.

Generalization is reported separately for the standard (Gen-Std) and randomized (Gen-Rand) settings. See the official RoboDojo leaderboard for detailed per-task results, rollout videos, metric definitions, and the latest rankings.

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:
A100, H100, H20
1 GPU is sufficient for deployment inference.

Download the Checkpoint

pip install -U "huggingface_hub[cli]"

hf download Dexmal/DM05-MEM-Robodojo-Sim \
  --local-dir ./checkpoints/DM05-MEM-Robodojo-Sim

The checkpoint directory must include its matching norm_stats.json.

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 .

RoboDojo-Sim Testing

Use RoboDojo's official evaluation workflow to test this checkpoint. RoboDojo runs the simulation benchmark client, while policy integration and serving are managed through XPolicyLab.

Follow these official guides in order:

1. Install RoboDojo and download its assets and data.
2. Set up or select the model integration in XPolicyLab.
3. Run the RoboDojo Quick Evaluation workflow.

Use the released checkpoint together with its bundled norm_stats.json. Keep the official camera order and absolute joint-position action mode unchanged, provide up to 20 head-camera history frames sampled at 1 FPS with left-padding during episode warm-up, predict 50-step action chunks, and execute the first 25 actions from each chunk.

For the model-specific XPolicyLab integration, see XPolicyLab PR #101. If the PR has already been merged, use the official XPolicyLab code from the main branch directly; otherwise, use the integration code provided by the PR.

For benchmark rules, multi-seed reporting, and verified leaderboard publication requirements, see the official evaluation protocol. Detailed results should be referenced directly from the official leaderboard.

Intended Use and Limitations

This checkpoint is intended for research evaluation on RoboDojo-Sim with the matching ARX X5 observation/action convention, normalization statistics, camera order, action horizon, and history-input policy. Using a different embodiment, state/action ordering, camera layout, action mode, or history sampling strategy requires an adapted configuration and may substantially reduce performance.

RoboDojo is an intentionally challenging benchmark. The aggregate results above do not imply reliable success on every task, and performance in simulation does not guarantee safe or successful real-world deployment.

Community and Support

• Learn more about Dexmal products and model updates on the Dexmal website.
• If you encounter OpenDM issues, please report them through GitHub Issues.
• For RoboDojo setup and evaluation questions, see the official documentation and community page.
• For further discussion with Dexmal, 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}
}

@article{chen2026robodojo,
    title   = {{RoboDojo}: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies},
    author  = {Chen, Tianxing and Chen, Yue and Li, Zixuan and Tang, Junyuan and Su, Kailun and Wan, Weijie and Chen, Baijun and Lu, Haoran and Yan, Haowen and Su, Honghao and others},
    journal = {arXiv preprint arXiv:2607.04434},
    year    = {2026}
}