MEVIUS2

This repository includes the hardware and software components of "MEVIUS2: Practical Open-Source Quadruped Robot with Sheet Metal Welding and Multimodal Perception"

Please refer to the following links for more information.

• Project Page
• arXiv
• YouTube

Hardware

You can find all the hardware components in the following link

• Goodle Drive

• RobStride03
• RobStride03 Manual
• Motor Index Order: FR(1-3) --> FL(4-6) --> BR(7-9) --> BL(10-12) / Proximal --> Distal
• Link Name: FR_scapula_link --> FR_thigh_link --> FR_calf_link
• Joint Name: FR_collar --> FR_hip --> FR_knee
• IsaacGym Joint Order: BR --> BL --> FR --> FL / collar --> hip --> knee

Software Dependencies

Please install the following dependencies.

• ROS # if ubuntu >= 22.04, please install ROS-O
• spacenav_node # optional
• virtual_joystick # optional
• legged_gym # for training
• livox_ros_driver2
• FAST_LIO
• elevation_mapping_cupy

Please install the following python packages.

• pytorch
• scipy
• can
• mujoco
• mujoco-python-viewer
• urdf_parser_py
• ds4drv (optional)

Build

mkdir -p ~/mevius2_ws/src
cd ~/mevius2_ws/src
git clone git@github.com:haraduka/mevius2.git
cd ..
catkin build
source ~/mevius2_ws/devel/setup.bash

Setup Can Devices

udevadm info -a -p /sys/class/net/can0 | grep serial # ATTRS{serial}=="002C00355457530220353930"

# see config/99-usb-can.rules
sudo vim /etc/udev/rules.d/99-usb-can.rules

# add the below command
SUBSYSTEM=="usb", ATTRS{serial}=="002C00355457530220353930", SYMLINK+="can_usb_0"
SUBSYSTEM=="usb", ATTRS{serial}=="001F00275457530120353930", SYMLINK+="can_usb_1"

# then (please plug out and in the device)
sudo udevadm control --reload-rules
sudo udevadm trigger

Reinforcement Learning

# in legged_gym
git remote add haraduka git@github.com:haraduka/legged_gym.git
git fetch haraduka
git checkout haraduka/mevius2

python3 legged_gym/scripts/train.py --task mevius2
python3 legged_gym/scripts/play.py --task mevius2 --load_run ( log_dir )
cp logs/mevius2/( log_dir )/exported/policies/policy_1.pt /tmp/policy.pt
roscd mevius2
cp /tmp/policy.pt models/policy.pt

Usage

Please test several codes by running the following command.

python3 scripts/mevius2_utils.py

Please test the motors by running the following command.

./bin/can_setup.sh
python scripts/xiaomimotor_test.py --ids 1 --task sense

Please test the real robot by running the following command.

# in the PC of MEVIUS2
roscore
sudo ds4drv
roslaunch mevius2 livox.launch
roslaunch mevius2 fast_lio.launch
./bin/can_setup.sh
python3 scripts/mevius2_main.py

# Cross: SITDOWN/STANDUP, Circle: STANDUP/WALK for ps4joy

# if you want to simulate the robot in mujoco
python3 scripts/mevius2_main.py --sim

# please press D to turn off every-step rendering in mujoco viewer

How to Make Models

• make detailed models in solidworks and export as 3dxml
• convert 3dxml to dae using convert-urdf-mesh with --force-zero-origin option
• make simple models in solidworks and export as 3dxml
• convert 3dxml to dae using convert-urdf-mesh with --force-zero-origin option
• make simple models in solidworks and export as stl
• use converted detailed model for URDF, and use simple models for STL and DAE files
• set joint limits, max velocity, max effort, and correct collision models in URDF
• make mevius2_dae.urdf and mevius2_stl.urdf (just each file path is different)
• use mevius2_dae.urdf for rviz and isaacgym (change mesh path), and use mevius2_stl.urdf for mujoco
• cp mevius2_stl.urdf to meshes/mevius2_mujoco.urdf, make collision model the same with visual model, add world link and floating joint
• compile urdf into xml, change a little (stl path, geom name, and default base_link pos) and load from scene.xml

Acknowledgements

• scripts/isaacgym_torch_utils.py is from IsaacGym
• scripts/legged_gym_math.py is from LeggedGym

Citation

@article{kawaharazuka2026mevius2,
  title={{MEVIUS2: Practical Open-Source Quadruped Robot with Sheet Metal Welding and Multimodal Perception}},
  author={Kento Kawaharazuka and Keita Yoneda and Shintaro Inoue and Temma Suzuki and Jun Oda and Kei Okada},
  journal={IEEE Robotics and Automation Practice},
  year={2026},
}