University of Tokyo MEVIUS

No image

Not for sale

Not for sale (research)Open-source designSheet-metal bodyParts ordered online

Not for saleShipping: N/A.

Overview

A 15.5 kg metal quadruped from Kento Kawaharazuka's group at the University of Tokyo, presented at Humanoids 2024. Every part can be ordered online, with the aluminum body and hip parts made by MISUMI's meviy sheet-metal welding service. Its frame uses only 10 part types, and only the nylon-composite calves and TPU feet are 3D printed. A reinforcement learning policy that allows for CAN bus delay let it walk over grass, slopes and cobblestone steps. In falls during tests only a PLA camera mount broke. CAD, parts list and code are free, but there is no step-by-step assembly guide. In December 2024 RT Corporation announced a plan to productize a quadruped based on MEVIUS for research, education and industry, aiming for build-to-order sales in fiscal 2025. RT's Mujina followed in November 2025, and RT has not said whether it is based on MEVIUS.

Type: Quadruped robot · Research & legacy platforms
Maker: University of Tokyo, JSK Robotics Laboratory (Japan)
Released: 2024
Price: Not for sale
Status: Not for sale (research)
Best for: Professionals & labs
Programming: Open source (hardware + software)

Price and shipping

Open design. The project's parts list totals ~$12,390 (¥1,734,612) for metal parts, bearings, motors and electronics, excluding 3D-printed parts and screws. A cheaper column with substitute motors, PC, CAN adapter and IMU totals ~$5,300.

Shipping: N/A. Not sold. The design files are free, and builders order the parts themselves.

Specifications

Category	Quadrupeds · Research & legacy platforms
Released	2024
Country	Japan
Availability	Not for sale (research)
DIY build plans	MEVIUS (metal quadruped) · parts $12,400 (estimate) · difficulty 4/5
Programmability	Open source (hardware + software)
Built-in autonomy	None / scripted · Moves only when controlled, or plays preset motions
LLM support	None · No built-in LLM. The open Python and ROS code could be extended to connect one, but none is provided.
MuJoCo simulation	Official model · The GitHub repo includes a MuJoCo model and scene, and the main control script has a simulation mode. haraduka/mevius (GitHub) ↗
NVIDIA Isaac simulation	Official model · Walking policies are trained in Isaac Gym with the authors' MEVIUS branch of legged_gym. haraduka/legged_gym, mevius branch ↗
Gazebo simulation	Maker URDF only · A URDF model is published in the repo and on the project's Google Drive. No Gazebo package was found. haraduka/mevius (GitHub) ↗
Best for	Professionals & labs
Vision sensors	Camera
Actuators	Brushless / direct-drive motors
Audio	No audio
Compute type	x86 PC (Intel / AMD)
Compute	Intel NUC 12 Pro mini PC (NUC12WSKi7, Core i7-1260P) running Python, linked to all 12 motors over one CAN bus through an ESD CAN-USB/2 adapter
Weight	15.5 kg (about 34 lb), with 0.25 m legs
Speed	No measured top speed published. The released walking policy is trained for commands up to 1 m/s forward and sideways
DOF	12 (3 per leg), all T-Motor AK70-10 brushless servo motors with a 10:1 gear ratio (25 N·m peak)
Sensors	Intel RealSense T265 tracking camera for body velocity, plus motor encoders. Logic runs on a 3S 6000 mAh LiPo and the motors on two 6S 3300 mAh LiPos in series. Wireless emergency stop
SDK	Open source under MIT on GitHub. Python control code using ROS 1, with reinforcement learning through a fork of legged_gym (Isaac Gym). A 3D mouse sends walking commands

Common questions

How much does MEVIUS cost?

Not for sale. Open design. The project's parts list totals ~$12,390 (¥1,734,612) for metal parts, bearings, motors and electronics, excluding 3D-printed parts and screws. A cheaper column with substitute motors, PC, CAN adapter and IMU totals ~$5,300.

Can MEVIUS be bought today?

No. It was never sold to the public.

Who makes MEVIUS?

MEVIUS is made by University of Tokyo, JSK Robotics Laboratory (Japan). It came out in 2024.

Can MEVIUS be programmed?

Open source (hardware + software). Open source under MIT on GitHub. Python control code using ROS 1, with reinforcement learning through a fork of legged_gym (Isaac Gym). A 3D mouse sends walking commands.