MEVIUS2 Quadruped Robot Platform

• MEVIUS2 is an open-source quadruped robot platform that bridges small-scale prototypes and commercial systems by combining durable metal fabrication with advanced multimodal sensing.
• It employs sheet metal welding and precision machining to simplify assembly and reduce failure points while ensuring robust structural integrity.
• The platform demonstrates effective locomotion and perception through terrain mapping, gait evaluation, and reinforcement learning control for research-grade performance.

MEVIUS2 is an open-source quadruped robot platform developed to provide a large, structurally robust, and perceptually capable alternative to both small-scale 3D-printed open-source robots and commercial systems such as Boston Dynamics’ Spot. MEVIUS2 employs sheet metal welding and precision metal machining to achieve a body scale and durability comparable to commercial robots, with all hardware and software designed for accessibility and reproducibility. The system integrates a comprehensive multimodal perception suite—including LiDARs and a high dynamic range camera—enabling advanced terrain mapping and environmental interaction. All hardware schematics, firmware, software, and simulation tools are publicly available, supporting research, education, and practical deployment at a fraction of the cost of proprietary platforms (Kawaharazuka et al., 23 Mar 2026).

1. System Specifications and Comparative Scale

MEVIUS2 is designed with a focus on bridging the gap between small, 3D-printed open-source quadruped robots and larger commercial products. Key parameters are summarized below:

Parameter	MEVIUS2	Boston Dynamics Spot	Solo-12/PAWDQ/Stanford Doggo
All-up weight	22.9 kg	33.8 kg	~2–5 kg
Leg length	0.30 m	0.34 m	0.16–0.22 m
Max joint torque	60 Nm	85 Nm	12–22 Nm
Main structure	Metal welded (Al)	Alloy/polymer	3D-printed plastic
Multimodal perception	Yes	Yes	No
Cost (approx.)	13 k USD	75 k USD	2–5 k USD

The payload area volume of MEVIUS2 (~0.05 m³) is usable for auxiliary equipment or expanded battery configuration. In terms of cost efficiency, MEVIUS2 provides a practical research-scale platform (leg length 0.30 m) with robust mechanics and perception capabilities, addressing critical shortcomings of plastic-printed designs—namely fragility, scaling limitations, and lack of environmental sensors (Kawaharazuka et al., 23 Mar 2026).

2. Mechanical Architecture and Fabrication

2.1 Material Selection and Manufacturing

MEVIUS2’s structure utilizes A7075 aluminum alloy for machined components, prioritizing stiffness-to-weight ratio. Sheet-metal elements—including the core Base-Link, Hip-Link, and Lidar-Cover—are fabricated from A5052 aluminum (t ≈ 3 mm), balancing machinability and yield strength. Every structural component is orderable via online services such as MISUMI meviy, using supplied STEP files.

Sheet parts are joined using MIG welding, resulting in large contiguous structures that reduce fastener count and interface complexity.

2.2 Component and Assembly Optimization

Reduction in unique metal components is a core design goal; the system integrates leg mounts, PCB trays, and sensor brackets into the Base-Link, consolidating more than 50 parts into a single welded entity. The parallel-link knee mechanism positions the actuating motor at the hip to minimize distal inertia.

The total number of unique metal parts is 16. This facilitates reliable large-scale assembly, minimizes potential failure points, and simplifies the supply chain.

2.3 Structural Analysis

Assuming a cantilevered beam of length $L$, width $w$, and thickness $t$ under end load $F$, section modulus is $Z = \frac{w t^2}{6}$, bending stress $\sigma = \frac{6 F L}{w t^2}$, and deflection $\delta = \frac{F L^3}{3 E I}$, with $I = \frac{w t^3}{12}$. For a critical segment ($w=50$ mm, $t=3$ mm, $w$0 mm, $w$1 GPa, $w$2 N): $w$3 MPa (well below A5052’s yield of $w$4200 MPa), $w$5 mm.

2.4 Actuation

The twelve primary joints are driven by Robstride03 brushless-DC actuators (continuous torque 20 Nm, peak 60 Nm, 8,000 rpm, 12-bit encoder, CAN interface). The architecture supports dynamic, torque-controlled locomotion with minimized weight distal to the hip.

3. Perception Hardware and Sensing Methods

3.1 Sensor Suite

MEVIUS2 integrates two Livox Mid-360 LiDARs (360° horizontal, ±38.4° vertical, single-beam nonrepetitive scan) and a Tier IV C1 HDR RGB camera (1 MP, >120 dB dynamic range, global shutter). An IMU (either integrated with LiDAR or via a separate Ouster unit) provides 200 Hz state feedback.

3.2 Sensor Layout and Calibration

Sensors are mounted on the main Base-Link, with LiDARs yawed at ±30° and an HDR camera facing front-center, yielding comprehensive coverage: LiDAR FOV is 360° (horizontal), ±19.2° (vertical per sensor), and camera FOV is 90° × 60°. Extrinsic calibration is explicit via measured transformations $w$6 defined as $w$7, with $w$8 parameterized as consecutive $w$9 and $t$0 axis rotations.

3.3 Perception Algorithms

MEVIUS2 employs elevation mapping wherein LiDAR point clouds and camera depth images are fused to construct voxel grid heightmaps. Terrain segmentation utilizes vertical gradient thresholding. Planned extensions include the use of YOLOv3 for object detection on HDR images.

4. Locomotion Control and Reinforcement Learning

4.1 Software Ecosystem

Simulation, training, and deployment flow across NVIDIA IsaacGym (for RL policy training with full Python API support), MuJoCo (for physics validation), and onboard NVIDIA Jetson edge computing (running Ubuntu, Python 3, PyTorch). Communication and high-level orchestration optionally leverage ROS2.

4.2 Reinforcement Learning Policy

States ($t$1) comprise base orientation $t$2, base linear velocity $t$3, joint angles $t$4, and velocities $t$5. Actions ($t$6) are 12-dimensional torque commands. The reward function is:

$t$7

where the terms regulate progression speed, energy consumption, control smoothness, and slip events. Policy training is performed using PPO algorithms for up to 12 million steps.

4.3 Whole-Body Stabilization

A 100 Hz control loop updates state estimation and gait references, with 500 Hz low-level torque control per actuator. Optionally, a QP-based whole-body controller solves for joint velocities $t$8 that minimize $t$9 subject to physical and actuator constraints.

5. Experimental Validation and Performance

5.1 Traversal and Gait Evaluation

MEVIUS2 demonstrates robust traversal across concrete, grass, and soil—maintaining a velocity near 0.3 m/s without slippage. Stair climbing trials (17 cm rise, 28 cm run) achieved a perfect 5/5 success rate; maximal single-step height was 20 cm. The robot manages slopes up to 20°, and for wet/slippery stairs, balance losses occurred in 2 of 10 runs but with 80% recovery.

5.2 Perception Benchmarking

Livox point cloud density reaches 10,000 points/m² at 2 m range, with 0.2° vertical resolution. Depth mapping achieves 3 cm RMSE, and lens undistortion introduces <0.5 pixel error. Indoor elevation maps match ground-truth topography within 5 cm.

5.3 Failure Modes

Extreme terrain (mud-drenched tires) can induce occasional foot slippage; the RL policy’s slip penalty and foot trajectory replanning provide mitigation. Sensor occlusion during stair ascent highlights the benefit of combined LiDAR and wide-angle imaging.

6. Open-Source Assets, Package Structure, and Replication

6.1 Repository and Documentation

All hardware, electronics, and software resources are maintained at [https://github.com/haraduka/mevius2]. Repository structure includes:

• /hardware/: STEP/solid models (16 unique metal parts), BOM.csv with supplier references.
• /electronics/: PCB schematics for CAN-USB, safety, and relays.
• /software/: RL policy source, gym wrappers, ROS2 nodes, perception (elevation mapping, point cloud fusion), and deployment scripts.
• /training/: IsaacGym and PPO hyperparameter configurations, example learning logs.

6.2 Replication Protocol

Construction involves direct ordering of all STEP-defined metal components via MISUMI meviy, 3D printing of TPU feet, and standard assembly of motors and battery electronics. A Jetson Xavier NX is provisioned with Ubuntu 20.04 and relevant software dependencies (Python 3.8, PyTorch, IsaacGym, MuJoCo). System launch and policy deployment are managed via a combination of ROS2 and Python scripts, with all configuration files and neural policy weights supplied.

By leveraging scalable sheet-metal fabrication and off-the-shelf components, MEVIUS2 enables reproducible research with advanced multimodal sensing at an accessible cost (~13 k USD), filling the open-source gap for large-format, high-performance quadruped robotics (Kawaharazuka et al., 23 Mar 2026).