MEVITA: Open-Source Bipedal Robot

• MEVITA is an open-source bipedal robot platform that integrates a minimal viable mechanism with welded metal construction and a 5-DoF per leg configuration.
• The design minimizes unique parts using e-commerce components, employing a hierarchical CAN-based control stack and sensor fusion for stability and performance.
• Extensive simulations with IsaacGym/LeggedGym and MuJoCo, along with hardware tests, demonstrate reliable locomotion across varied terrains with reinforcement learning.

MEVITA is an open-source bipedal robot platform constructed entirely from components readily available through e-commerce channels and assembled via sheet metal welding. The system is designed to overcome scalability and robustness limitations typical of 3D-printed bipedal robots, while minimizing the number of unique parts and supporting reproducibility for research and educational purposes. MEVITA combines a minimal viable mechanism (5 degrees of freedom per leg), welded metal links for structural efficiency, a CAN-based electronics/control stack, and reinforcement learning (RL) for gait generation. All design files, electronics schematics, control software, and RL training environments are openly released (Kawaharazuka et al., 25 Aug 2025).

1. Mechanical Design and Fabrication

MEVITA adopts a minimal-viable configuration with five degrees of freedom (DoF) per leg: Hip-Yaw, Hip-Roll, Hip-Pitch, Knee-Pitch, and Ankle-Pitch, deliberately omitting ankle-roll for design simplicity. The structural solution emphasizes scalability and robustness by using an all-metal frame composed of A7075 aluminum machined blocks, A5052 aluminum sheet, and SUS304 stainless sheet. Critical load-bearing and geometrically complex links—including the Base-Link, Hip1-Link, Hip2-Link, and Calf-Motor-Mount—are fabricated by welding together multiple laser-cut sheets, a process that enables the realization of large and complex forms (e.g., the 300×320×150 mm Base-Link torso) as single, monolithic components.

All welded and machined parts are provided as STEP files suitable for direct upload to services such as MISUMI’s meviy platform, enabling instant online quoting and procurement. This design philosophy minimizes the component count; after trade-offs for cost and maintainability (Configuration-4 in the original data), the final mechanical design uses 18 unique metal parts, compared to a theoretical lower bound of 13. All actuators (T-MOTOR AK10-9, AK70-10), bearings, shafts, and fasteners are specified to be e-commerce available.

Kinematic and dynamic modeling relies on a Denavit–Hartenberg chain for each leg: $$T_{\mathrm{foot}/\mathrm{base}} = \prod_{i=1}^5\left[\mathrm{Rot}_{\mathrm{axis}_i}(\theta_i)\,\mathrm{Trans}_{i}\right].$$ Center of mass (CoM) in the sagittal plane is computed by

$$x_\mathrm{CoM} = \frac{1}{M}\sum_{i} m_i\,x_i,\quad z_\mathrm{CoM} = \frac{1}{M}\sum_{i} m_i\,z_i,$$

where $M = \sum_i m_i$. The zero-moment point (ZMP) for stability analysis is given by

$$x_{\mathrm{ZMP}} = \frac{\sum_i m_i(\ddot z_i+g)x_i - \sum_i m_i \ddot x_i z_i}{\sum_i m_i (\ddot z_i + g)}.$$

Joint actuator torques for static and dynamic scenarios follow the classical expressions: $$\tau_j = I_j\ddot\theta_j + \sum_{k\ge j} m_k g l_{jk} + f_\mathrm{external}l_{\mathrm{contact},j}.$$

2. Electronics Architecture and Control Stack

MEVITA integrates joint encoders within each AK series actuator, supplementing these with an onboard IMU (from the Livox Mid-360 LiDAR) for body orientation and angular velocity. The electronic architecture is based on CAN for actuator communication, with two USB↔CAN adapters managing 10 motors per bus. Low-level position-PID control operates at $\sim$1 kHz on a Jetson Orin Nano via CAN-USB interfacing.

The control hierarchy consists of:

• Low-level: motor position-PID (∼1 kHz loop).
• Mid-level: sensor fusion for pose and CoM (encoder + IMU).
• High-level: RL-based policy generates velocity commands $(v_x, v_y, \omega_z)$, converted to joint trajectories and sent to the PID loop.

System communication achieves CAN loop times around 1 ms per cycle, with higher-level PC controller tasks running at 100–200 Hz. A wireless emergency-stop circuit is hardwired to the power system for safety.

3. Simulation, Reinforcement Learning, and Sim-to-Real Transfer

Training and validation follow a hierarchical pipeline:

• Simulation: Training is performed using IsaacGym/LeggedGym for highly parallel GPU-based simulation, and additional verification in MuJoCo for accurate contact dynamics. The contact model uses randomized Coulomb friction with near-zero restitution.
• Reinforcement learning formulation:
  • State: $$\{ \omega, \hat g, v_x^*, v_y^*, \omega_z^*, q, \dot q \}$$
  • Action: 5 joint targets $\Delta\theta$ per leg.
  • Reward:

  $$r = w_1(-\|v_x-v_x^*\|^2 - \|v_y-v_y^*\|^2 - \|\omega_z-\omega_z^*\|^2) + w_2(-\textstyle\sum |\text{torque}|) + w_3(-\text{energy}_\text{penalty}) + w_4(\text{alive}_\text{bonus})$$ - Algorithm: Proximal Policy Optimization (PPO) via Stable-Baselines3.

Domain randomization covers link mass/inertia, friction, sensor noise, and actuator latency. Sim-to-Real transfer proceeds via Sim-to-Sim verification in MuJoCo, followed by direct hardware deployment with minimal fine-tuning.

4. Experimental Validation and Comparative Analysis

MEVITA’s RL-trained policy was deployed without retuning across diverse surfaces: uneven indoor tiles, grass, dirt, smooth concrete, and gentle (5°) slope. The robot demonstrates robust walking, maintaining tracking error within ±0.05 m/s RMS forward velocity and yaw error under 0.02 rad/s. Stable locomotion is observed up to ~0.3 m/s, with ZMP inside the support polygon for lateral sway up to ±0.05 m. Typical motor power draw averages ~200 W (total), yielding a cost of transport (CoT) near 2.5 J/N·m.

A comparative summary across several bipedal platforms:

Platform	Mass (kg)	Leg Length (m)	Degrees of Freedom (per leg)	Unique Metal Parts
MEVITA	19.8	0.32	5	18
Duke Humanoid	30	0.25	5	24
Bolt (plastic)	1.3	—	—	—

MEVITA approaches the theoretical minimum component count (within 25%) and offers greater scalability and impact resilience compared to plastic-based open-source platforms.

5. Open-Source Release and Community Provisions

All design assets are openly available at https://github.com/haraduka/mevita. The repository includes:

• Hardware: STEP/DXF for all metal parts, welding profiles, and 3D-printable TPU parts.
• Electronics: Wiring diagrams, CAN bus configuration, battery and PCB designs.
• Software: ROS-native drivers (AK-motors, IMU/LiDAR), state estimation, control stacks.
• Reinforcement Learning: IsaacGym environments, MuJoCo XMLs, PPO scripts, reference policy checkpoints.
• Documentation: Tutorials, assembly guides, calibration procedures, operation instructions.

The release is covered by an MIT-style license. Community standards are enforced via a CONTRIBUTING.md detailing coding standards, submission criteria, and e-commerce sourcing guidelines, complemented by templated issue forms for hardware, software, or training concerns.

6. Context and Significance

MEVITA addresses a critical bottleneck in open-source bipedal robotics: the trade-off between scalability, manufacturability, and robustness. By leveraging e-commerce provisioning and sheet-metal welding, it sets a reproducible standard for medium-to-large scale bipedal robots that can be assembled outside institutional machine shops. The use of reinforcement learning for gait generation, coupled with a minimal-parts strategy, facilitates both rapid experimentation and practical deployment. Its modular and community-oriented release strategy positions it as a reference platform for research spanning locomotion control, sim-to-real transfer, hardware design, and leg