- The paper introduces MEVITA, a bipedal robot platform that uses sheet metal welding and e-commerce-sourced parts to reduce part count while ensuring robust dynamics.
- It employs a reinforcement learning pipeline with domain randomization for effective Sim-to-Real transfer, achieving stable locomotion across varied terrains.
- Comparative analysis demonstrates that MEVITA significantly lowers metal part count versus traditional designs, enhancing manufacturability and ease of maintenance.
Introduction
The paper presents MEVITA, an open-source bipedal robot platform designed for rapid, scalable, and robust assembly using exclusively e-commerce-sourced components and sheet metal welding. The work addresses critical limitations in current open-source bipedal robots, particularly the fragility and scalability constraints of 3D-printed designs and the complexity and procurement challenges of traditional metal-machined robots. MEVITA's architecture is optimized for minimal part count, manufacturability, and ease of assembly, while maintaining mechanical robustness suitable for dynamic locomotion tasks.
Design Principles and Mechanical Architecture
Minimal Viable Configuration
The authors formalize the minimal viable configuration for a 5-DoF bipedal leg (Hip-Y, Hip-R, Hip-P, Knee-P, Ankle-P), excluding the ankle roll, following conventions in recent platforms (e.g., Cassie, MIT Humanoid, Duke Humanoid). The design process systematically explores single- and double-supported joint architectures, parallel linkages for inertia reduction, and practical trade-offs in part splitting for maintainability and cost. The theoretical minimum for a double-supported, parallel-link bipedal leg is 13 unique link components; MEVITA's practical design uses 18, balancing manufacturability and maintainability.
MEVITA's frame is constructed entirely from aluminum and stainless steel, with four critical components (Base-Link, Hip1-Link, Hip2-Link, Calf-Motor-Mount) fabricated via sheet metal welding. This approach enables the integration of complex geometries into single parts, drastically reducing the total part count compared to conventional CNC-machined designs. All structural, mechanical, and electronic components are sourced from e-commerce platforms (notably MISUMI and meviy), ensuring global accessibility and reproducibility. The design is fully compatible with automated online quoting and ordering workflows using STEP files.
Actuation and Sensing
The robot employs T-MOTOR AK10-9 actuators for Hip-P and Knee-P, and AK70-10 for Hip-Y, Hip-R, and Ankle-P. The sensor suite includes a Livox Mid-360 LiDAR/IMU for environmental perception and proprioception. The circuit architecture is CAN-based, with power management via wireless emergency stop and relay, and logic/motor power supplied by LiPo batteries.
Control System and Learning-Based Locomotion
Reinforcement Learning Pipeline
Locomotion policies are trained in IsaacGym/LeggedGym, validated in MuJoCo for Sim-to-Sim transfer, and deployed to hardware for Sim-to-Real evaluation. The control input is joint angle commands for PID control; the state vector includes base angular velocity, gravity direction, desired base velocity, joint positions, and velocities. The reward function is adapted from quadrupedal locomotion literature, with domain randomization applied to mass, inertia, center of mass, friction, and latency to facilitate robust Sim-to-Real transfer.
Empirical results demonstrate that MEVITA tracks commanded linear and angular velocities with minimal discrepancy between simulation and hardware, particularly for rotational commands. Translational velocity tracking exhibits slightly larger errors in hardware, consistent with domain transfer limitations. The robot achieves stable walking across diverse environments (uneven indoor terrain, grass, dirt, concrete, slope), confirming the effectiveness of the design and learning pipeline.
Comparative Analysis
MEVITA is compared against seven contemporary bipedal robots, including Cassie, MIT Humanoid, Berkeley Humanoid, Bolt, Berkeley Humanoid Lite, and Duke Humanoid. MEVITA achieves a significant reduction in metal part count (18 vs. 24 for Duke Humanoid), approaching the theoretical minimum while maintaining practical manufacturability. Among open-source platforms, MEVITA is unique in its exclusive use of metal components and e-commerce sourcing, overcoming the scalability and robustness limitations of 3D-printed designs.
Implementation and Deployment Considerations
- Manufacturing: All CAD files and BOMs are open-source, enabling direct ordering from e-commerce platforms. Sheet metal welding is leveraged for large, complex parts, reducing CNC machining requirements and cost.
- Assembly: The reduced part count and modular design facilitate rapid assembly and maintenance.
- Control: The CAN-based architecture and Jetson Orin Nano onboard computer support real-time control and future expansion.
- Learning Environment: The RL training pipeline is fully open-source, supporting reproducibility and extension to other bipedal platforms.
Limitations and Future Directions
While MEVITA demonstrates robust walking, the policy is less stable during static standing and at high speeds, with occasional foot-ground interference during lateral motion. Further work is needed on policy refinement, Real-to-Sim transfer, and hardware-in-the-loop adaptation to improve stability and expand behavioral repertoire. The open-source nature of MEVITA provides a foundation for community-driven research in physical intelligence, hardware optimization, and learning-based control.
Conclusion
MEVITA represents a significant advancement in open-source bipedal robotics, combining manufacturability, robustness, and accessibility through e-commerce sourcing and sheet metal welding. The platform enables scalable, reproducible research in learning-based locomotion and physical intelligence, lowering the barrier to entry for hardware experimentation. The open-source release of all hardware, software, and training environments positions MEVITA as a reference architecture for future developments in bipedal robot design, control, and learning.
