MEVIUS2: Practical Open-Source Quadruped Robot with Sheet Metal Welding and Multimodal Perception

Sign up for free to create and run prompts on this paper using GPT-5.

MEVIUS2: A simple explanation

What is this paper about?

This paper introduces MEVIUS2, a four‑legged “robot dog” that anyone can build and modify because all the designs and software are open and free to use. Unlike many hobby robots that are small and mostly plastic, MEVIUS2 is metal, about the size of Boston Dynamics’ Spot, and packed with sensors so it can understand its surroundings. The goal is to make a tough, practical, and affordable research robot that more people can build and improve.

What were the researchers trying to do?

In easy terms, they wanted to:

• Build a strong, larger, dog‑sized robot that doesn’t break easily.
• Make it open‑source so students and researchers can customize everything—from body parts to low‑level motor control.
• Give it “good eyes” and “good senses” (multiple sensors) so it can see and understand the world in detail.
• Train it to walk over rough ground, stairs, and slippery surfaces using modern AI training methods.
• Make the whole thing orderable online, like shopping for parts on regular e‑commerce sites.

How did they build and train it?

Think of building MEVIUS2 like making a sturdy metal backpack frame instead of a plastic toy:

• Strong body with sheet metal welding: They formed the robot’s body from flat metal plates that are welded together. It’s like folding and gluing cardboard into a tough box—except with metal and welding—so it’s strong but uses fewer parts.
• Metal parts you can order online: Most pieces are standard machined aluminum or welded sheet metal. The team designed every part so you can upload the files and order them from a manufacturing service (like buying custom parts online).
• Powerful legs with light lower sections: The knee is driven by a mechanism that lets the motor sit higher up. Imagine a bike where the power is transferred through linkages so the “lower leg” stays light—this helps the robot move more efficiently.
• Multimodal perception (multiple senses): MEVIUS2 uses two LiDARs and a high‑dynamic‑range camera.
  • LiDAR is like a laser ruler that scans the environment to make a 3D map (similar to how bats “see” with sound, but using light).
  • A high‑dynamic‑range camera helps the robot see both bright and dark areas clearly.
• Simple electronics layout: Twelve motors are controlled by a small onboard computer (an NVIDIA Jetson). There’s a separate battery for the motors and one for the computer, plus a wireless emergency stop for safety.
• Training with reinforcement learning: They taught the robot to walk in a simulator (like a realistic video game) using trial and error—this is called reinforcement learning. After the robot “learned” in simulation, the same walking policy was tested in another simulator and then run on the real robot. This speeds up training and keeps the real robot safe while learning.

What did they find, and why is it important?

Key results:

• It’s robust and practical: The metal body and welded structure made a larger, tougher robot that’s still relatively simple to assemble.
• It sees its environment well: With two LiDARs and a quality camera, the robot can build detailed maps of the ground around it.
• It can handle real‑world terrain: In tests, MEVIUS2 walked on concrete, grass, soil, steep slopes, indoor floors, and even slippery, wet stairs—without falling. It could climb human‑sized stairs, which smaller open‑source robots struggled with.
• It’s accessible and affordable for research: While commercial robots of this size are very expensive and closed‑source, MEVIUS2’s full design and code are released, and the total cost is around $13,000—much lower than commercial options of similar size.

Why that matters:

• A strong, sensor‑rich, open robot lets more schools and labs try advanced experiments (like outdoor navigation, mapping, or new leg control methods) without buying a closed, very expensive platform.
• Being able to order all parts online lowers the barrier for building a capable robot from scratch.

What could this mean for the future?

• More innovation: Because everything is open, students and researchers can tweak the design, add new sensors, try new AI methods, and share improvements.
• Better real‑world robot helpers: Stronger bodies and better “senses” mean robots can be more useful outdoors—climbing stairs, handling rough ground, or carrying equipment.
• Community growth: An affordable, practical robot platform encourages a community that builds, tests, and teaches with the same base robot.

The authors also note current limits—like no waterproofing yet, welding being harder for some builders, and the need for more sensors and safety work. But by releasing all hardware files, software, and training environments, they’ve provided a solid, ready‑to‑build foundation for others to push legged robotics further.

Knowledge gaps, limitations, and open questions

The following list identifies what remains missing, uncertain, or unexplored in the paper and supplementary materials, framed as concrete, actionable items for future work.

Mechanical design and manufacturing

• Absence of quantitative structural analysis (e.g., FEA, deflection, stress, factor-of-safety) for the welded A5052/A7075 chassis and links under dynamic loads, impacts, and worst-case postures.
• No fatigue, vibration, shock, or drop testing of welded joints and machined parts; long‑term durability of sheet‑metal welds under cyclic legged locomotion loads is unverified.
• Tolerance sensitivity and repeatability of the welded Base‑Link/Hip‑Link assemblies across different vendors and batches are not characterized; no jigs/fixtures, weld sequence, or post‑weld distortion control processes are provided.
• Repairability and maintainability of the single, highly integrated Base‑Link are not evaluated (e.g., ease and cost of replacing subcomponents vs. replacing the entire welded assembly).
• Lack of a waterproof/dustproof mechanical design and sealing strategy (gaskets, cable glands, IP rating targets, test protocols) despite stated importance for outdoor use.
• No assessment of foot material wear, abrasion, or performance vs. TPU hardness, tread geometry, or environmental conditions; no replaceability guidelines or field service intervals.
• Kinematic/workspace limits, singularities, and torque transmission characteristics of the parallel-link knee are not analyzed (e.g., joint torque vs. pose, effective inertia, distal mass trade‑offs).

Electronics, power, and thermal

• No power budget or energy efficiency metrics (e.g., average/peak power draw, cost of transport, Wh/km) across gaits and terrains; only a rough “~1 hour” runtime is reported.
• Thermal behavior of motors, drivers, and Jetson under sustained load is unmeasured (e.g., overheating thresholds, derating, ambient temperature effects, cooling requirements).
• Battery system details (BMS, cell chemistry, protections, pre‑charge, charging protocols, inrush control) and their safety validation are absent.
• Impact of using two LiDARs and an HDR camera on power, compute load, and runtime is not quantified; no real‑time budget or latency breakdown is provided.

Sensing and perception

• Sensor calibration and synchronization pipeline (camera/LiDAR/IMU extrinsics, time stamping, drift handling) is unspecified; no calibration accuracy or repeatability metrics.
• Elevation mapping accuracy, latency, and robustness are not benchmarked (e.g., map error vs. ground truth, performance under motion blur, foliage, rain, or low‑texture scenes).
• The role of the RGB camera in elevation mapping is unclear; no ablation to justify multimodal fusion vs. LiDAR-only, or to compare single vs. dual LiDAR coverage trade‑offs.
• Sensitivity to environmental conditions (rain, fog, snow, dust, direct sun HDR, LiDAR cross‑talk) is not evaluated; no mitigation strategies (filters, enclosures, sensor heaters) are presented.
• Proposed future sensors (RTK‑GNSS, thermal, foot contact) are not integrated or tested; no interface, mounting, or data fusion plan is specified.

Control, RL policy, and sim-to-real

• The locomotion controller’s inputs/outputs, control rates, and torque/impedance loops are not described; it is unclear whether motors are operated in torque, current, or position mode and how low‑level control is implemented.
• Sim‑to‑real methodology is not detailed (e.g., domain randomization parameters, actuator/latency models, contact/friction modeling, sensor noise); no ablation on what was necessary for transfer.
• No quantitative locomotion benchmarks (max speed, step height, slope angle, turning rate, success rate, recovery time) or comparisons against baselines (e.g., classical MPC, other RL policies).
• Absence of perceptive locomotion integration: elevation maps are demonstrated, but there is no evidence that perception informs foot placement, gait adaptation, or planning in closed loop.
• No evaluation of robustness to disturbances (external pushes, slips), sensor dropouts, or communication faults; fall detection and recovery strategies are not presented.
• Lack of policy generalization tests across payload variations, terrain parameter changes, or hardware modifications common in an open-source platform.

Experimental evaluation and benchmarking

• Terrain trials are qualitative with no standardized protocols (e.g., angles, dimensions, repetitions); success/failure rates, statistical variability, and confidence intervals are not reported.
• No endurance or long-duration field trials to establish mean time between failures (MTBF), component wear rates, or maintenance schedules.
• Gait diversity and dynamic agility (e.g., running, jumping, bounding) are not explored; only “walking” is demonstrated without speed or agility metrics.
• Payload capacity and its impact on performance (speed, energy, stability) are not measured despite mentioning available payload space.

Open-source reproducibility and scalability

• Reproducibility across different fabricators and regions (cost variance, lead times, alternative suppliers to MISUMI/meviy) is unaddressed; supply-chain risk and substitutions are not mapped.
• Build documentation (fixtures, welding instructions, QA checks, torque specs) is not provided in the paper; risk of build-to-build performance variability is not quantified.
• Software licensing and long‑term maintenance plans (issue tracking, versioning, CI, tested OS/driver versions) are unspecified; reproducibility may depend on unpinned dependencies.

Safety and compliance

• No formal risk assessment or functional safety architecture (SIL/PL classifications, failure modes and effects analysis, fault injection tests); emergency stop behavior and stop categories are unvalidated.
• Pinch/crush hazard mitigations (coverings, interlocks), power isolation, and safe-state strategies on communication loss are not defined or tested.
• EMC/EMI considerations and regulatory compliance for wireless and power systems are not discussed.

Design trade-offs and optimization

• Weight–strength–cost trade-offs of A7075/A5052 + welding vs. alternative manufacturing routes (e.g., castings, extrusions, modular machined frames, topology-optimized prints) are not compared.
• Sensor configuration optimization (number, placement, and types of LiDAR/cameras) is not explored relative to coverage, occlusion, cost, and compute constraints.
• No quantitative comparison to prior MEVIUS or other open-source platforms on identical standardized tasks to isolate the benefits of welding, larger size, or added sensors.

These gaps suggest clear next steps: define standardized tests and metrics; detail and validate sim‑to‑real and low‑level control; quantify structural, power, and perception performance; integrate perception into locomotion; formalize safety; and harden the design and build process for reproducibility and field use.

Immediate Applications

The following applications can be deployed now based on the paper’s open-source hardware/software, procurement workflow, and demonstrated locomotion and perception.

• Open, modifiable quadruped research platform (Academia, Robotics)
  • Use case: Labs can replicate and extend legged-locomotion, perception, and sim-to-real research on a Spot-scale, metal platform with accessible low-level control.
  • Tools/products/workflows: Provided STEP/URDF/MuJoCo models, LeggedGym environment, ROS/Rviz launch files, common Python deployment scripts; “click-to-order” metal parts via meviy part numbers; Jetson-based perception stack; elevation mapping pipeline.
  • Assumptions/dependencies: Access to MISUMI/meviy (or equivalent) for machining/welding; NVIDIA GPUs for IsaacGym; staff expertise in ROS, CAN, and safety; 1-hour battery runtime; operation in dry, dust-free environments; no formal risk assessment yet.
• Courseware and hands-on education at universities and makerspaces (Education, Academia)
  • Use case: Capstone courses, robotics clubs, and maker programs can build and operate a full-scale quadruped, compare policy learning vs. classical control, and teach safety engineering.
  • Tools/products/workflows: Predefined BOM with online quoting, “build labs” (mechanical assembly, CAN bus debugging, RL training), standardized assignments using provided simulation and perception pipelines.
  • Assumptions/dependencies: Budget (~$13k per unit), access to fabrication vendors, supervision for safe operation, campus facilities for testing.
• Rapid prototyping for indoor inspection and mapping pilots (Construction, Manufacturing, Facilities Management, Software)
  • Use case: Proofs-of-concept for factory/facility walkthroughs, warehouse aisle mapping, stair-inclusive route validation, and as-built floor mapping in dry indoor sites.
  • Tools/products/workflows: Out-of-the-box LiDAR + HDR camera, ROS-based elevation mapping, Jetson GPU processing, templated inspection “missions.”
  • Assumptions/dependencies: Non-waterproof/dustproof; requires safety cordons and E-stop SOPs; limited runtime; integration with facility CAD/BIM and data logging.
• Sensor and algorithm benchmarking on a full-scale legged platform (Robotics, Software/AI)
  • Use case: Evaluate LiDAR-camera fusion, elevation mapping, visual RL, foothold planners, and state estimation in realistic locomotion scenarios (stairs, slopes, slippery floors).
  • Tools/products/workflows: Swap-in perception modules via the Base-Link sensor mounts; compare policies trained in IsaacGym and validated in MuJoCo; dataset collection for multimodal perception.
  • Assumptions/dependencies: Reproducible terrains/test courses; synchronization and calibration of sensors; data governance for collected video/point clouds.
• Component/vendor evaluation and digital procurement workflows (Manufacturing, Supply Chain)
  • Use case: Demonstrate reduction of part count and lead time using sheet metal welding and online quoting; benchmark suppliers for small-batch robotic frames.
  • Tools/products/workflows: meviy auto-quoting from STEP, standardized part numbers, parallel-link leg design as a reference for DFM/DFA.
  • Assumptions/dependencies: Regional availability of online fabrication; welding QA; tolerance control for assemblies; shipping/import considerations.
• Security and telepresence pilots in controlled indoor environments (Security, IT/AV)
  • Use case: After-hours patrols on multi-level sites; remote tour/inspection with camera stream; event venue walkthroughs.
  • Tools/products/workflows: Add teleop UI over ROS, Wi-Fi/Private 5G streaming, logging; waypoint-based patrol scripts.
  • Assumptions/dependencies: Indoor-only, dry conditions; safety procedures; network coverage; privacy policies for video capture.
• Reproducible benchmarking and open tests for legged RL (Academia, AI)
  • Use case: Cross-lab benchmarks of locomotion policies on a standardized, Spot-scale platform; open datasets and leaderboards for rough-terrain performance.
  • Tools/products/workflows: Shared reward configs, policy release, evaluation scripts, terrain suites; community CI for sim-to-real regressions.
  • Assumptions/dependencies: Community governance, licensing clarity, and standardized reporting; reproducible hardware configs.
• Outreach and public demos (STEM engagement, Daily life)
  • Use case: Museum/university demos of legged locomotion and mapping; STEM outreach showcasing open-source mechatronics at human-scale.
  • Tools/products/workflows: Pre-scripted demos and safety playbooks; quick-swap batteries; transport cases.
  • Assumptions/dependencies: Supervised, controlled venues; insurance and risk mitigation.

Long-Term Applications

The following require additional engineering (e.g., IP ratings, certification), expanded sensing, and/or scaling to production.

• Outdoor industrial inspection and maintenance (Energy, Utilities, Oil & Gas, Construction)
  • Use case: Substation/solar/wind farm rounds, pipeline corridor checks, construction progress monitoring, slope/embankment assessment.
  • Tools/products/workflows: IP65+ sealing, RTK-GNSS, thermal cameras, gas sensors, foot contact/force sensing; fleet management, mission planning, hotspot detection analytics.
  • Assumptions/dependencies: Environmental hardening, safety certification (e.g., ISO 13482 for personal care robots/ISO 3691-4/IEC 61508 context-dependent), long-duration power (swappable packs/docks), robust comms (LTE/5G), operator training, site permissions.
• Disaster response and search-and-rescue (Public Safety)
  • Use case: Rubble navigation, stairwell reconnaissance, victim localization with thermal/audio sensors, post-flood/slippery environments.
  • Tools/products/workflows: Teleop/assisted autonomy, redundancy in sensing, protective enclosures, tethered power/comms options.
  • Assumptions/dependencies: Ruggedization, interoperability with incident command systems, liability coverage, extensive operator training.
• Healthcare logistics and hospital support (Healthcare)
  • Use case: Inter-ward transport, lab sample delivery over stairs, night-shift supply runs.
  • Tools/products/workflows: Hospital-grade safety features (people-detection, geofencing, bumpers), disinfection-friendly materials, EM interference compliance, fleet scheduling.
  • Assumptions/dependencies: Regulatory approval, infection control protocols, quiet/clean operation, reliability beyond 1-hour runtime.
• Agricultural field scouting on slopes and uneven terrain (Agriculture)
  • Use case: Vineyard/orchard scouting, crop stress mapping with multispectral/thermal, fence-line inspection.
  • Tools/products/workflows: Weatherproofing, extended autonomy, RTK-GNSS, crop-analytics pipeline, docking/charging in the field.
  • Assumptions/dependencies: IP rating, dust/mud resilience, seasonal maintenance, rural connectivity.
• Outdoor security patrol and smart-city surveying (Security, Public Sector)
  • Use case: Park/campus perimeter patrol; sidewalk-level asset surveys; ADA-related accessibility assessments (ramp/stair mapping).
  • Tools/products/workflows: Safety-rated perception, citizen interaction policies, encrypted teleop, integration with VMS/PSIM platforms; GIS-based reporting.
  • Assumptions/dependencies: Municipal permits, privacy compliance, insurance, noise/light pollution limits, public acceptance.
• Standardized open reference platform for national R&D and workforce programs (Policy, Academia, Industry Consortia)
  • Use case: Government-funded, open hardware baselines to reduce vendor lock-in; shared curricula and testbeds for AI/robotics upskilling.
  • Tools/products/workflows: Certified builds, long-term support branches, parts availability guarantees, aligned standards and evaluation protocols.
  • Assumptions/dependencies: Policy backing and funding, IP stewardship, supply continuity, governance of safety and data standards.
• Commercial developer kits and verticalized “app packs” (Robotics, Software)
  • Use case: MEVIUS2-based kits with vertical modules: “Inspection Pack,” “Ag Pack,” “Security Pack,” each bundling sensors, software, and workflows.
  • Tools/products/workflows: Modular sensor bays, hot-swappable batteries, app store for missions, cloud-based fleet management and analytics.
  • Assumptions/dependencies: Productization, documentation, customer support, warranty/QA, compliance testing.
• Foundation-model and multimodal learning on legged robots (AI/Software)
  • Use case: Vision-language-action policies for terrain-aware navigation and task following; self-supervised mapping from LiDAR+RGB at scale.
  • Tools/products/workflows: Data collection at fleet scale, on-device acceleration, sim-to-real with photorealistic rendering, continual learning pipelines.
  • Assumptions/dependencies: Compute budgets, data governance, safety guardrails for LLM-in-the-loop control, robust fallback controllers.
• Hybrid manipulation–locomotion platforms (Robotics, Manufacturing/Service)
  • Use case: Add a lightweight arm for door operation, valve turning, sample collection during inspection.
  • Tools/products/workflows: Mechatronic integration (mount points, power budget), whole-body control, safety interlocks, task libraries.
  • Assumptions/dependencies: Payload and stability margins, additional sensing, expanded safety certification scope.
• Digital manufacturing playbooks for low-volume robotics (Manufacturing, Supply Chain, Policy)
  • Use case: Codified “design-to-click-to-order” workflows (DFM for welded sheet metal, tolerance stacks, QA) to accelerate small-batch robot production.
  • Tools/products/workflows: Parametric CAD templates, automated quoting pipelines, vendor-agnostic part libraries, variant management.
  • Assumptions/dependencies: Cross-vendor interoperability, workforce training, quality standards, alternative processes where welding access is limited.

Cross-cutting assumptions and dependencies (impacting feasibility across long-term items):

• Environmental sealing (water/dust), formal risk assessment, and compliance to relevant service-robot safety standards are not yet implemented.
• Power: baseline ~1 hour runtime; many applications need battery swaps, larger packs, or docks.
• Supply chain: availability of specific motors (Robstride03), Livox LiDARs, Jetson modules; regional access to online fabrication.
• Legal/policy: site permissions, data/privacy compliance for imaging and mapping, and insurance.
• Talent: teams need multi-disciplinary expertise (mechanical, controls, perception, safety) to deploy and maintain fleets.

• A5052: An aluminum-magnesium alloy commonly used for sheet metal due to its formability and corrosion resistance. "made from A5052."
• A7075 aluminum alloy: A high-strength aluminum alloy (Al-Zn-Mg-Cu) often used where high strength-to-weight is needed. "11 parts are machined from A7075 aluminum alloy"
• Base-Link: The central structural body component that integrates mount points for legs, electronics, and sensors. "The Base-Link consists of five welded sheet metal plates"
• CAN-USB interface: A hardware adapter that bridges a Controller Area Network (CAN) bus to a computer’s USB port. "via two CAN-USB interfaces"
• Continuous torque: The torque a motor can sustain indefinitely without overheating or damage. "continuous torque of 20. Nm"
• Distal (leg segments): Refers to parts farther from the robot body; distal leg segments are the lower portions of the limb. "lighter distal leg segments."
• Elevation mapping: A grid-based 2.5D map that maintains local terrain height estimates from range sensors for locomotion and navigation. "For perception, elevation mapping [26] is employed"
• High dynamic range camera: A camera capable of capturing scenes with very bright and very dark regions without losing detail. "a high dynamic range camera"
• IMU: Inertial Measurement Unit; a sensor module that measures acceleration and angular rate (and sometimes magnetic field) for pose estimation. "LiDAR/IMU"
• IsaacGym: NVIDIA’s GPU-accelerated physics simulation platform for large-scale reinforcement learning. "reinforcement learning conducted in IsaacGym [25]."
• LeggedGym: A reinforcement learning environment tailored for legged robots built on top of Isaac Gym. "based on LeggedGym"
• LiDAR: Light Detection and Ranging; a laser-based ranging sensor that produces precise 3D measurements of the environment. "LiDARs"
• LiPo battery: Lithium-polymer rechargeable battery known for high energy density and discharge rates. "two 24V, 3300mAh LiPo batteries are connected in series"
• meviy: MISUMI’s online machining and fabrication service that auto-quotes and allows ordering from CAD (STEP) files. "meviy [24], a machining and fabrication service by MISUMI."
• MuJoCo: A high-performance physics engine commonly used for robotics simulation and control validation. "validated on MuJoCo"
• Multimodal perception: Combining multiple sensor modalities (e.g., LiDAR and camera) to perceive the environment more robustly. "equipped with multimodal perception"
• NVIDIA Jetson: An embedded GPU computing platform for edge AI and robotics. "The onboard PC is a NVIDIA Jetson"
• Parallel-link mechanism: A linkage design that maintains parallelism between links to transmit motion efficiently (e.g., to a knee joint). "A parallel-link mechanism is employed"
• Peak torque: The maximum torque a motor can deliver for short durations. "peak torque of 60. Nm."
• Point cloud: A set of 3D points measured by sensors (like LiDAR) representing surfaces in the environment. "point cloud processing is performed"
• POTICON: A specialized engineering plastic referenced in the MEVIUS platform’s materials. "Metal / POTICON"
• Reinforcement learning: A machine learning paradigm where agents learn policies by maximizing cumulative reward through interaction. "thanks to reinforcement learning"
• RGB camera: A standard color camera that captures red, green, and blue channels. "an RGB camera"
• ROS: Robot Operating System; a middleware framework for building and integrating robot software. "ROS and Rviz"
• Rviz: ROS’s 3D visualization tool for sensor data and robot models. "Rviz visualization"
• RTK-GNSS: Real-Time Kinematic Global Navigation Satellite System; high-precision satellite positioning using carrier-phase corrections. "RTK-GNSS"
• Servo motor: An actuator with closed-loop control of position, speed, or torque, often via internal sensing and control electronics. "12 servo motors are controlled"
• Sheet metal welding: Joining sheet metal parts by welding to form strong, integrated structures. "sheet metal welding"
• Sim-to-Real: Transferring policies or controllers trained in simulation to real hardware. "Sim-to-Real deployment."
• Sim-to-Sim: Validating or transferring a controller across different simulators before deploying to hardware. "Sim-to- Sim simulation"
• STEP files: ISO 10303 neutral CAD exchange files used for sharing and ordering fabricated parts. "STEP files"
• TPU: Thermoplastic polyurethane; a flexible, abrasion-resistant polymer used for soft printed components. "3D printed in TPU"
• URDF: Unified Robot Description Format; an XML specification for representing robot models. "includes the URDF model of MEVIUS2."
• Wireless emergency stop: A remote, wireless safety mechanism to immediately cut power or halt motion. "wireless emergency stop"

Sign up for free to add this paper to one or more collections.