SO-ARM 102 Open-Source Robotic Arm Released: 2.5x Payload Boost, Fully Compatible with LeRobot Ecosystem

The open-source robotic arm space has a major update. The latest generation of the SO-ARM series, SO-ARM 102, has officially launched. The project was reviewed and approved by Rob Knight, with the GitHub repository now publicly available. As the direct successor to SO-ARM 101, version 102 brings systematic upgrades in mechanical structure, materials, and motion performance, while maintaining full compatibility with the Hugging Face LeRobot ecosystem.

Structural Design: From Gripper to Topology Optimization

The most visible change in SO-ARM 102 is the gripper. The newly designed gripper jaws maintain parallel motion throughout the entire stroke, rather than letting objects roll during closure as traditional designs do. This means noticeably improved positioning accuracy and stability when grasping cylindrical or irregularly shaped objects.

A deeper change comes from the structure itself. SO-ARM 102 adopts a topology-optimized design—the lattice structures seen in the video are a product of this approach. Topology optimization uses algorithms to find the optimal material distribution under given load conditions, removing redundant material and reinforcing critical load-bearing areas, thereby reducing weight while improving overall stiffness.

For material selection, the official recommendation is PET-CF (carbon fiber reinforced PET) for the best stiffness performance, but PLA has also been confirmed to print well. This material compatibility matters for community users: PLA is the most widely available desktop 3D printing material, while PET-CF requires higher printing temperatures and stricter process control. Users can flexibly choose based on their equipment and performance needs.

Rotation and folding angles have also been widened compared to the 101, giving the arm a more flexible range of motion. Feetech 3250 servos are used at the critical joints—the positions with the highest demands for precision and torque output.

Topology Optimization is a numerical computation-based structural design method. Its core idea is to automatically solve for the optimal material distribution within a given design space, load conditions, and constraints (such as material usage and stress limits) through iterative finite element analysis and optimization algorithms. Unlike traditional sizing or shape optimization, topology optimization can determine the overall structural skeleton at the conceptual design stage, often generating biomimetic, organic lattice or truss forms. It is an ideal partner for additive manufacturing (3D printing), because the printing process is not constrained by traditional subtractive manufacturing and can directly realize complex topology results. Adopting this method in an open-source robotic arm means designers can remove inefficient material along specific load paths, significantly reducing self-weight while maintaining stiffness—thereby indirectly improving payload capacity and motion performance.

PET-CF (carbon fiber reinforced PET) is a composite printing material formed by mixing chopped or continuous carbon fibers into a polyethylene terephthalate (PET) matrix. The addition of carbon fibers significantly improves the matrix's tensile strength, flexural modulus, and heat resistance while reducing the coefficient of thermal expansion, making it commonly used for mechanical structural parts requiring high stiffness and low creep. Its printing difficulty is also higher than ordinary PLA: it typically requires a nozzle temperature of 260–300°C, a heated chamber, and a hardened nozzle, otherwise poor interlayer bonding or clogs can occur. By contrast, PLA prints at lower temperatures, has less warping and better surface quality, but is prone to creep under long-term loads. SO-ARM 102 confirming both materials are usable essentially hands the choice between "performance ceiling" and "entry barrier" to users based on their equipment conditions.

Performance Data: Comprehensive Quantitative Improvements

Compared to SO-ARM 101, SO-ARM 102's core performance metrics have improved as follows:

• Payload capacity: 2.5x increase
• Positioning accuracy: approximately 2x increase
• Motion speed: approximately 1.6x increase
• Working range: approximately 36 mm additional reach

These figures cover the most critical performance dimensions of a robotic arm. A 2.5x payload increase means it can operate heavier end effectors or grip heavier objects; doubled positioning accuracy is especially critical for fine manipulation tasks (such as assembly, insertion/removal); and increased speed directly impacts task execution efficiency.

The extra 36 mm of working range may not seem like much, but in desktop-level robotic arm applications, it often determines whether the arm can reach target positions at the edge of its workspace.

Feetech 3250 servos belong to Feetech's serial bus digital servo series, using TTL half-duplex bus communication. Multiple servos can be cascaded and can read back feedback information such as position, temperature, and load. Their rated stall torque is approximately 20–30 kg·cm (depending on the specific model and voltage), with metal gears combined with magnetic encoder or potentiometer feedback, providing better positioning accuracy and consistency than ordinary analog servos. In SO-ARM 102, they are used at the critical joints with the highest precision and torque requirements. These joints must both withstand the leverage torque from end loads and ensure repeatable positioning accuracy, so servo selection directly determines the arm's performance ceiling. Bus-based connections also simplify multi-joint wiring, and when paired with LeRobot's software control stack, commands can be read and sent in batches through a unified protocol.

Ecosystem Compatibility: Zero Migration Cost

SO-ARM 102 continues its deep integration with Hugging Face LeRobot. The official statement is clear: with the same toolchain, testing framework, and workflow, the entire setup existing SO-ARM 101 users already have will work from day one.

The value of this strategy lies in lowering the upgrade barrier. For users who have already built data collection pipelines, training workflows, and deployment solutions in the LeRobot ecosystem, changing hardware requires no rewriting of anything at the software layer. The fragmentation problem that open-source robotics projects often face—where each new version brings incompatible APIs and toolchains—does not appear in the SO-ARM series.

The project code is hosted in the roboninecom/SO-ARM-102 repository on GitHub, released as open source, allowing the community to freely view design files, submit issues, and contribute improvements.

Hugging Face LeRobot is an open-source library for real-world robotics, providing dataset formats, imitation learning and reinforcement learning algorithms (such as ACT, Diffusion Policy, TD-MPC, etc.), training scripts, and hardware driver templates. Its core value lies in unifying the interfaces for "data collection—model training—real-world deployment," enabling low-cost robots of different form factors to share the same toolchain and pre-trained resources. As one of the robotic arm platforms officially supported by LeRobot, the SO-ARM series' joint configurations, camera calibration, and teleoperation workflows have all been incorporated into LeRobot's configuration system. Therefore, SO-ARM 102 maintaining the same protocols and interfaces means that demonstration data, training checkpoints, and deployment scripts accumulated by 101 users can be seamlessly reused—this is the technical basis for the "zero migration cost" claim.

Significance for the Open-Source Robotics Community

The iteration pace of the SO-ARM series reflects a trend: low-cost open-source robotic arms are rapidly approaching performance levels that previously only industrial-grade products possessed. Topology optimization, carbon fiber materials, high-precision servos—these elements appeared mainly in high-end research equipment just a few years ago, and are now integrated into an open-source project that the community can print and assemble themselves.

The binding with the LeRobot ecosystem solves another key problem—software stack continuity. If hardware performance improvements come with a broken software ecosystem, their practical value is greatly diminished. SO-ARM 102 chose to upgrade hardware while maintaining compatibility, and this trade-off is worth noting.

For researchers, educators, or makers evaluating entry-level robotic arms, SO-ARM 102 offers a choice with significantly improved performance and extremely low migration cost.

One-line Summary

SO-ARM 102 open-source robotic arm released, with major performance gains over 101 and full LeRobot ecosystem compatibility.

Paragraph Summary

SO-ARM 102 is the latest generation of the open-source SO-ARM robotic arm series, reviewed by Rob Knight and open-sourced on GitHub. Compared to its predecessor 101, it brings systematic upgrades in gripper, structure, and materials: the gripper jaws maintain parallel motion throughout to prevent object rolling, the body uses topology optimization and lattice design, and PET-CF is recommended while PLA printing remains compatible. Performance gains include 2.5x payload, 2x positioning accuracy, 1.6x speed, and approximately 36 mm more working range, with Feetech 3250 servos at critical joints. It is fully compatible with the Hugging Face LeRobot ecosystem, allowing 101 users to migrate directly without modifying software configurations, offering researchers, educators, and makers a choice with significantly improved performance and extremely low migration cost.