Product Selection Guide | SO-ARM101 vs NexArm: Which Open-Source 6-Axis LeRobot Arm Should You Choose?

The rapid growth of Hugging Face’s LeRobot open-source ecosystem has shifted robot development from pure rule-based control toward data-driven learning. By collecting demonstration data through leader-follower teleoperation and training with algorithms such as ACT or Diffusion Policy, developers can teach a robot a task from human demonstrations instead of writing every motion by hand.

In practice, though, not every LeRobot arm is the same. Differences in structure, servo strength, vision, and expandability mean some platforms are better for getting started, while others are needed for more demanding research. Matching the arm to your budget, use case, and the complexity of the algorithms you want to run is the key to choosing the right platform.

Hiwonder’s SO-ARM101 and NexArm are both open-source 6-axis arms that natively support the Hugging Face LeRobot ecosystem. Both support leader-follower teleoperation, dual-view data collection, and end-to-end examples with ACT and Diffusion Policy. They are aimed at different levels of work.

SO-ARM101: Affordable LeRobot Entry Point

Core positioning: Learning LeRobot, robot data collection, and getting started with imitation learning.

Hardware: Community-inspired open-source structure with 6 intelligent encoded bus servos. Maximum servo torque is 30 kg·cm, end-effector payload is 200 g, and working radius is 400 mm. Dual-view data collection system: first-person camera (2 MP) and third-person camera (0.3 MP).

Algorithms: Forward and inverse kinematics; trapezoidal acceleration/deceleration control built into the servos.

Key advantage: Improves on the underpowered servos and jitter common in many early open-source arms, so you can quickly run the full loop—teleoperation → data collection → model training → policy execution—at very low cost. Suitable for university teaching, algorithm beginners, and learning the LeRobot ecosystem.

NexArm: Higher-End Embodied AI Open-Source Arm

Core positioning: Longer reach and higher payload for precise grasping, complex tasks, and embodied AI development.

Hardware: ESP32 + AT32 dual-chip architecture. 6-DOF arm with aviation-grade metal body and dual-output-shaft joints, 6 intelligent encoded bus servos, maximum stall torque 65 kg·cm, 500 mm working radius, and 500 g end-effector payload. Dual-view data collection system: first-person camera (2 MP) and third-person camera (0.3 MP). Modular design that can be expanded with a K230 vision module, hanging Mecanum chassis, electric conveyor, electric slide rail, and other peripherals.

Algorithms: Forward and inverse kinematics; trapezoidal acceleration/deceleration control on the servos.

Key advantage: Fully supports the basic LeRobot learning flow, and its stronger structure and motion performance make fine manipulation, continuous action sequences, and policy validation more practical. The expandable peripherals also open more research directions and multimodal AI use cases, making it a strong option for university labs, robot developers, and embodied AI researchers.

Structure and materials

SO-ARM101 uses an optimized Hugging Face community open-source structure. It is light and compact, balancing cost and practicality for desktop learning and experiments. NexArm uses an aviation-grade metal shell with dual-output-shaft joints, a parallel-guide gripper, and a dual-controller architecture. The higher rigidity and stability make it more durable under longer, higher-load grasping.

Payload and working range

These directly affect the tasks you can run. SO-ARM101’s ~200 g payload and 400 mm radius are better for small-object grasping, action data collection, and imitation learning experiments. NexArm’s ~500 g payload and 500 mm radius give more workspace and carrying capacity for larger objects, continuous action tasks, and more complex robot applications.

Magnetic encoder bus servos

Servo choice affects motion accuracy, response, and load capacity. SO-ARM101 uses 6 × HX-10HM on the leader arm and 6 × HX-30HM on the follower arm; maximum stall torque on the follower is 30 kg·cm, suitable for lighter manipulation. NexArm uses 5 × HX-10HM + 1 × HX-12HM on the leader arm, and 1 × HX-65HM + 3 × HX-30HM + 2 × HX-12HM on the follower arm. Maximum servo torque reaches 65 kg·cm, which improves grasp stability and the ability to execute more complex motions.

Multimodal peripheral expansion

Unlike the desktop-focused SO-ARM101, NexArm can be expanded with a hanging Mecanum chassis, electric slide rail, electric conveyor, and other peripherals. This expands the workspace and motion range and opens more scenarios for LeRobot algorithm validation and embodied AI research.

Q1: I’m a complete beginner with LeRobot. Which one should I pick?

A: SO-ARM101. It offers strong value and deep compatibility with the official LeRobot ecosystem. It helps you quickly understand leader-follower teleoperation, data collection, dataset building, and ACT imitation learning training—an excellent entry point for embodied AI development.

Q2: What if I don’t know how to assemble a LeRobot arm?

A: NexArm ships largely pre-assembled; you mainly complete simple wiring. SO-ARM101 is a DIY kit, but Hiwonder provides assembly tutorials so you can learn the structure while building it and get started quickly.

Q3: I want to learn ACT, VLA, and other embodied AI algorithms. Which should I choose?

A: Both can serve as learning platforms. If your focus is LeRobot data collection and imitation learning workflows, SO-ARM101 is enough. If you also want finer manipulation, more complex task execution, and embodied AI application research, NexArm is the better choice.

Q4: How does the leader-follower teleoperation experience differ?

A: NexArm’s industrial-grade metal structure and stronger servo set (max 65 kg·cm torque) give better leader-follower sync, motion stability, and load capacity. SO-ARM101 is lighter, with a 30 kg·cm max torque on the follower arm, and is better suited to desktop teaching and light grasping.

Q5: How is a LeRobot arm different from a traditional ROS arm?

A: Traditional ROS arms rely on inverse kinematics and hand-written control logic. LeRobot arms are built around data-driven imitation learning: you collect human demonstrations via leader-follower teleoperation, then train end-to-end policies (ACT, Diffusion Policy, etc.) so the robot learns the task. They are complementary rather than replacements. LeRobot is especially useful for exploring robot learning, embodied AI, and end-to-end policy development.

Q6: What is the practical value of NexArm’s modular design?

A: NexArm supports expansion with an electric slide rail, mobile chassis, electric conveyor, and other modules. These broaden data-collection and task-validation scenarios. The slide rail extends workspace, the mobile chassis enables mobile grasping experiments, and the conveyor can simulate dynamic target manipulation—useful for more complex robot operation and embodied AI research.

Q7: How is Hiwonder’s SO-ARM101 different from other SO-ARM101 arms on the market?

A: While fully following the LeRobot open-source protocol, Hiwonder optimizes the hardware so motion is more precise, smoother, and more durable. It uses six 30 kg·cm torque servos plus trapezoidal acceleration/deceleration control for fluid movement, and a dual-camera system: a 2 MP first-person camera for fine motion capture and a 0.3 MP third-person camera for global awareness, supporting higher-quality data for embodied AI algorithms.