CNY
Product Overview
SO-ARM101 is an open-source robotic arm deeply optimized from SO-ARM100. Revised cable routing and motor/gear pairings eliminate the joint cable-breakage problem, and the performance upgrades support real-time leader-follower following. This is the 7-DOF version: it adds a wrist roll (yaw) servo over the 6-axis model, giving the wrist far greater freedom of posture, more reachable points and multi-angle precision grasping. It adapts perfectly to the Hugging Face LeRobot toolkit and connects directly to PyTorch models and shared datasets, so imitation learning and reinforcement learning are easy to put into practice. A complete assembly tutorial and DIY kit are included — students, researchers and makers can all work with intelligent robotics for learning, research and creation.
What is Hugging Face LeRobot? LeRobot is the open-source robotics ecosystem from the world-renowned AI community Hugging Face. It combines a Python library with a model-sharing platform, packaging standardized tools, datasets and pretrained models so that frontier AI techniques such as imitation learning and reinforcement learning are within reach — developers, students and hobbyists can all get hands-on with robotics development.
What can it do? Assembly is simple and requires no complex debugging. Using the leader-follower structure, the leader arm performs a grasp or placement and the follower arm mirrors it in real time; a small amount of training is enough for it to complete preset tasks independently. It connects directly to shared AI models and datasets with no need to write code from scratch, making it suitable for grasping and light assembly experiments. Low cost and low barrier make it an efficient route into embodied AI practice for students, researchers and makers.
Ships factory assembled via SF Express with free shipping. Invoices are available, and a KVM switch is included as a gift with finished-unit packages.
At a glance: 7DOF with 90° wrist roll · 12V high-torque servos at 30kg.cm · optional TPU flexible gripper · dual-view data collection · NVIDIA and D-Robotics RDK onboard inference · factory assembled and ready to use · supports ACT / SmolVLA / Pi0 / Pi0.5 / GR00T N1.5 model training.
High-precision magnetic encoding · End-to-end imitation learning · Hugging Face LeRobot · Teleoperation collaboration algorithms · Precise data collection · ACT model training · Arm-mounted camera vision · Scene camera perspective · Precise positioning with real-time feedback · 7DOF (7 degrees of freedom) · Leader-follower structure · 30KG high-torque bus servos · HD wide-angle camera · Supports secondary development · Textbook-grade tutorials.
Key Advantages
-
7 Degrees of Freedom with 90° Wrist Roll
The 7-DOF version adds a left/right wrist roll (yaw) servo to the 6-axis design, so the wrist can approach a target from far more angles and reach more points in the workspace. That extra freedom is what makes multi-angle, fine-grained grasping possible.
-
Leader-Follower Teleoperation & Imitation Learning
The arm integrates the Hugging Face AI framework. Teleoperate the leader arm to record demonstration motions, then train an imitation learning model in one pass and deploy the optimized policy. It can take on complex tasks and adapt to its environment, closing the automation loop end to end.
-
Dual-View Global Coverage
The SO-ARM101 arm-mounted camera captures the target's spatial position, angle and surface texture at close range for higher-fidelity data collection, while a desk-stand scene camera reads the workspace environment in real time. Together they keep operation precise, respond quickly to changes and prevent drift or stalling.
-
30kg.cm High-Torque 12V Bus Servos
The follower arm is unified at 7 × Feetech STS3215-C018 (12V, 30kg.cm, 1:345) to guarantee sufficient grasping torque under multi-axis load, while the leader arm stays at 7.4V to balance hand feel against cost. A 12-bit magnetic encoder delivers 0.088° precision on every axis.
-
Multi-Model Training Support
Train and deploy ACT, SmolVLA, Pi0, Pi0.5 and GR00T N1.5 policies on the same hardware, and reuse pretrained models such as lerobot/smolvla_base, lerobot/pi0_base, lerobot/pi05_base and lerobot/xvla-widowx directly from the LeRobot model hub.
-
Onboard Inference on NVIDIA and D-Robotics RDK
Connect a single cable to a Raspberry Pi, D-Robotics RDK or NVIDIA Jetson controller to run inference on the arm itself, with real-time motor control and encoder feedback.
-
Factory Assembled, Ready to Use Out of the Box
Every unit ships assembled, wired and calibrated — connect power and USB and the platform is ready for teleoperation and data collection.
Upgradeable Flexible Gripper
The flexible gripper is an upgrade of the standard rigid gripper, 3D printed in a flexible TPU material. It uses a hollow design with internal reinforcing ribs and works on a fin-type gripper principle: it conforms to the shape of the object being grasped, reducing the contact force applied to it. That makes the gripper ideal for complex shapes or objects with fragile components — fruit, for example — because it lowers the control precision needed to handle them safely.
How it works: on contact, the top and bottom of the fin-shaped jaws bend toward the object, so the gripper wraps around its surface. That increases the load-bearing contact area and produces a firm hold.
-
Ideal for: soft or easily damaged items that cannot be handled by a conventional rigid gripper — fruit, glassware, eggs and food processing.
-
Adaptive to shape: cylinders, spheres, ellipses and polygons — the full family of circle-based object shapes.