ThinkRobotics LeRobot SO-101 6-Axis Robotic Arm Review: AI-Ready, Open-Source, and Built for Learning
by Gaurav Sarraf
Most desktop robotic arms at accessible price points are designed for pick-and-place demonstrations or basic scripted motion. The SO-101 is designed for something different: teaching a robot to learn. It is the hardware half of an end-to-end AI robotics workflow that integrates directly with the Hugging Face LeRobot library, allowing users to collect motion data via teleoperation and train imitation-learning models that run on the arm itself.
At Think Robotics, the SO-101 is one of the most technically interesting products in the catalog because it bridges hardware assembly, servo programming, and applied machine learning in a single kit. This review covers the full specification, what the kit includes, how the leader-follower teleoperation system works, the learning workflow, who it is designed for, pricing in India, and how it compares to similar platforms.
What Is the SO-101?
The SO-101 is the second-generation version of the SO-100 arm, created by TheRobotStudio in collaboration with Hugging Face. The SO-101 improves on the SO-100 with better wiring routing that prevents disconnection at joint 3, easier assembly that does not require removing gears during the build, and updated motors for the leader arm with optimized gear ratios. The hardware design, 3D-printed structural files, bill of materials, and assembly guide are all publicly available under an open-source license.
DIY unassembled kit: The ThinkRobotics kit ships with 3D-printed structural parts, STS3215 bus servos, wiring, BusLinker controller board, power supplies, and table clamps. Assembly follows the official Hugging Face SO-101 documentation. A camera is not included and must be sourced separately.
6-axis / 6 DOFSTS3215 bus servos30 kg.cm torque @ 12V360° magnetic encoder per servoBusLinker V3.0 controllerHugging Face LeRobot compatibleOpen-source hardware and softwareRs. 27,999
The Leader-Follower Concept
This is the core idea that makes the SO-101 different from a standard programmable robotic arm. The typical setup consists of two arms: a leader arm controlled by a human operator, and a follower arm that mirrors the leader's movements in real time.
Operated by human
Leader Arm
Moved by hand to demonstrate tasks
Operator moves the leader arm through a task manually
Different gear ratios per joint for comfortable human control
3x STS3215 with 1/147 ratio, 2x with 1/191, 1x with 1/345
Sends positional signals to the follower arm over USB
Records and learns
Follower Arm
Replicates leader in real time and trains
Mirrors the leader arm's movements in real time
All 6x STS3215 servos use uniform 1/345 gear ratio
Joint states and camera frames are captured as a training dataset
Once trained, executes tasks autonomously without human input
As the operator moves the leader arm to perform a task, the follower arm replicates the motion while cameras capture both arm movements and the task environment. This captured motion data becomes the training dataset for an imitation learning model. Once the model is trained on sufficient demonstrations, the follower arm can autonomously execute the task without further human input.
Key Specifications
Specification
Detail
Degrees of Freedom
6 (6-axis)
Follower Arm Motors
6x STS3215 bus servos, 1/345 gear ratio
Leader Arm Motors
3x STS3215 with 1/147 ratio, 2x with 1/191 ratio, 1x with 1/345 ratio
Follower Servo Torque (12V)
30 kg.cm
Follower Servo Torque (7.4V)
16.5 kg.cm
Encoder
360-degree magnetic encoder per servo
Controller Board
BusLinker V3.0 (Hiwonder) or Waveshare Bus Servo Adapter
Power Supply
5V 6A x2 included (12V version optional for higher torque)
Structural Parts
3D-printed (photosensitive resin or PLA)
Communication
USB to PC (Windows / Ubuntu / macOS)
Software
Hugging Face LeRobot (Python, open-source)
Camera
Not included (USB webcam 720p+ required)
Table Clamps
4x included
Kit Type
Unassembled DIY
The STS3215 Bus Servo
The STS3215 is a programmable bus servo from Feetech. Unlike standard hobby servos that take a PWM signal and have no position feedback, the STS3215 communicates over a half-duplex serial bus, reports its current position through a 360-degree magnetic encoder, and can be daisy-chained with other servos on a single data line.
Standard hobby servo
InterfacePWM signal
Position feedbackNone
Encoder typePotentiometer
Daisy-chainNot supported
Wiring per arm6 separate signal wires
STS3215 (used in SO-101)
InterfaceHalf-duplex serial bus
Position feedbackReal-time, continuous
Encoder type360° magnetic encoder
Daisy-chainYes, single data line
Wiring per arm1 shared bus per chain
The magnetic encoder provides positional accuracy significantly better than the potentiometer-based feedback found in consumer hobby servos. This accuracy matters for imitation learning because small positional errors during data collection translate directly into training noise, which degrades model performance. At a 12V supply, the follower arm servos produce 30 kg.cm of torque, sufficient for typical tabletop manipulation tasks: picking up small objects, moving blocks, or handling lightweight tools.
Software: The Hugging Face LeRobot Library
LeRobot is a Python library maintained by Hugging Face that provides the complete software stack for the SO-101 workflow. It handles servo configuration, motor ID assignment, teleoperation data collection, dataset creation, model training, and autonomous policy inference, all through a command-line interface.
1
Motor Setup
Configure motor IDs and baud rates on each servo. Each STS3215 servo gets a unique address on the bus. This is done using the LeRobot setup command, which auto-assigns IDs as you connect servos one at a time.
Both arms are calibrated by moving each joint through its full range of motion while the library records the endpoint positions. This establishes the zero positions and joint limits used during teleoperation and inference.
The operator moves the leader arm to demonstrate tasks while the library records the follower arm's joint states and camera frames as a dataset. Each teleoperation session produces one or more episodes. Datasets are stored locally and can be pushed to the Hugging Face Hub.
A policy model (e.g. ACT or Diffusion Policy) is trained on the collected dataset using the library's training scripts. The resulting model then runs inference on the follower arm for autonomous task execution without further human input. An NVIDIA GPU speeds up training, but CPU-only training is possible.
Community datasets on the Hub: All datasets and trained models can be shared on the Hugging Face Hub. Community members have published datasets from their SO-101 deployments, allowing a new user to download existing training data and train a model without performing teleoperation as a starting point. This collaborative sharing is a key advantage of the platform.
Use Cases
🔬
Robotics Research and Embodied AI
One of the few low-cost platforms where imitation learning and reinforcement learning policies can be trained, tested, and iterated on physical hardware without access to a research lab. For university students, independent researchers, and AI hobbyists working on manipulation tasks.
🎓
AI and Robotics Courses
Educators teaching applied machine learning, robotics programming, or human-robot interaction can use the SO-101 as a hands-on lab platform. The official Hugging Face course materials provide students with a structured learning path from assembly to model training.
⚗️
Algorithm Prototyping
For developers building new imitation-learning or reinforcement-learning algorithms, the SO-101 provides a physical test bench far cheaper than commercial manipulators. The standard Python library makes it straightforward to swap in custom training pipelines.
🛠️
DIY AI Robotics Projects
Hobbyists who want to go beyond scripted robot motions and explore policy-based control have in the SO-101 a platform specifically designed for this. The Hugging Face Discord community provides active support for assembly, calibration, and model training troubleshooting.
What the Kit Includes
The ThinkRobotics SO-101 kit is the unassembled DIY version. For a complete leader-follower teleoperation setup, two arm kits are needed: one configured as the leader and one as the follower.
Included in the kit
3D-printed structural parts for one arm
STS3215 bus servo motors (6 per arm)
BusLinker V3.0 controller board
Power supply: 5V 6A x2
Table clamps x4
Wiring and cables
Camera (not included, must be sourced separately)
Second arm (two kits needed for full teleoperation)
Camera required for full workflow: The dual-camera setup referenced in the Hugging Face documentation (one camera mounted at the wrist, one external overview camera) must be sourced separately. Standard USB webcams at 720p or higher resolution are sufficient and are used in the official tutorials. Identical camera models can cause USB path confusion on some systems; use two different camera models or manage USB paths carefully.
Pricing and Availability
The ThinkRobotics LeRobot SO-101 is priced at Rs. 27,999 as a DIY unassembled kit. This price covers one arm (leader or follower configuration). For a complete leader-follower teleoperation pair, two kits are required.
For reference, assembled kits from international vendors such as WowRobo and PartaBot ship at approximately $200 to $400 USD for a complete leader-follower pair with a camera. The DIY kit from Think Robotics provides the same hardware in India with local availability, local support, and no import wait times.
What Users Are Saying
Feedback from the Hugging Face community blog, Hiwonder product listings, Hackster.io, and CNX Software covers both the genuine strengths of the SO-101 platform and the practical challenges real users have encountered.
This SO-ARM101 kit is fantastic! Super clear instructions, quality components, and it integrates perfectly with LeRobot. Got my imitation learning prototype up and running over a weekend.
PositiveHiwonder product listing, verified buyer
The arm can learn to place a Lego brick into a container after a single demonstration. According to Hugging Face, the arm can master such tasks within minutes using standard consumer computers. The leader-arm-follows-follower feature is also a major improvement for reinforcement learning workflows.
PositiveTech Funding News, SO-101 launch coverage, May 2025
The training script was pretty easy to set up, thanks to the training and logging pipeline provided by LeRobot. I kicked off training before going to sleep. My ACT model (52M parameters) trained in about 4 hours on a 12GB NVIDIA RTX 3080.
PositiveSherry Chen, Hugging Face blog post, Sep 2025
The SO-101 improves the wiring to prevent disconnection issues previously seen at joint 3 and adds motors with optimized gear ratios. The leader arm can now follow the follower arm in real-time, which is useful for reinforcement learning where a human can intervene and correct robot actions.
Improvement notedCNX Software, SO-ARM101 coverage, May 2025
The data collection itself was unexpectedly time-consuming. My cameras would disconnect every 5 episodes and crash the recorder script. After investigation, I realized it was because my two webcams were identical to each other, which confused the computer and caused USB paths to be reassigned randomly.
Setup challengeSherry Chen, Hugging Face blog, Sep 2025
My SO-101 didn't learn to pick up the block after the first training run. After investigation, I realized I was cheating during data collection by looking over my shoulder at the follower arm while recording, giving myself information the robot would never have from camera footage alone. The LeRobot tutorial actually warns about this.
Training pitfallSherry Chen, Hugging Face blog, Sep 2025
How It Compares
SO-101 vs
SO-100 (Previous Gen)
The SO-100 is the predecessor on which the SO-101 is based. The SO-101 improves wiring management at joint 3, removes the gear disassembly step from the build, and uses updated leader arm motors with better gear ratios for easier manual control. The SO-100 is now considered deprecated in the official documentation.
Choose SO-101 for all new builds
SO-101 vs
Koch v1.1 Arm
The Koch v1.1 is another low-cost open-source arm in the LeRobot ecosystem, using Dynamixel servos instead of STS3215. Dynamixel servos are more widely used in academic robotics research and have a larger ecosystem of compatible hardware, but are significantly more expensive. The SO-101 offers a lower entry cost for the same LeRobot workflow.
SO-101 for lower cost access
SO-101 vs
Commercial Arms (myCobot, xArm)
Commercial arms come pre-assembled, include vendor-specific control software, and carry manufacturer warranties. They are designed for deployment and integration rather than learning and experimentation. The SO-101 is an open-source research platform whose design can be modified and extended. For learning, research, and algorithm development, the SO-101 is the lower-cost and more extensible choice.
SO-101 for research, commercial for deployment
Before You Buy
✅
Jetson-compatible for untethered AI. The LeRobot library runs on Ubuntu, and Jetson boards running JetPack with Ubuntu support the full stack. Running inference on a Jetson allows autonomous operation without a laptop, useful for untethered deployments.
✅
Open-source, modifiable, community-backed. The hardware design files, BOM, and software are all public. The Hugging Face community on Discord provides active support and shared datasets. You can download community datasets and train without starting from scratch.
✅
Local India availability. The ThinkRobotics kit ships domestically with local support. No import wait times or customs processing compared to international vendors at $200 to $400 USD for an assembled pair.
✅
CPU training is possible. An NVIDIA GPU on the host machine speeds up model training considerably, but CPU-only training is possible for smaller datasets. The ACT model (52M parameters) trains in approximately 4 hours on an RTX 3080.
⚠️
Two kits needed for teleoperation. A single kit is sufficient for scripted motion programming and servo debugging only. The full leader-follower data collection workflow requires two arm kits: one as leader, one as follower.
⚠️
Camera not included. Any USB webcam at 720p or higher works, but avoid using two identical camera models. Identical models can cause USB path reassignment that crashes the data recorder. Use different camera brands or models for the wrist and overview positions.
⚠️
Assembly takes 3 to 6 hours. This is an unassembled DIY kit. Motor ID configuration and calibration add time before the arm is operational. First-time builders should follow the official Hugging Face SO-101 documentation carefully.
⚠️
Data collection quality matters enormously. Looking at the follower arm while recording teleoperation data introduces information the robot will not have from cameras alone. This causes model training to fail. Follow the LeRobot tutorial guidance on proper data collection posture.
Conclusion
The Most Accessible Entry Point into Physical Imitation Learning in India
The ThinkRobotics LeRobot SO-101 is the most accessible entry point into physical imitation learning robotics currently available in India. The combination of STS3215 bus servos with magnetic encoder feedback, a fully open-source hardware design, and direct compatibility with the Hugging Face LeRobot Python library means a developer, student, or researcher can go from an unbuilt kit to a trained autonomous manipulation policy in a single weekend.
It is not a plug-and-use device. Assembly, motor configuration, calibration, and model training all require time and technical effort. For buyers willing to invest that effort, the SO-101 provides a hands-on platform for embodied AI research that previously required either a research lab or a budget far beyond what most individuals or student teams can access. Find it at Think Robotics with local support and availability across India.
6-axis, 6 DOFSTS3215 magnetic encoder servosHugging Face LeRobot nativeImitation learning readyFully open-sourceJetson-compatible
Frequently Asked Questions
No. The camera is not included and must be sourced separately. Any USB webcam at 720p or higher works with the official Hugging Face tutorials. Avoid using two identical camera models for the wrist and overview positions, as this can cause USB path conflicts that crash the data recorder. Use two cameras from different manufacturers to prevent this.
Yes. The leader-follower workflow requires two arms. One is configured as the leader (moved by hand) and one as the follower (mirrors and records motion). The motor configurations differ between the two: the leader arm uses three different gear ratios across its six joints, while the follower arm uses a uniform 1/345 ratio. A single kit is sufficient for scripted motion programming and servo testing only.
Any laptop or desktop running Ubuntu, macOS, or Windows with Python 3.10 or higher. An NVIDIA GPU speeds up model training considerably: a 52M parameter ACT model trains in approximately 4 hours on a 12GB RTX 3080. CPU-only training is possible for smaller datasets but is significantly slower. For production deployment without a tethered laptop, the LeRobot library runs on NVIDIA Jetson boards running JetPack with Ubuntu, enabling autonomous untethered operation.
Assembly typically takes 3 to 6 hours for a first-time builder following the official guide. Motor ID configuration and calibration add additional time before the arm is operational. The SO-101 improves on the SO-100 by eliminating the gear disassembly step, making the build process more straightforward. Video references in the Hugging Face documentation cover each assembly stage.
Yes. The LeRobot library runs on Ubuntu, and Jetson boards running JetPack with Ubuntu support the full stack. Running inference on a Jetson allows autonomous operation without a laptop, which is useful for untethered deployments. The SO-ARM101 was specifically designed to interface with NVIDIA Jetson AI modules, making it a natural pairing for edge AI robotics applications.
Get the ThinkRobotics LeRobot SO-101 in India
Open-source 6-axis arm for imitation learning and embodied AI research. Local support, fast delivery, and no import wait times. Rs. 27,999 per kit.
Please note that although my status in your website shows fulfilled, my components haven't arrived yet. I am extremely disappointed by the poor coordination between your team and the delivery team.
Three of the four ports stopped working after usage for sometime. After more usage using the fourth port only, the fourth port also stopped working in 3 days. Very bad experience. Please return or replace.
