Koch v1.1

Home > Hardware > Arms > Koch

Overview

Koch v1.1 is an improved version of the low-cost open-source robot arm designed by Alexander Koch and refined by HuggingFace engineer Jess Moss. For approximately $250-$480, you can build a Leader-Follower teleoperation system with complete integration with the LeRobot framework.

Item	Details
Original Designer	Alexander Koch (Co-founder of Tau Robotics)
v1.1 Improvements	Jess Moss (HuggingFace)
License	Open Source (GitHub public)
Type	6-DoF low-cost 3D printed robot arm
DoF	6 (Shoulder Pan, Shoulder Lift, Elbow, Wrist Flex, Wrist Roll, Gripper)
Motors	Dynamixel XL430-W250-T + XL330-M288-T
Parts Cost	~$258 (Follower) / ~$183 (Leader) / ~$441 (pair)
Release Date	Original: Early 2024, v1.1: July 2024

Key Significance

Low-Cost Data Collection Platform for VLA Research

The most important contribution of the Koch robot arm is dramatically lowering the data collection cost for VLA (Vision-Language-Action) research.

Ultra-low-cost Hardware: 1/100 the cost compared to existing research robot arms (Franka $30K+, UR $25K+)
Fully Open Source: CAD files, SolidWorks models, BOM, and control code all publicly available
3D Printing Based: Manufacturable with standard FDM printers (Prusa, Ender, Bambu)
LeRobot Integration: Perfect compatibility with HuggingFace’s robotics ML framework

Tau Robotics’ Vision

Alexander Koch presented the vision of “millions of robot arms learning in the real world” through Tau Robotics. This low-cost robot arm is the core hardware for that vision, enabling distributed data collection and large-scale imitation learning.

“We are building a general AI for robots. We start by building millions of robot arms that learn in the real world.”

Alexander Koch, Tau Robotics

Technical Specifications

Motor Configuration

Follower Arm (Controlled Target)

Joint	Motor	Quantity	Torque	Features
Shoulder Pan	XL430-W250-T	1	1.4 N.m	High torque, base rotation
Shoulder Lift	XL430-W250-T	1	1.4 N.m	High torque, shoulder lift
Elbow Flex	XL330-M288-T	1	0.52 N.m	Lightweight (18g)
Wrist Flex	XL330-M288-T	1	0.52 N.m	Lightweight (18g)
Wrist Roll	XL330-M288-T	1	0.52 N.m	Lightweight (18g)
Gripper	XL330-M288-T	1	0.52 N.m	Lightweight (18g)

Design Philosophy: XL430 motors are about 2x more powerful than XL330, used for shoulder joints supporting heavy loads. XL330s are ultra-lightweight at 18g to minimize arm end inertia.

Leader Arm (Teleoperation Controller)

Joint	Motor	Quantity	Features
All joints	XL330-M077-T	6	5V unified, simplified assembly

The Leader arm is configured with a handle and trigger instead of a gripper for intuitive control by human operators.

Dynamixel Motor Details

Model	Voltage	Stall Torque	Weight	Price
XL430-W250-T	11.1V	1.4 N.m	57.2g	~$50 ($27.50 on sale)
XL330-M288-T	5.0V	0.52 N.m	18g	~$24
XL330-M077-T	5.0V	0.34 N.m	18g	~$24

Bill of Materials (BOM)

Follower Arm (Controlled Target)

Part	Quantity	Price (USD)	Notes
Dynamixel XL430-W250-T	2	$100	Shoulder joints (high torque)
Dynamixel XL330-M288-T	4	$96	Elbow, wrist, gripper
XL330 Frame & Idler Wheel Set	1	$10	4pc set, use 2-3
XL430 Idler Wheel Set	1	$7
Waveshare Serial Bus Servo Driver	1	$10	USB-C connection
12V to 5V Voltage Reducer	1	$10	Voltage conversion for XL330
12V Power Supply	1	$12	Main power
Table Clamp	1	$6	Mounting
Wires/Connectors	-	$7	Wiring
Total		~$258	3D printing cost separate

Leader Arm (Teleoperation Controller)

Part	Quantity	Price (USD)	Notes
Dynamixel XL330-M077-T	6	$144	Unified for all joints
XL330 Frame	1	$7
XL330 Idler Wheel Set	1	$10
Waveshare Serial Bus Servo Driver	1	$10
5V Power Supply	1	$6	Leader uses 5V only
Table Clamp	1	$6
Total		~$183	3D printing cost separate

Total System Cost

Configuration	Cost	Use Case
Follower only	~$258	Autonomous control research
Leader + Follower	~$441	Teleoperation data collection
Dual arm system	~$882	Bimanual manipulation

Where to Buy

ROBOTIS: Koch v1.1 Leader Kit, Follower Kit
PartaBot: Complete products and parts kits
Tau Robotics: Waiting list - One-stop package

Tip: ROBOTIS official shop frequently offers 10% discount codes.

v1.0 vs v1.1 Comparison

Item	v1.0 (Original)	v1.1 (Improved)
Designer	Alexander Koch	Jess Moss (HuggingFace)
Assembly Difficulty	Medium (soldering required)	Low (no soldering required)
Voltage Converter	Manual adjustment needed	Pre-set DC converter
Screw Interference	Some present	Fixed
Unnecessary Materials	Present	Removed
Hole Sizes	Non-standard	Standardized
Plastic Screws	Used	Removed (replaced with metal screws)
Leader Board Platform	None	Added
CAD Format	STL only	STL + SolidWorks (easier community contribution)

Key Improvements in v1.1

No Soldering: DC converter replacement improves accessibility for general users
Standardization: Unified hole sizes and screw specifications reduce assembly errors
SolidWorks Models: Easier community modification and contribution
LeRobot Optimization: ML workflow optimization by HuggingFace engineers

Assembly and Build

3D Printing Requirements

Item	Recommended Specs
Material	PLA+, ABS, PETG (reasonably strong plastic)
Nozzle	0.4mm
Layer Height	0.2mm
Infill	~30%
Support	Gripper only

Recommended Printers: Prusa Mini+, Bambu P1, Creality Ender 3

Assembly Process

3D Printing: Print all structural parts with STL files
Motor Setup: Set unique IDs (1-6) and 1M baudrate for each motor with Dynamixel Wizard
Voltage Converter Connection: 12V → 5V conversion (for XL330 on Follower)
Serial Bus Connection: Connect motors in daisy chain
Board Mounting: Mount controller board to base
Calibration: Set joint ranges with LeRobot

Power Configuration

Arm	Voltage	Reason
Leader	5V	All XL330-M077 motors use 5V
Follower	12V + 5V	XL430 uses 12V, XL330 uses 5V (with converter)

Warning: Applying 12V to Leader risks motor damage

LeRobot Ecosystem Integration

Installation

There are two ways to install LeRobot:

Method 1: Source Installation (Recommended - for development/contribution)

# Clone repository
git clone https://github.com/huggingface/lerobot.git
cd lerobot

# Create and activate virtual environment
python -m venv .venv
source .venv/bin/activate  # Linux/macOS
# .venv\Scripts\activate   # Windows

# Install with Dynamixel SDK
pip install -e ".[dynamixel]"

Method 2: pip Installation (Simple use)

pip install "lerobot[dynamixel]"

Note: Both methods install CLI commands (lerobot-find-port, lerobot-calibrate, etc.). With source installation, you can also run via python lerobot/scripts/....

Finding Ports

lerobot-find-port

macOS Output Example

Finding all available ports for the MotorBus.
['/dev/tty.usbmodem575E0032081', '/dev/tty.usbmodem575E0031751']
Remove the USB cable from your MotorsBus and press Enter when done.
...
The port of this MotorsBus is /dev/tty.usbmodem575E0032081

Leader arm: /dev/tty.usbmodem<SERIAL_A>
Follower arm: /dev/tty.usbmodem<SERIAL_B>

Linux Output Example

Finding all available ports for the MotorBus.
['/dev/ttyACM0', '/dev/ttyACM1']
Remove the USB cable from your MotorsBus and press Enter when done.
...
The port of this MotorsBus is /dev/ttyACM1

Leader arm: /dev/ttyACM0 or /dev/ttyACM1
Follower arm: remaining port

Linux USB Permission Setup

USB device access permissions are required. Using the dialout group is recommended for security:

Recommended: Add to dialout group (permanent, safe)

# Add current user to dialout group
sudo usermod -aG dialout $USER

# Log out and log back in to apply
# Or reboot

Alternative: Add udev rules (permanent, safe)

# Create udev rule file (dialout group based, security recommended)
# For Waveshare board and general USB-serial adapters
cat << 'EOF' | sudo tee /etc/udev/rules.d/99-dynamixel.rules
# Waveshare Serial Bus Servo Driver (CH340/CH341)
SUBSYSTEM=="tty", ATTRS{idVendor}=="1a86", ATTRS{idProduct}=="7523", GROUP="dialout", MODE="0660"
SUBSYSTEM=="tty", ATTRS{idVendor}=="1a86", ATTRS{idProduct}=="5523", GROUP="dialout", MODE="0660"
# FTDI USB-serial adapters
SUBSYSTEM=="tty", ATTRS{idVendor}=="0403", GROUP="dialout", MODE="0660"
# CP210x USB-serial adapters
SUBSYSTEM=="tty", ATTRS{idVendor}=="10c4", GROUP="dialout", MODE="0660"
# Generic CDC ACM devices
SUBSYSTEM=="tty", KERNEL=="ttyACM*", GROUP="dialout", MODE="0660"
EOF

# Reload udev rules
sudo udevadm control --reload-rules
sudo udevadm trigger

Note: The above rules only allow access for users in the dialout group. You must add your user to the dialout group.

Temporary Solution (development/testing only, security caution)

# Warning: world-writable permissions, security risk
# Use only temporarily in development environment

# CDC ACM devices (typically Koch board)
sudo chmod 666 /dev/ttyACM0
sudo chmod 666 /dev/ttyACM1

# USB-serial adapters (some environments)
sudo chmod 666 /dev/ttyUSB0
sudo chmod 666 /dev/ttyUSB1

Security Warning: chmod 666 grants read/write permissions to all users. Use dialout group or udev rules in production environments.

Motor Setup

# Follower arm motor ID setup
# macOS example
lerobot-setup-motors \
    --robot.type=koch_follower \
    --robot.port=/dev/tty.usbmodem575E0031751

# Linux example
lerobot-setup-motors \
    --robot.type=koch_follower \
    --robot.port=/dev/ttyACM0

# Leader arm motor ID setup
# macOS example
lerobot-setup-motors \
    --teleop.type=koch_leader \
    --teleop.port=/dev/tty.usbmodem575E0032081

# Linux example
lerobot-setup-motors \
    --teleop.type=koch_leader \
    --teleop.port=/dev/ttyACM1

Calibration

Calibration is essential for neural network transfer between robots. For policies trained on one robot to work on another, they must return the same values at the same physical positions.

# Follower arm calibration
# Replace <PORT> with actual port (e.g., /dev/ttyACM0 or /dev/tty.usbmodem...)
lerobot-calibrate \
    --robot.type=koch_follower \
    --robot.port=<PORT> \
    --robot.id=my_koch_follower

# Leader arm calibration
lerobot-calibrate \
    --teleop.type=koch_leader \
    --teleop.port=<PORT> \
    --teleop.id=my_koch_leader

Teleoperation and Data Collection

Important: In the commands below, replace <FOLLOWER_PORT> and <LEADER_PORT> with actual ports:

macOS: /dev/tty.usbmodem...

Linux: /dev/ttyACM0, /dev/ttyACM1 or /dev/ttyUSB0, /dev/ttyUSB1

Note: robot.type and teleop.type use koch_follower and koch_leader as in motor setup/calibration above.

With Source Installation (from lerobot directory)

# Teleoperation test (control Follower with Leader)
python lerobot/scripts/control_robot.py teleoperate \
    --robot.type=koch_follower \
    --robot.port=<FOLLOWER_PORT> \
    --teleop.type=koch_leader \
    --teleop.port=<LEADER_PORT> \
    --robot.cameras="{'cam': {'type': 'opencv', 'index': 0}}"

# Record dataset
python lerobot/scripts/control_robot.py record \
    --robot.type=koch_follower \
    --robot.port=<FOLLOWER_PORT> \
    --teleop.type=koch_leader \
    --teleop.port=<LEADER_PORT> \
    --repo-id=your-username/koch-picking-task

# Train policy
python lerobot/scripts/train.py \
    --policy.type=act \
    --dataset.repo_id=your-username/koch-picking-task

With pip Installation (module execution)

# Teleoperation test (control Follower with Leader)
python -m lerobot.scripts.control_robot teleoperate \
    --robot.type=koch_follower \
    --robot.port=<FOLLOWER_PORT> \
    --teleop.type=koch_leader \
    --teleop.port=<LEADER_PORT> \
    --robot.cameras="{'cam': {'type': 'opencv', 'index': 0}}"

# Record dataset
python -m lerobot.scripts.control_robot record \
    --robot.type=koch_follower \
    --robot.port=<FOLLOWER_PORT> \
    --teleop.type=koch_leader \
    --teleop.port=<LEADER_PORT> \
    --repo-id=your-username/koch-picking-task

# Train policy
python -m lerobot.scripts.train \
    --policy.type=act \
    --dataset.repo_id=your-username/koch-picking-task

VLA Research Applications

Imitation Learning Demos

Alexander Koch successfully demonstrated the following tasks with the Koch robot arm:

Simple Picking Task: Object picking learned from camera images and joint states only
Clothes Folding: Clothes folding task with dual arm configuration
Multi-Robot Control: Single neural network controlling multiple robot arms with language conditioning

Supported Policy Architectures

The following policies can be trained with Koch + LeRobot:

Policy	Description
ACT	Action Chunking with Transformers
Diffusion Policy	Diffusion model-based policy
SmolVLA	450M lightweight VLA model
TDMPC	Model-based RL

Community Data Collection

Koch’s low-cost structure enables distributed community data collection:

HuggingFace Hub: 487+ community datasets tagged with lerobot
Open X-Embodiment: Can contribute to large-scale robot datasets
Tau Robotics: Vision of millions of robots learning in the real world

Troubleshooting

Problem	Solution
Motor not recognized	Check USB cable, power cable, 3-pin cable
Waveshare board not recognized	Ensure jumper is set to ‘B’ channel (USB)
Motor ID conflict	Connect only one motor at a time to set ID
Linux permission error	Add user to `dialout` group or set udev rules (see above)
Calibration failed	Move each joint slowly through full range of motion
Insufficient torque	Check power supply, ensure 12V is being supplied properly

Comparison with SO-100/SO-101

Item	Koch v1.1	SO-100/SO-101
Motors	Dynamixel XL430/XL330	Feetech STS3215
Cost	~$440 (pair)	~$230 (pair)
Torque	Higher	Medium
Precision	Higher	Medium
Assembly Difficulty	Medium	Low
Community	Growing	Active
LeRobot Support	Official support	Official support

Koch v1.1

Overview

Key Significance

Low-Cost Data Collection Platform for VLA Research

Tau Robotics’ Vision

Technical Specifications

Motor Configuration

Follower Arm (Controlled Target)

Leader Arm (Teleoperation Controller)

Dynamixel Motor Details

Bill of Materials (BOM)

Follower Arm (Controlled Target)

Leader Arm (Teleoperation Controller)

Total System Cost

Where to Buy

v1.0 vs v1.1 Comparison

Key Improvements in v1.1

Assembly and Build

3D Printing Requirements

Assembly Process

Power Configuration

LeRobot Ecosystem Integration

Installation

Method 1: Source Installation (Recommended - for development/contribution)

Method 2: pip Installation (Simple use)

Finding Ports

macOS Output Example

Linux Output Example

Linux USB Permission Setup

Recommended: Add to dialout group (permanent, safe)

Alternative: Add udev rules (permanent, safe)

Temporary Solution (development/testing only, security caution)

Motor Setup

Calibration

Teleoperation and Data Collection

With Source Installation (from lerobot directory)

With pip Installation (module execution)

VLA Research Applications

Imitation Learning Demos

Supported Policy Architectures

Community Data Collection

Troubleshooting

Comparison with SO-100/SO-101

References

Official Resources

Where to Buy

3D Print Files

Related Projects

Community

See Also