Skip to content

Open-Source Robot Hardware

The explosion of open-source hardware has dramatically lowered the barrier to robot learning research. From $110 desktop arms to $32K mobile dual-arm platforms, researchers now have a rich set of choices. This article compiles the current mainstream open-source robot hardware platforms.

Platform Overview

Platform Type Cost DOF Key Features Developer
ALOHA Desktop dual-arm teleoperation ~$20K 2x6+2 grippers Bilateral teleoperation, original ACT policy platform Stanford
ALOHA 2 Desktop dual-arm teleoperation ~$20K 2x6+2 grippers Engineering-improved ALOHA, DeepMind companion Google DeepMind
Mobile ALOHA Mobile dual-arm ~$32K 2x6+mobile base ALOHA on mobile base, whole-body teleoperation Stanford
UMI Gripper data collection tool ~$200 1 gripper iPhone ARKit tracking, portable data collection Columbia
GELLO Joint-space teleoperation device ~$200 Matches target arm Kinematic matching, direct joint-space mapping Berkeley
LEAP Hand Dexterous hand ~$2K 16 3D printed, direct drive, RL-friendly CMU
Koch v1.1 Low-cost single arm ~$250 6+gripper Official LeRobot support platform Jess Moss / Community
SO-100 Low-cost single arm ~$110 5+gripper Currently the cheapest research-grade arm HuggingFace / Community
Open Manipulator X Single arm ~$500 4+gripper Dynamixel driven, native ROS2 support ROBOTIS

Detailed Introductions

ALOHA / ALOHA 2

ALOHA (A Low-cost Open-source Hardware System for Bimanual Teleoperation) was released by Stanford's Zhao et al. in 2023 and is the de facto standard for dual-arm manipulation research.

Hardware Components

Component Specification
Follower Arms 2x ViperX 300 6DOF (Interbotix)
Leader Arms (Teleop) 2x WidowX 250 6DOF (Interbotix)
Gripper Parallel gripper
Motors Dynamixel XM/XH series
Cameras 4x Logitech C922 (top + wrist)
Control Frequency 50Hz
Frame 80/20 aluminum extrusion

Bilateral Teleoperation Principle

The operator (Leader) and follower use robot arms with identical kinematic structure. The operator moves the Leader arm, and joint angles are mapped in real time to the Follower arm:

\[\theta_{follower}(t) = \theta_{leader}(t)\]

Since both arms have identical kinematics (joint-space mapping), no inverse kinematics computation is needed, and the mapping is precise with minimal latency.

ALOHA 2 Improvements

Google DeepMind made engineering improvements on top of ALOHA:

Improvement Description
Better camera mounting More stable, better viewing angle coverage
Improved gripper Wider grasp range
Better cable management Reduced interference during joint motion
Data collection software Compatible with RT-2, ALOHA Unleashed algorithms

Associated Algorithms

Algorithm Paper Description
ACT Learning Fine-Grained Bimanual Manipulation CVAE + Transformer, original ALOHA paper
Diffusion Policy Diffusion Policy Diffusion model generates action sequences
ALOHA Unleashed — Google DeepMind large-scale training

Build Notes

  • Assembly time: ~2-3 days (including 3D printing wait time)
  • Key tool: Dynamixel Wizard 2.0 (set motor IDs and baud rates)
  • Common issues: Leader/Follower zero-point calibration, cable routing to prevent tangling
  • GitHub: https://github.com/tonyzhaozh/aloha

Mobile ALOHA

Adds a mobile base to ALOHA for whole-body teleoperation.

Component Specification
Base AgileX Tracer
Robot Arms 2x ViperX 300
Teleoperation 2x WidowX 250 + base joystick
Additional Sensors IMU, wheel odometry
Total Cost ~$32K

Research Significance: Demonstrated the feasibility of "whole-body teleoperation" — the operator simultaneously controls base movement and dual-arm manipulation, collecting mobile manipulation data.

Build Notes

  • Mechanical mounting between base and dual arms requires custom mounting plates
  • Power system needs unification (base battery powers both arms and computing platform)
  • GitHub: https://github.com/MarkFzp/mobile-aloha

UMI (Universal Manipulation Interface)

UMI is an ultra-low-cost data collection solution proposed by Columbia University. The core idea is tracking the 6-DoF trajectory of a hand-held gripper using iPhone.

Working Principle

  1. 3D print a hand-held gripper (with iPhone mount)
  2. iPhone uses ARKit to provide 6-DoF pose tracking
  3. During collection, the operator manipulates the hand-held gripper in real scenes
  4. Record gripper pose trajectory + open/close state
  5. During playback, map trajectory to target robot

Core Advantages

Advantage Description
Ultra-low cost ~$200 (iPhone not included)
Platform-agnostic Data can be replayed on different robots
Natural operation Operator holds directly, no teleoperation learning needed
Scalable Anyone can participate in data collection

Limitations

  • Depends on iPhone ARKit accuracy (~1cm position error)
  • No force feedback, not suitable for precision assembly
  • Requires Cartesian control capability on the target robot

Build Notes

GELLO

GELLO is a low-cost joint-space teleoperation device developed by Berkeley, building a kinematically scaled-down teleoperation handle for the target robot.

Design Philosophy

Feature Description
Joint count Same as target robot (6 or 7)
Motors Dynamixel XL330-M077 (~$25 each)
3D printed All links fully 3D printed
Total cost ~$200
Mapping method Direct joint-space mapping (no IK needed)

Adapted Robots

Target Robot GELLO Version
Franka Panda 7-DOF GELLO
UR5 6-DOF GELLO
xArm 6-DOF GELLO
Koch 6-DOF GELLO

Build Notes

  • Link proportions need customization for the target robot (open-source repository provides multiple versions)
  • Motor ID setup and zero-point calibration are critical steps
  • GitHub: https://github.com/wuphilipp/gello_software

LEAP Hand

LEAP (Low-cost, Efficient, and Agile Prosthetic) Hand is an open-source dexterous hand developed by CMU and is currently the most active open-source dexterous hand platform.

Feature Specification
DOF 16 DOF (4 fingers x 4 joints)
Actuators 16x Dynamixel XC330-T288-T
Drive Method Direct drive (no tendons)
Material 3D printed (PLA/PETG)
Weight ~500g
Grip Force ~15N fingertip force
Control Frequency 200Hz
Cost ~$2,000

Research Value:

  • 16 DOF provides the redundancy needed for dexterous manipulation
  • Direct drive simplifies control (no tendon breakage risk)
  • Low cost makes multiple experimental setups feasible
  • Adopted by multiple research groups for dexterous manipulation RL research

Build Notes

  • Print time: ~20-30 hours (all parts)
  • Assembly time: ~1 day
  • Key note: Motor wiring needs to be compact; finger joint alignment precision affects grasp quality
  • GitHub/Homepage: https://leap-hand.github.io/

Koch v1.1

Koch is a low-cost 6-DOF robot arm designed for the LeRobot ecosystem.

Feature Specification
DOF 6 + gripper
Motors Leader: XL330-M077, Follower: XL430-W250
Workspace Radius ~30cm
Material 3D printed
Cost ~$250 (single arm)
LeRobot Integration Official support

Suitable for: Entry-level researchers wanting to do real-robot experiments with LeRobot.

Build Notes

  • BOM Example (dual-arm system):
Category Component Qty Unit Price Subtotal
Motor (Follower) Dynamixel XL430-W250 6 ~$50 $300
Motor (Leader) Dynamixel XL330-M077 6 ~$25 $150
Structural parts 3D printed parts — — ~$20
Communication U2D2 USB adapter 2 ~$30 $60
Power 12V 5A power supply 2 ~$15 $30
Cables Dynamixel cables Misc. — ~$20
Camera Logitech C920 2 ~$50 $100
Total ~$680 (dual arm)

SO-100

SO-100 is currently the lowest-cost research-grade robot arm, driven by the HuggingFace community.

Feature Specification
DOF 5 + gripper
Motors Feetech STS3215 (bus servo)
Control Frequency 30-50Hz
Material 3D printed
Cost ~$110 (single arm)
LeRobot Integration Official support
Communication Half-duplex UART

Strengths: Extremely low cost, official LeRobot adaptation, active community.

Weaknesses: Low torque, limited precision, average servo quality.

Build Notes

  • Feetech servos use half-duplex UART, requiring a dedicated serial adapter board
  • Not compatible with Dynamixel series (different protocols)
  • LeRobot provides complete calibration and usage tutorials
  • Reference: LeRobot SO-100 Tutorial

Open Manipulator X

An officially produced ROS2-native robot arm by ROBOTIS.

Feature Specification
DOF 4 + gripper
Motors 4x Dynamixel XM430-W350 + 1x XM430-W350
Workspace Radius 38cm
Payload 500g
Communication OpenCR (STM32) + U2D2
ROS2 Support MoveIt 2 + ros2_control
Cost ~$500

Strengths: Best ROS2 integration, has MoveIt 2 configuration, comprehensive official documentation.

Weaknesses: Only 4 DOF, not flexible enough.

Build Notes

General 3D Printing Recommendations

Most open-source hardware platforms rely on 3D printed structural parts. General recommendations:

Parameter Recommended Value
Material PLA+ or PETG (PETG is more durable)
Layer height 0.2mm
Infill 30-50%
Wall count 3-4
Printer Any FDM (Bambu Lab P1S recommended)

General Assembly Notes

  1. Motor ID setup: Each Dynamixel needs a unique ID (use Dynamixel Wizard 2.0)
  2. Communication baud rate: 1Mbps (1000000) recommended
  3. Origin calibration: Mark zero positions before assembly
  4. Cable routing: Prevent cables from being tangled by joints
  5. Gripper adjustment: Spring pre-tension affects grasp success rate

Quick Selection Guide

Goal Recommended Platform Rationale
Lowest-cost entry SO-100 ($110) Cheapest, LeRobot support
LeRobot research Koch v1.1 ($250) 6DOF + official support
Low-cost teleop device GELLO ($200) Adapts to multiple target arms
Large-scale data collection UMI ($200) Ultra-low cost, platform-agnostic
Dual-arm manipulation research ALOHA ($20K) Academic standard, original ACT platform
Mobile manipulation Mobile ALOHA ($32K) Mobile + dual arm
Dexterous hand research LEAP Hand ($2K) 16DOF, direct drive
ROS2 education Open Manipulator X ($500) Native ROS2

GitHub Repositories:

Related Notes:

评论 #