OpenAMRobot 2.0: the open dual-arm robot anyone can teach

Update, October 2026: OpenAMRobot 2.0

OpenAMRobot is now becoming a general-purpose, open-source dual-arm mobile manipulator with an embodied AI software layer. The idea has not changed: people who know the work teach the robot by demonstration (show, correct, judge) instead of writing robot code.

About the grocery video. The renders and the video show the robot picking two customer orders in a simulated dark store. Grocery order picking is one use case we use to demonstrate and test the platform. It is not what the robot is limited to. The base, lift, arms, sensors and software are not tied to totes or shelves, and the same robot can be taught other mobile manipulation jobs, from lab and hospital logistics to kitting, replenishment, research and teaching. Renders and video come from our simulation; the physical 2.0 robot is integrated and tested in Berlin from late November.

Video: https://www.youtube.com/watch?v=rQ5obdNkIZE

What is new in 2.0

• Two OpenArm 2.0 arms by Enactic (7 DoF each) on a lift column, shoulder height 1000 to 1350 mm with three calibrated stops
• Stereo head camera and in-hand cameras from the OpenArm set, Orbbec Gemini 336L depth camera at the front of the base
• NVIDIA Jetson Orin NX as the single high-level computer; STM32H723 (NUCLEO-H723ZG) base controller with micro-ROS over Ethernet
• ZLTECH 200 mm hub motors with fail-safe holding brakes and a CANopen driver
• 25.6 V LiFePO4 pack (8S, 105 Ah) with BMS on CAN, designed for TZBOT wireless charging
• Safety layer: two E-stops cut motor power through relays, electronics stay powered, one reset button
• SLAMTEC RPLIDAR S3, 4 ultrasonic sensors, 2 ToF cliff sensors, ICM-42688-P IMU
• CAN only on the robot, 1.5 m/s software speed ceiling, under 120 kg gross, up to 1.7 m tall
• ROS 2, Nav2, MoveIt 2, ros2_control, an arm-agnostic manipulation server, the browser operator UI, and a new embodied AI repo for demonstration recording, an ACT baseline and evaluation

Timeline

• 20 Nov 2026: v2.0.0-rc.1 pre-release (design, software and manufacturing package)
• 23 Nov to 18 Dec: physical integration and acceptance testing in Berlin
• 18 Dec 2026: v2.0.0 final release
• December: hands-on hackathon with the robot at MotionLab.Berlin
• January 2027: OpenAMRobot 3.0 development cycle starts

Still fully open: software MIT, documentation CC BY 4.0, hardware CERN-OHL-P-2.0. Around 20 contributors work on 2.0 in the open. Announcement and discussion: https://github.com/orgs/openAMRobot/discussions/27 . Docs: https://docs.openamrobot.ai

The original v0.1 story from July 2026 follows below.

OpenAMRobot - the open robot anyone can teach

Most robots still need a robotics engineer, a system integrator, and months of custom work before they do anything useful. That's the wall I'm trying to knock down. OpenAMRobot is an open, affordable robot designed so the person who actually knows the work - not a roboticist - can put it to work.

It's a full system, not a one-off demo: an autonomous mobile platform with a complete ROS 2 architecture underneath, built to be modular, reproducible, and eventually deployable in the real world.

What it does today

Right now it's a mobile platform you can direct in the simplest possible ways:

• Drop a waypoint on a map → it navigates there autonomously, avoiding obstacles.
• Drag-and-drop blocks → snap together a task sequence, no code.
• Plain language → tell it what to do and it goes.

And I'll be honest - none of those, on their own, is bleeding-edge in 2026. Waypoint nav is old. Blockly is older. Voice control is everywhere. That's kind of the point: the boring foundation works, it's all open, and it proves something that matters - a person with zero robotics background can direct a real robot.

Under the hood

• Compute: Raspberry Pi 5 (navigation / system), with room for a Jetson for AI perception
• Software: ROS 2 - Nav2 navigation, control, and system orchestration
• Perception: LiDAR + cameras for SLAM, mapping, and obstacle avoidance
• Interface: a block-based (Blockly-style) UI + natural-language command layer for the operator
• Design: modular, header-based, community-reproducible - built to be built by other people

Where it's headed (the actual point)

Driving somewhere is the easy case. The hard, valuable case is manipulation - teaching a robot to do a real physical task with two arms. So the next step puts a dual-arm setup on this platform (in development), where you don't program it - you show it a task a few times and it learns: imitation learning, multiplied in simulation, then refined with reinforcement learning until it's reliable.

To the person using it, it just feels like: "I showed it, and it learned." The goal is simple and a little stubborn: domain expertise becomes the way you program a robot, and all the AI complexity stays under the hood, where it belongs. The long-term direction is a mobile humanoid and embodied AI anyone can direct.

Open-source, and honestly early

Everything is open-source and MIT-licensed. It's early, and genuinely a bit rough - sometimes it's not even wearing its cover. I'm building it in the open, posting the wins and the faceplants, because I think this kind of technology shouldn't stay locked inside billion-dollar labs.

Feedback, questions, and contributors are very welcome.

👉 Repo & build details: github.com/openAMRobot

Built by Alex Reznichenko: founder, system architecture, hardware and firmware, ROS 2, the learning platform, and maintainer of the open-source repository.