Qubit Is the Desk Robot That Cooks and Cleans

There are all sorts of desktop AI assistants these days. Some of them answer our questions and help us manage our calendar and email, while others are designed more for entertainment or stress relief. One thing nearly all of these assistants have in common is that they are only useful for digital tasks. They can look up information stored in their huge models and spit it back out, but they are lost if you ask them to hold something while you solder it or fold your laundry.

Samuel Boylan-Sajous is seeking to change this with his recent project. He has designed and built an AI-powered assistant for his desk. Unlike other builds, this one, called Qubit, has robot arms. Boylan-Sajous has also developed a method to teach Qubit to use these arms for complex tasks to give its owner more than just a virtual hand.

The project begins not with exotic hardware, but with standard pegboard mounted to the wall. This pegboard acts as the backbone of the system, providing a modular structure for mounting tools and the robot’s components. Attached to it are a pair of SO-100 robotic arms, which are themselves open source. The arms were 3D printed using a Bambu Lab A1 printer, with parts sourced directly through MakerWorld, allowing the entire kit to be fabricated and assembled with relative ease.

Assembly involves installing motors into the printed housings and securing them in place. To give Qubit some personality, Boylan-Sajous added an LED matrix capable of displaying more than 72 different facial expressions. The animated “face” can show emotions like happiness or surprise and shift its eyes along X and Y axes to track movement. LED light tubes complete the build, giving the wall-mounted robot a polished, cinematic aesthetic.

Control is handled through a custom-built dashboard. Users can manually manipulate each motor via on-screen sliders, or record animations and play them back later. Because the arms are mounted against a wall, the software includes safety constraints that define minimum and maximum motion ranges, preventing costly collisions.

But prerecorded animations can only go so far. Tasks like whisking eggs (who does that at their desk?) or folding laundry introduce too many variables. To address this, Boylan-Sajous integrated cameras and computer vision, feeding data into AI models trained in simulation. Using NVIDIA Isaac Sim and NVIDIA Isaac Lab, he created a high-fidelity virtual version of Qubit. In this environment, the robot can practice complex tasks at accelerated speeds, compressing what would be years of physical trial and error into minutes of simulated learning.

Qubit’s source code and design files are available on GitHub for anyone who would like to make their own desk robot.