Free Toy Robot Arm Simulator Trains AI on a Normal PC

A small robotics company has released a free browser-based demo that lets anyone control a toy robot arm using simple text or voice commands. The AI model behind it is tiny—just 6.5 million parameters—and was trained on a regular desktop computer, not a powerful server. Despite its size, it can successfully follow instructions like 'pick up the red block' about 95% of the time, even in scenes it hasn't seen before. This shows that you don't need expensive hardware to experiment with robot AI.

The project started because the company needed small AI models that could run directly on robot hardware. Training larger models was painful on their workstations, and standard robotics datasets filled up their hard drives. So they built a tool that generates training data on the fly—images and language instructions—allowing them to test if a model can learn a simple task before spending days training on real data. The entire process was coded with help from AI assistants like Qwen and Claude.

The demo also includes a 3D model viewer that lets you fly through the layers of the AI model in a Gibson-style virtual reality. This isn't just for fun; it helps people understand how real models work. For example, seeing the repeated passes of a larger model called SmolVLA laid out in space made the concept of flow matching easier to grasp. The code is open-source and extends to other PyTorch and LeRobot models, so you can explore different architectures.

The data generator and model are both open-source under the Apache license, meaning anyone can use, modify, and share them. The company is asking for feedback on which models to add and how to make the tool more useful. This project lowers the barrier to entry for robotics AI, allowing hobbyists, students, and researchers to experiment without expensive equipment. It's a step toward making robot programming more accessible and understandable for everyone.