“Explain what you do and clean up the table,” the researcher instructs.
“I’ll use the robot arm to clean the table surface,” the robot replies. “First, I’ll check for available motion skills to pick up the sponge and wipe the table, then I’ll execute the sequence to clean the spill.”
The robot arm grabs the cleaning sponge and cleans the surface. It doesn’t do a fantastic job, but it does work on the task. The most surprising part is that this whole system was built in only a few days, for less than the cost of a Taylor Swift ticket.
Imagine a robot arm that can clean up a spill, understand instructions, and explain its actions — all for around $120. In a groundbreaking project, researchers at UC Berkeley and ETH Zurich accomplished just that. Using OpenAI’s GPT-4o and a pair of budget-friendly robot arms, Jannik Grothusen and Kaspar Janssen successfully trained their robots to handle tasks like spill cleanup in only four days.
The researchers started with cheap, open-source robot arms (open-source refers to software or hardware whose design is publicly accessible, allowing anyone to view, modify, and distribute the source code or design).
By feeding the robot arms around 100 demonstrations, they developed motion policies using an already existing framework called the ACT (Actor-Critic Training) framework. This approach teaches the robots to mimic human movements, improving their efficiency in handling tasks. The result? A robot that can recognize a spill, pick up a sponge and clean it up — all while explaining its actions.
Robotics becomes more open-source
One of the most exciting aspects of this project is its accessibility. The SO-100 robot arms used in the experiment are affordable and fully open-source. Detailed assembly instructions and all necessary files are shared on YouTube and GitHub. This makes it possible for anyone with basic 3D printing skills to replicate the project at home.
This “proof-of-concept for a robot control architecture” demonstrates how open-source is beginning to democratize the field of robotics, notes Grothusen in a LinkedIn post. The entire project took four days to complete and can even be replicated by inexperienced people.
This democratization of robotics is a significant shift in the field. Finally, with 3D printing and AI, we have a recipe to make cheap, customizable robots. By combining open-source hardware with AI-powered software, even small teams or individual enthusiasts can now develop robots capable of performing complex tasks. Grothusen and Janssen’s work exemplifies how collaboration, open-source platforms, and cutting-edge AI can transform the future of robotics.
The 3D-printed components are optimized for durability, using carbon fiber materials to ensure longevity. Adjustable tolerances and refined joint designs allow for precise movements, making the robot both robust and agile. Assembly is straightforward, requiring only common tools like screwdrivers. The entire process can be completed in about 30 minutes, even by those with minimal technical expertise.
GPT is an important tool
GPT-4o might seem like an unusual tool for training robots, and it’s definitely not a perfect one. Yet, its ability to process language and context proves invaluable. By using a visual language model (VLM), the researchers could guide the robot through tasks that require both reasoning and physical interaction. The robot didn’t just follow rigid pre-programmed instructions; it adapted to changes in its environment and responded intelligently to commands.
It’s also a tool that’s now widely available to anyone.
As open-source platforms and affordable hardware continue to evolve, the barriers to entry for robotics are lowering. This means that innovative solutions like these might not be confined to research labs for long. Soon, we could see affordable, AI-powered robots assisting in everyday life, transforming our homes and workplaces.
With the help of AI models like GPT-4o, what was once considered science fiction is becoming reality. These robots, powered by language models, are breaking into the physical world, demonstrating how AI can extend its capabilities beyond text-based tasks.
So, the next time you spill a drink, don’t be surprised if a robot arm swoops in to clean it up — courtesy of cutting-edge AI and a few 3D-printed parts.
