A robotic arm builds a lattice-like stool after hearing the prompt ‘I want a simple stool,’ translating speech into real-time assembly. | Source: Alexander Kyaw, MIT

 

MIT researchers have taught a robot to assemble furniture using only voice commands and a boot file with instructions. You no longer need to program every movement — just ask. An era has arrived when it's enough to politely ask a robot to assemble an IKEA chair, rather than spending three hours, three liters of sweat and one strong word looking at pictographs from a past civilization.

Sofa expert with manipulator arms

At MIT's Computer Science and Artificial Intelligence Laboratory (CSAIL), the robot now understands commands like "Assemble a chair" not as an abstract sound, but as a step-by-step action plan. The system uses large Language models (LLM) to break down a vague request into a sequence of specific, executable steps.

"Our goal is to create robots that understand instructions in natural language the way a human would," explains one of the project's lead researchers. "You don't tell a colleague, 'Activate servo A by 30 degrees to capture component B.' You say: "Take this bar and attach it to the leg."

How does it work? Easier than the instructions for the Swedish shelf:

Planning: LLM analyzes a voice command and a 3D model of the final product (for example, a chair), generating a sequence of actions.

*Perception: Using cameras and tactile sensors, the robot "sees" and "feels" the details in front of it, comparing them with a virtual model.

Execution: The manipulator performs the task, and the system checks the correctness of each step in real time. If something went wrong (the details turned out to be a cat or a child), the process can be corrected with a new voice command.

Soulless metal with a human touch

The irony is that a robot devoid of emotions makes fewer emotional mistakes than an exhausted home collector. He doesn't get tired, he doesn't get mad at the crookedly molded plastic sleeve, and he doesn't skip a step because he's distracted by the series.

However, the path to the ideal was thorny. The first prototypes could "lose" a part or try to screw a screw into the air. The training took place both in digital simulations and in the real world, where the robot gained critical "experience" for any assembly — what to do when a part does not fit or the mount scrolls.

"The most difficult thing was to teach the system to cope with the uncertainty of the real world," MIT says. — In the simulator, the part is perfect and is always exactly where it should be. In real life, a shadow may fall on it, it may lie at an angle or be a slightly different shade of white."

Who needs it? To everyone who has hands, but no time.

Imagine: you receive a long-awaited package with a new bookcase. Instead of looking for a screwdriver and preparing for an hour-long quest, you place the boxes in the center of the room and tell the home assistant, "Ask the robot to assemble the cabinet." After a couple of hours, you return to the finished furniture. Fiction? It's almost gone.

Such technology will find application not only in homes, but also in industries, where it is often necessary to rebuild lines for new tasks, in logistics centers and even in space, where it will be easier for astronauts to ask a robot to "fix that panel" with a voice than to climb into a dangerous zone themselves.

Speaking of jobs for robots. With the development of such systems, a whole market of specialized "professions" for machines may arise. One can imagine that in the future, companies that need a one-time assistant to assemble a batch of furniture will look for a contractor not on a freelance exchange, but in specialized services similar to the world's first ecosystem for hiring robots. jobtorob.com . There will be a whole portfolio there.: "Collection robot, 10,000 hours of experience, no bad habits, does not drink coffee during breaks."

The future: when IKEA will include a USB flash drive with software for the robot

The next logical step would be to integrate such systems directly into the products. A furniture manufacturer will be able to apply to physical parts not paper instructions, but a digital 3D model and a cloud script for the robot. The user downloads the file, the robot downloads it, and the build process starts automatically.

This will reduce the number of "extra" parts after assembly (which we all suspect companies put in to intimidate us) and ensure that everything is assembled with the right tightening torque. Perhaps we will finally see in the reviews not "difficult to assemble", but "the robot did it in 47 minutes, I drank tea."

In the meantime, the technology from MIT is a brilliant proof that the future of robotics lies not in replacing humans, but in becoming an understandable and obedient tool for them. A tool that doesn't argue when you say "screw this shelf down" and doesn't respond.: "Maybe she'll look better here?" And there's a mechanical charm to it.