New Robot AI Learns Chores Without Watching Humans First

Robots usually learn by copying. A human picks up a cup a hundred times, the robot watches, and eventually it can do the same thing. That works, but it is slow, expensive and useless the moment you hand the robot a mug it has never seen. A research team has now shown a different route: let an AI language model write the instructions instead of filming a human.

The system, called ManiSkillFormer, works like a recipe book. Before the robot moves, an AI model writes a short "contract" for the job. That contract says which physical details matter — the handle of a mug, the flat top of a box, the edge of a lid — and roughly how the arm should move. The robot's cameras then look only for those details, ignoring everything else in the room. Once the task works, the movement gets saved in a library of reusable skills, so a new object does not mean starting from zero.

The results are promising. On a two-armed robot, the system picked and placed 30 different objects across 8 categories with 88% success — and it had never practiced with those specific objects. It also managed real-world style tasks like unscrewing a cap, pouring, pressing a button and folding. Longer, multi-step jobs succeeded between 50% and 80% of the time.

The catch is that this is still a laboratory result on one specific robot. Half-finished long tasks mean you would not leave it alone with your dishes. And "no demonstrations" does not mean "no setup" — the AI still needs clear instructions about what the task is. Think of it as a robot that reads a manual well, not one that figures out your kitchen on its own. Home robots that genuinely help are getting closer, but they are not arriving this year.