Robots Can Now Learn New Skills by Reusing Old Ones

Today's humanoid robots learn one skill at a time — walking, picking up a cup, waving hello. The problem is that each of those skills is built with different software, so you can't easily snap them together into a longer routine. A team of researchers says the fix is simple: no matter how a robot learned a move, it ultimately boils down to a path of motion through space. So they built CHOREO, a system that converts every skill into one shared file format they call SkillMotion, which records the movement, when the feet touch the ground, and what the move is supposed to mean.

The clever part is what you don't have to do: retrain the robot. Normally, teaching a robot a new combination of actions means hours of expensive computing and re-teaching, like reshooting a whole movie to add one scene. CHOREO just stitches existing skills together instead — more like editing clips in a video. The team loaded 2,950 skills into a simulated Unitree G1 robot and ran 130 multi-step tasks, hitting 95.4% success overall and 93.8% on the hardest tests, which chained eight actions in a row.

The honest catch: all of this happened inside a computer simulation, not on a real robot in your kitchen. Simulated physics is tidy — no slippery floors, no uneven carpet, no toddler grabbing the robot's arm mid-step. The skill library also has to be built by people first, so this isn't a robot that teaches itself overnight. And a research paper is a long way from a product you can buy.

Still, the direction matters. If robot skills can be stored and mixed like apps on a phone, then adding a new ability to a robot gets much cheaper and faster. That accelerates the arrival of robots that genuinely help around warehouses, hospitals, and eventually homes — because the hard part stops being "teach it everything from scratch" and starts being "plug in one more move."