Robots told to avoid an obstacle hit it anyway — until a checking step was addedtest headline
An AI that writes robot control code kept crashing into objects it had been told to avoid. Adding a tool that checks every planned path before the arm moves raised collision-free runs from 59% to 87.5%.test gist
ترجمهٔ ماشینی است؛ برای دقت به متن اصلی انگلیسی مراجعه کنید.
Why it matters
Robot arms are starting to work in warehouses, kitchens and hospitals, next to people and fragile things. A robot that finishes its job but knocks over a bottle on the way is not safe enough. The useful lesson here is simple: a safety rule written in words at the start of a long job can slip out of the AI's attention, so it has to be re-checked by a separate step each time the arm moves. Be careful with the numbers, though — all of this happened in a computer simulation, not with a real arm, and the robot still displaced the obstacle in about one run in eight.test why
Who's behind it: Bingxin Xu, Yuzhang Shang, Zhen Dong and Emilio Ferrara, Unknown.
Summary by the Lemma AI · how we grade