New Test Asks Robots to Climb a Ladder and Change a Bulb

Changing a ceiling light bulb sounds like the easiest chore in the house. For a robot, it's a nightmare. You have to carry a ladder, place it under the right spot, climb it without tipping over, twist out a hot bulb without crushing it, screw in a new one, and carry the dead bulb back down to a trash bin. That's balance, grip, vision, and planning all at once — and no existing robot test measures the whole thing from start to finish.

So a team of researchers built Fiatlux, a measuring stick for exactly that job. It runs inside NVIDIA's Isaac Lab, a simulation environment where a virtual Unitree G1 humanoid — a real commercial robot, roughly human-sized — attempts the full chore. The task is split into twelve smaller steps, each scored separately. Robots that only manage to place the ladder or climb halfway still earn partial points, which tells engineers where the weak links are. The rules are strict: the fresh bulb must be seated properly, the old one must land in the disposal crate, neither can be dropped, and the bulb can't be squeezed hard enough to break.

The team also released recordings of a human teleoperating the robot through the task, plus baseline results from three different AI approaches — including a vision-language-action model (AI that sees, reads instructions, and moves the robot at once). Those baselines show how far today's best systems get. Spoiler: not all the way.

Why does this matter if it's only a simulation? Because benchmarks like this become the scoreboard that funding, hiring, and product roadmaps follow. A robot that can reliably handle a fragile bulb on a ladder is also a robot that can stock shelves, reach high windows, or work in warehouses. This test frames the goal as a whole job, not a party trick — and that's the gap between robot demos and robots you'd actually pay.