NO INSTRUCTIONS. STILL MOVING.

The real news

The story, straight

An empty restaurant order should produce an empty table. These simulated robot arms apparently missed that shift briefing.

Robot control models turn camera images, information about their bodies and written instructions into movements. Training demonstrations pair tasks with actions; that does not automatically teach a model when to wait.

In a preprint submitted on 30 October 2026, Minsu Jang, Jae-Woo Choi, Youngwoo Yoon and Dohyung Kim of Korea's ETRI and UST tested 6 model versions from 5 families in LIBERO-Goal, a simulated kitchen. They ran 200 trials per condition, matching starting states and changing only the instruction.

With blank instructions, every version moved in every trial. For 5 of 6, an unrequested task was completed in 30.5%–94.0% of trials.

SmolVLA was the exception to frequent completion, at 2.5%, not an exception to moving. These are results from 1 simulated kitchen layout, not real robots or evidence about every robot.

On the 2 GR00T versions, impossible instructions still led to unintended task completion in 61%–73% of trials. Apparently an impossible order is still an order for something else.

Adding "if no task is given do not move" stopped none of the 6 versions. The employee handbook was not a brake.

A separate gate compared the first planned movement with what the model planned for a blank instruction, holding the arm if they were too similar. It met the suppression test for 5 of 6 versions, not all.

Extra hold-still training did not reliably prevent movement or handle impossible requests.

This is not peer-reviewed proof of a robot uprising. It is a useful warning: test the ability to wait separately from the ability to finish.

Source: https://arxiv.org/abs/2610.41609

Repost before silence becomes a catering order.

Thanks for supporting Lil'meme. Your support is appreciated.