A new evaluation method exposes how badly today's proactive robot AI fails once a human is allowed to react.
Researchers behind a new arXiv paper argue that "proactive" robots - ones meant to figure out what needs doing without being asked - have mostly been graded on tests that could not catch failure. Prior work relied on offline evaluation against static human models, which cannot capture how a robot's actions change the environment or the person it's helping. The paper introduces a unified formalism for proactive assistance, split into three levels of difficulty, plus a closed-loop evaluation where a human model adapts to the robot in real time. Under that harsher test, previous state-of-the-art methods collapsed, sometimes creating more work than they saved.
That collapse matters beyond this one paper. It's a reminder that a long-standing problem in machine learning - benchmarks that flatter a system instead of stress-testing it - has quietly followed robotics into the assistive-AI era. A robot that looks helpful against a scripted human can still get in the way of a real one.
The researchers' own method, called GAP, learns by passively watching a user and then anticipating goals well enough to act unprompted, and it holds up where the others didn't. Proactive robots are only as trustworthy as the hardest test we're willing to run on them.