Researchers introduced MessyNav, a framework that helps mobile robots navigate cluttered spaces by deciding what to do with obstacles in their way. It uses a vision-language model—a system that interprets images and text—to assess objects before planning whether to swipe them aside, pick them up or avoid them.
In real-robot trials, the team reports a 27-percentage-point improvement in success rate over its strongest baseline. The researchers also report statistically significant gains in a larger offline evaluation, without training specifically for the tasks.
Why it matters
For teams building robots for homes or other cluttered spaces, MessyNav offers a way to choose between moving and avoiding objects without task-specific training; the abstract does not establish how well it works outside the reported evaluations.
