New AI Trick Makes Home Robots Actually Finish Their Chores

Ask an AI chatbot to plan your Saturday and it sounds brilliant. Ask a robot to actually do it — find the sponge, fill the bucket, mop the kitchen, put everything back — and things fall apart fast. Robots powered by large language models (AI trained on huge amounts of text) often invent steps that are impossible, forget what they already did, or wander around looking for objects that aren't where they assumed. In tests, these robots finished hard, multi-step chores only about 20 to 41 percent of the time.

A research team from Duke University and collaborators built a fix called GAVEL. Think of it as giving the robot a checklist and a mental map instead of asking it to improvise. GAVEL builds a simple graph — a web of connections — showing which objects relate to each other, what each action requires before it can happen, and what it changes afterward. Before the robot moves, GAVEL checks its plan against that map and silently repairs obvious mistakes, saving the expensive AI thinking for genuinely tricky decisions. When a task has several parts, GAVEL also reasons about where objects probably are, reordering steps to search less.

The results are striking. On BEHAVIOR-1K, a widely used simulated household-chore test, single-task success rose from 41.2% to 91.8%. Multi-task instructions — like 'clean the living room and then set the table' — jumped from 19.9% to 92.6%. The robot also travelled about 5.4% less distance, meaning less wasted time and battery. Notably, this worked with a relatively small, cheap AI model, suggesting you won't need a giant data center in your hallway.

The honest catch: BEHAVIOR-1K is a computer simulation, not a real apartment with stairs, pets, and a toddler moving your keys. Real-world messiness — dropped objects, broken handles, a human in the way — is the next hurdle. Still, this is the kind of plumbing work that turns flashy robot demos into something you'd actually trust with the dishes.