AI/ robotics · ai · dataset · household-robots

AI Model Learns to Run Home Appliances by Reading Manuals

A new dataset and 7B parameter model teach robots to operate 22 types of appliances by parsing their manuals, beating prior systems by 10x in tests.

A new AI system reads appliance manuals, then tells a robot how to actually run the appliance.

Researchers built a pipeline called MAGE that mines appliance manuals for instructions, using what they call a Hierarchical Appliance Graph to automatically generate three kinds of training data: which physical part does what, how to plan a multi-step task, and how to recover when a step fails. That pipeline produced UseAppliance, a dataset spanning 22 appliance categories with more than 89,000 labeled parts, 53,000 manipulation tasks, and 33,000 closed-loop correction steps. The team then trained a 7-billion-parameter model, AppliancePlan, on that data. On a benchmark called RealAppliance-Bench, it beat the strongest existing baseline by more than 10x on open-loop planning, and the researchers also ran it on a real robot across six household appliances.

The real bottleneck for appliance-operating robots has never been the robot arm, it's the lack of task data, and manuals exist for nearly every consumer product already. Auto-generating training data from manuals, instead of hand-collecting demonstrations, is the kind of scaling trick that could move household robotics forward faster than manual data collection ever would.

A 10x improvement over a self-selected best baseline is the sort of number that deserves scrutiny once outside reviewers get a look, and going from six appliances in a lab to your actual kitchen remains a long way off.

TR

The Revision

Written by an AI system from the public sources credited above. How we write →