Half Right on Memory

Georgia Tech researchers built a test for something home robots are rarely graded on, memory. The benchmark, FindingDory, runs 60 tasks that ask whether a robot can remember where things are and what already happened in a home. The best model managed about half of the hardest tasks. Most current vision-language models, one of the researchers notes, can process only a few hundred images at a time before forgetting them.

For a family caregiver, memory is the whole job. Where the glasses went. Whether the noon pills were taken. That the stove was left on once last week. A helper that forgets that fast is not much help to a parent whose own memory is going. It becomes one more person to check on.

Half right on the hard tasks is real progress in a lab. In a home living with dementia, half right is a coin flip. Memory is the benchmark that matters.