The Instructive Failure

A Furhat social robot served coffee in an over-50s housing facility for five weeks. Repeat interaction was low. Most residents tried the robot once or twice and then stopped. The deployment was, by standard HRI metrics, a failure.

Wilson, Robb, Lim, Hastie, Aylett, and Georgiou (arXiv:2603.16336) make the failure modes the primary finding. The social dynamics of the housing complex overrode individual curiosity about the robot. Who watched whom interact with the machine mattered. A technical breakdown witnessed by neighbors created lasting avoidance — not because the breakdown was severe, but because being seen with a malfunctioning robot was socially costly. The physical accessibility of the common room, the social hierarchy among residents, and the implicit meaning of “using the new technology” in a community where technological competence signals independence — all shaped adoption more than the robot’s actual capabilities.

In controlled lab studies, these dynamics are invisible. The lab isolates the human-robot dyad from its social context, producing clean interaction data that tells you about the robot and the individual but nothing about the environment. The housing facility study produces messy, mostly negative data that tells you about the environment — which is where robots will actually live.

Every HRI study that “works” tells you about the robot. This study’s failure tells you about the world. The conditions for adoption are not technical (the robot worked when it worked) and not individual (most residents were curious enough to try). They are social, spatial, and reputational. The failure is not a noise term to be reduced; it is the signal that controlled studies are designed to suppress.

The flaw in the deployment is the feature of the research. You can only learn what kills adoption by attempting adoption and watching it die.