Fast or Accurate? How Intention-Recognition Models Shape Human Perception of a Mobile Robot

Abstract

Robots that operate alongside people increasingly depend on intention-recognition models to anticipate human motion and adapt their behavior in socially appropriate ways. However, these models vary widely in both latency and accuracy, leading to different trade-offs between reacting quickly and correctly. Although these technical differences are well documented, it remains unclear how they shape the user’s experience of interacting with a robot. To examine how these translate into human perception, we conduct a preliminary user study comparing three intention-recognition models: a fast but low-accuracy model (Geo), an intermediate model (LSTM), and a slower but highly accurate model (Fusion). Participants interacted with a mobile robot controlled by each model and rated their experience across key dimensions of social interaction. Overall, the findings suggest that socially fluent interaction does not emerge from speed or accuracy alone, but from the balance of timely, reliable, and predictable robot behavior.