Human Navigational Intent Inference with Probabilistic and Optimal Approaches
Presented at ICRA 2022, this paper formalizes pedestrian intent inference as Bayesian goal inference, enabling robots to reason about human motion under uncertainty rather than relying on pattern matching alone. The probabilistic framework makes prediction uncertainty explicit and auditable - a property that matters whenever an AI system's decision to act (or not act) depends on uncertain predictions about human behavior.
A robot sharing space with humans needs to predict not just where people are, but where they are going and why. Simple velocity extrapolation fails in complex environments where people adjust routes around obstacles, other people, and social norms.
Framework
We formulate navigational intent inference as Bayesian goal inference: given a partial trajectory and a map, maintain a distribution over possible goal locations that explains the observed motion.
Likelihood model: We use an optimal planning model as the likelihood - a person heading to goal G would take a path close to optimal under their utility function. The likelihood of an observed trajectory under goal G is proportional to how well it approximates the optimal path to G.
Prior over goals: Semantic priors from the map (exits, commonly-used destinations) combined with uniform uncertainty over novel environments.
Posterior update: As the trajectory unfolds, the posterior over goals updates via Bayes' rule. Early in an interaction, the distribution is broad; as the trajectory reveals preference, it concentrates.
Key Properties
This framework makes uncertainty explicit in a useful way: the entropy of the posterior distribution over goals is a direct measure of how confident the system is about human intent. A robot can use this to make conservative decisions (wait, yield) when the posterior is diffuse and confident decisions when it has concentrated on a single goal.
The optimality assumption - people take near-optimal paths - is testable and calibratable. If the assumption is violated (a person wanders inefficiently due to distraction), the posterior will remain diffuse rather than producing a confidently wrong prediction.
Relation to AI Reliability
The design principle here is directly applicable to AI systems in regulated environments: a system should communicate uncertainty in a form that allows a human overseer to intervene. A diffuse probability distribution is more useful than a confident point estimate that happens to be wrong.