Hugging Face researchers have released a preprint extending their earlier work on pedestrian archetypes, a framework designed to improve how autonomous vehicles (AVs) anticipate and respond to real-world pedestrian behavior. The original taxonomy, published in a prior paper, defined 12 archetypes such as the Wanderer, Drunk, Distracted, and Jaywalker, moving beyond simple behavior labels to capture progressive and often dangerous actions.
In this extension, the team analyzed additional YouTube dash-cam footage and identified seven new archetypes that exhibit significant behavioral differences not covered by the original set. These new categories address patterns that could not be fully explained by the previous framework, filling critical gaps in AV safety testing scenarios.
For each new archetype, the paper defines essential and optional behaviors, contrasts them with existing archetypes, and provides video-frame evidence to illustrate the behavior in action. This detailed approach aims to give AV developers a more nuanced toolkit for simulating and testing edge-case pedestrian interactions.
The work builds on the original Pedestrian Archetypes paper published in IEEE (document 11097414) and is part of ongoing efforts to enhance the realism and safety of autonomous driving systems.