A driver slows at a crosswalk. A pedestrian hesitates, catches the driver's eye, and steps out. A cyclist glances over a shoulder and drifts left. None of these people say a word, but they are communicating constantly: with eye contact, posture, speed and small gestures.
Human drivers learn this social layer over years. Autonomous vehicles have to learn it from data.
What the numbers show
Waymo, the most widely deployed robotaxi service in the U.S., has published safety comparisons against human driving. A peer-reviewed paper accepted in Traffic Injury Prevention, covering 56.7 million miles, reported that the Waymo Driver had 92% fewer injury-involving pedestrian crashes and 82% fewer cyclist crashes than human benchmarks.
By December 2025, Waymo reported 170.7 million fully driverless miles, with 92% fewer pedestrian injury crashes and 85% fewer cyclist injury crashes than human benchmarks. Those newer figures are company-reported.
The part that is hard to measure
Avoiding crashes is the floor, not the ceiling. A vehicle can be perfectly safe and still be a bad participant in traffic: too timid at a four-way stop, confusing to a pedestrian waving it through, unable to tell a person hailing a ride from a person waving to a friend.
These are social judgments. They depend on reading intent from people who are not following any script, and on behaving in ways that other people can read in return.
Why it matters for Social AI
Roads are one of the first places where AI and the public interact physically at scale. The same skills apply wherever machines move among people: sidewalks, warehouses, hospitals and homes. In each case, the machine needs to understand human signals and send clear signals back.
Learning those skills depends on real recordings of people interacting in shared physical spaces, captured from several viewpoints, with the timing preserved and with consent from the people recorded.
Crosstalk sources physical AI and social settings data, from egocentric video to in-person encounters, with consent from every participant. Explore the datasets or read the paper.
Sources: Insurance-Canada on the peer-reviewed study, Pulse 2.0 on 170 million miles, NBC Bay Area.