Perception — seeing the cars, pedestrians, and cyclists around you — is the part of self-driving that gets demonstrated. Prediction is the part that is actually hard. Knowing where another car is now is comparatively easy; knowing where it will be in three seconds requires guessing what its driver intends to do, and intentions are invisible. A 2025 Waymo grant tackles this head-on, and its approach reveals how the problem is really framed.
The record: on June 10, 2025, Waymo LLC was granted US12325452B1, “Agent trajectory prediction using dynamic candidate future intents.” The CPC classes are autonomous-driving prediction classes — B60W 60/00274 and 60/0011 (autonomous decision functions), B60W 40/04 and 40/06 (estimating road and traffic conditions), and the learning class G06N 3/084. The key phrase is “candidate future intents” — plural, dynamic.
Here is the insight in the design. A naive predictor outputs one guess: that car will continue straight. But other drivers are ambiguous — a car approaching an intersection might go straight, turn, or stop, and you do not know which. Predicting a single trajectory and being wrong is dangerous. Instead, this approach reasons over multiple candidate intents simultaneously, each with its own likely trajectory, and the candidates are “dynamic” — they update as new evidence arrives. The system holds several possible futures at once.
Why is multi-intent the right framing? Because driving safely among humans means planning for what they might do, not betting on what they probably will do. If there is a real chance the car ahead brakes, you must leave room for it even if continuing is more likely. Predicting a distribution of intents — and the trajectories each implies — lets the planner account for the dangerous possibilities, not just the expected one. It is prediction built for safety, not just accuracy.
Trace it to the product and the significance is the heart of why autonomy is hard in mixed traffic. A self-driving car shares the road with unpredictable humans, and its safety depends on anticipating them well. Waymo — which operates actual driverless vehicles in real cities — patenting dynamic multi-intent prediction is a window into the machinery that makes that operation possible: not a single confident guess, but a constantly updated set of what-ifs.
The skeptic's caveat: a granted prediction method is a technique, not proof of safe behavior in every situation, and prediction remains genuinely unsolved — humans surprise the best systems. The demo shows the car reacting; the deployment has to anticipate, which is far harder. But the framing is honest about the problem. Coordination with human drivers is the quiet frontier of autonomy, and a 2025 Waymo grant shows the real work happening there — in reasoning about intentions you cannot see.