Waymo Patents an AI That Predicts Collisions by Mapping Every Object's Next Move
Waymo's latest patent describes a system that doesn't just track where other cars and pedestrians are right now, it predicts where they'll be, then checks whether your car's own predicted path crosses theirs.
How Waymo's car spots a crash before it happens
You're in a self-driving car cruising through an intersection when a cyclist cuts left across the road. By the time most systems register the danger, it may already be too late to stop smoothly.
Waymo's patent describes an AI that runs ahead of that problem. Instead of only watching where things are at this moment, the system draws a predicted future path for both the car and every object around it at the same time. If those two paths are going to cross, the car's planning system gets an early warning and adjusts speed or direction before the conflict becomes an emergency.
The system pulls together data from multiple sensors, which means it isn't guessing from a single camera or lidar beam alone. It combines those sensor readings into a shared picture of the environment and then uses that picture to run its forward-looking predictions. The result is a car that effectively asks "where is everyone going?" rather than just "where is everyone now?"
generating, using a machine learning model (MLM) and one or more embeddings characterizing an environment around an autonomous vehicle (AV), a first polyline representing a predicted future trajectory of the AV and a second polyline representing a predicted future trajectory of an object in the environment around the AV …
Translation: An AI uses surroundings data to draw future path lines for both the robot car and nearby objects.
How the model plots two paths and compares them
The core of the patent is a machine learning model that takes in sensor data from around the autonomous vehicle and produces what the filing calls polylines, essentially plotted lines on a map showing the predicted future route of the car itself and each nearby object (other vehicles, pedestrians, cyclists, and so on).
Once both trajectories are drawn, the system checks mathematically whether they intersect or come dangerously close within a relevant time window. If a potential collision is detected, a signal goes to the vehicle's planning system (the module that decides speed, steering, and braking) so it can reroute or slow down.
A key detail is that the model works from embeddings, compact, structured representations of the environment built by fusing data from multiple sensors (camera, lidar, radar, or some combination). Rather than each sensor working separately, their inputs are combined into one unified environmental description before prediction happens.
- Sensor data is fused into a shared environment model
- The ML model generates a predicted path for the AV and for each detected object simultaneously
- The system checks whether any predicted paths intersect
- The planning module adjusts driving behavior proactively if a collision risk is found
… determining, based on the first polyline and the second polyline, a potential collision between the AV and the object, and causing a planning system of the AV to adjust an operation of the AV to avoid the potential collision with the object.
Translation: The system checks if those predicted lines cross and steers the vehicle out of danger.
What this means for self-driving safety decisions
Collision avoidance in self-driving cars has historically been reactive, the car detects something close and brakes hard. Predicting future trajectories for both the car and surrounding objects at the same time allows the system to act earlier, with smaller and smoother corrections rather than panic stops.
For everyday riders, this means fewer jarring emergency maneuvers and, more importantly, fewer situations where no maneuver would have been enough. Waymo's consistent filing in multi-sensor prediction suggests the company sees this layer of forward-looking awareness as a foundation for higher-speed and more complex driving environments, not just city crawls.
That makes this Google's 54th filing we've tracked since May in our self-driving sensing race watchlist, building on work like laser-based weather detection and driver reaction time prediction.
Intersection collisions, pedestrian strikes, and sideswipes often unfold in under two seconds, faster than any driver can meaningfully respond once the danger is already visible. A self-driving car that cannot reliably anticipate where every nearby person and vehicle is heading will always be chasing events rather than preparing for them.
Waymo's approach addresses that gap by predicting the car's own future path and the paths of surrounding objects together, in a single calculation, rather than as separate tasks stitched together afterward. That matters because conflicts between two moving things only become visible when you model both at once.
The patent lays out a credible structure for doing this, but how accurately those forecasts hold up when a cyclist swerves or a driver cuts a corner remains the real test, and that answer lives in performance data rather than a filing.
Get our take in your Top Stories
Liked this breakdown? Add Patentlyze as a preferred source on Google, and our plain-English take shows up more often in your Top Stories the next time Google patent news breaks.
There are more where this came from
We read every patent application Big Tech publishes and send you the ones worth knowing. Plain English, free, every week.
The drawings
10 drawing sheets from US 2026/0264723 A1 · click any drawing to enlarge
Show the other 2 sheets
Want this weekly breakdown for a company we don't cover? Patentlyze Pro →
Be the first to weigh in