Waymo Patents a System That Predicts When Cyclists Are About to Lose Control

By Patentlyze Team · Updated May 22, 2026

Most self-driving safety systems react to people who are already in the road. Waymo's new patent is about something harder: predicting that a cyclist or pedestrian is *about* to lose control — and reacting before they hit the ground.

What Waymo's loss-of-control detection actually does

Imagine a cyclist ahead of your car suddenly wobbles, clips a curb, and goes down. By the time a human driver processes what happened, it may be too late to stop. Waymo is working on a system that watches for exactly that kind of loss-of-control moment — not after the fall, but as it's starting.

The idea is to track key points on a person's body — think shoulders, hips, head — and watch how their height changes relative to each other over time. A pedestrian stumbling, a cyclist tipping sideways, or a scooter rider going down all create a distinctive pattern in those height readings. When the system detects that pattern, it flags the person as a loss-of-control risk and tells the car to steer clear or slow down.

The system also uses a separate machine learning model to assign a "prone state score" — essentially a probability that someone is in the process of falling flat. That score feeds into the same decision to trigger an avoidance action. It's a two-pronged approach: track the geometry of the body, and separately estimate whether a fall is in progress.

How the system tracks body height to predict a fall

The patent describes a perception pipeline built into Waymo's autonomous vehicle stack that specifically targets vulnerable road users (VRUs) — pedestrians, cyclists, scooter riders, and similar — who are in danger of losing control of their movement.

Here's how the pipeline works:

• The vehicle's sensing system (likely LiDAR, cameras, or a combination) collects environmental data continuously.
• One or more machine learning models process that data to identify a set of reference points on a VRU's body — think of these as a simplified skeleton or keypoint map.
• The system calculates height differentials between those reference points — the vertical distance between, say, a person's head and their center of mass — and tracks how those differentials change over time.
• A separate ML model outputs a prone state score (a confidence estimate that the person is falling or has fallen horizontal), which can be combined with the height-differential signal.

If the system determines that the height changes and/or prone state score indicate a loss-of-control risk, it sends a command to the vehicle's control system to perform an avoidance action — slowing, steering, or stopping. The claim is deliberately broad about what sensing modality is used, making it applicable to LiDAR point clouds, camera frames, or fused sensor inputs.

Why catching a fall before it happens changes AV safety

Most AV perception systems are optimized to track where people are and predict where they're going based on their current trajectory. A person who's upright and walking is easy to model. A person who's suddenly collapsing or flying off a bike is much harder — their trajectory becomes chaotic and unpredictable in a fraction of a second. Catching that transition early is the whole point of this patent.

For Waymo's robotaxi operations in dense urban environments, this is a genuinely practical problem. Cities have cyclists, e-scooters, and pedestrians on uneven terrain. If your AV can detect a wobbling cyclist before they hit the ground, you've bought yourself an extra second or two of reaction time — which at low speeds is often the difference between a near-miss and a collision. This kind of proactive safety logic is what separates a well-tuned AV from a system that merely reacts to obstacles.

Editorial commentary on a publicly published patent application. Not legal advice.