In June 2026, joint research published in Nature Communications by Waymo and the Delft University of Technology introduced the Waymo Reference Driver (ReD) model to evaluate the safety performance of Waymo’s autonomous vehicles.
This framework establishes an objective behavioral benchmark by simulating how human drivers avoid collisions. Rather than relying on rigid, reactive parameters, the system models human cognitive processing to capture how a driver anticipates risks and manages road uncertainty. This approach allows developers to compare autonomous vehicle performance directly against a standardized human safety baseline within virtual simulation environments.
Inside Waymo’s human-inspired collision avoidance research
The joint research paper introduces a computational cognitive model that brings together human perception, decision-making, and vehicle control to improve autonomous vehicle safety. Developed by researchers at the Delft University of Technology and Waymo, the active inference framework views collision avoidance as a continuous effort to reduce cognitive surprise over time. Rather than focusing on fixed reaction times, the closed-loop system continuously updates a driver’s internal understanding of the road by interpreting the movements and intentions of surrounding vehicles. This framework accurately models how humans select, time, and execute defensive actions across three high-stakes scenarios: a leading vehicle braking suddenly, an oncoming vehicle swerving into the lane, and a vehicle failing to yield at an intersection.
To make the simulation more realistic, the model incorporates the perceptual and cognitive limitations that influence real-world driving behavior. For instance, visual perception relies on visual looming cues, which track how an object expands in a driver’s field of view, and includes a built-in detection threshold to explain why humans experience natural recognition delays at longer distances. To predict what other cars will do, the system uses a norm-conditioned particle filter. Under normal conditions, it assumes other drivers will follow basic traffic rules. When unexpected events occur, however, it adapts to predict highly unpredictable trajectories. Instead of instantly defaulting to rigid, pre-programmed maneuvers, the model constantly tracks a surprise signal that serves as evidence that the current path is no longer safe. A full defensive re-plan triggers only when this accumulated surprise crosses a specific threshold, factoring in realistic physical limits such as the time required to move a driver’s foot from the accelerator to the brake.
The next section explores how these technical capabilities are reflected in Waymo’s patent portfolio, examining the company’s historical patenting trends and the core innovations that support its autonomous driving technology.
Waymo: Patenting Activity
Waymo rapidly expanded its patent portfolio after becoming an independent Alphabet subsidiary in late 2016. The company increased annual priority filings from 340 in 2016 to 612 in 2017, before reaching a peak of 620 filings in 2019. During this period, Waymo aggressively protected its autonomous driving platform, filing patents covering proprietary lidar, radar, vision systems, vehicle control software, and self-driving technologies. The 2019 peak also coincided with the launch of Waymo One, marking the company’s transition from intensive technology development to early commercial deployment.
In 2021, the company implemented an executive leadership restructuring and secured a $2.5 billion investment round. During this period, the company expanded its robotaxi operations into additional cities, with engineering efforts increasingly focused on software refinement, fleet operations, and system integration to support commercial deployment.
Waymo: Top Jurisdictions
The United States, the European Patent Office (EPO), and China represent Waymo’s three largest patent filing jurisdictions. The company’s strong concentration of U.S. filings reflects its headquarters, primary engineering operations, and commercial autonomous vehicle deployments. The prominent placement of the European Patent Office reflects strategic alliances with European automotive manufacturers, such as an exclusive partnership with the Volvo Car Group and fleet vehicle integrations with Jaguar Land Rover. Meanwhile, filings in China support the company’s manufacturing and vehicle development partnerships, notably with the Chinese electric vehicle brand Zeekr, a subsidiary of Geely Automobile, which builds the platform for the Waymo Ojai robotaxi.
Waymo’s filings in Japan correspond to the company’s operational mapping and public road testing through partnerships with local mobility platforms, including GO and Nihon Kotsu. By securing patent protection in these secondary markets, Waymo mitigates infringement risks, protects its supply chains from localized replication, and strengthens its position for future technology licensing and cross-licensing agreements as autonomous vehicle markets continue to evolve.
Waymo: Top Technology Areas
Waymo’s largest patent application categories are integrated vehicle drive and hybrid control systems (B60W), radio navigation and wave-based detection systems (G01S), and image or video recognition or understanding (G06V). Together, these classifications highlight the company’s focus on the three core pillars of autonomous driving: vehicle control, environmental sensing, and perception. Applications in B60W describe motion control systems for steering, acceleration, and braking functions. Filings within G01S involve physical sensing hardware, including radar and lidar architectures. Classifications in G06V cover image analysis methods used to identify objects, traffic signs, and other road actors.
Beyond these leading categories, Waymo maintains substantial patent activity in computing arrangements based on specific computational models (G06N), image data processing or generation (G06T), and electrical digital data processing (G06F). These categories relate to machine learning algorithms and computational data pipelines that support autonomous driving decisions. The portfolio then extends to surveying, navigation, and distance measurement systems (G01C), traffic management and control systems (G08G), systems for controlling or regulating non-electric variables (G05D), and pictorial communication (H04N). These sections address spatial mapping, routing functions, mechanical actuator regulation, and data transmission interfaces, illustrating Waymo’s approach to autonomous vehicle development.
Waymo: Top Law Firms
Botos Churchill IP Law, McDonnell Boehnen Hulbert & Berghoff, and Liu, Shen & Associates handle the largest share of Waymo’s patent filings. This distribution reflects the volume of applications processed through domestic firms alongside technical oversight for vehicle manufacturing and co-development partnerships in China.
The remaining patent filings are distributed across Venner Shipley in the United Kingdom; TMI Associates through patent attorneys, Yoshiyuki Inaba & Mutsumi Sato in Japan; and Fish & Richardson, Banner Witcoff, and Lowenstein Sandler in the United States; Kim & Chang through patent attorneys Young Jun Yang, Minho Lee & Man-ki Baek in South Korea; and Reinhold Cohn and Partners in Israel,. Together, these firms manage patent prosecution across key international jurisdictions, helping Waymo navigate country-specific patent requirements while maintaining consistent protection for its autonomous driving technologies in major automotive and technology markets.
How Waymo’s patents enable autonomous driving
Waymo’s patent portfolio reveals how its autonomous vehicles perceive their surroundings, predict the behavior of other road users, and make driving decisions in real time. The following patents describe systems that compare expected and actual road conditions, evaluate multiple driving paths, and select the safest and most efficient action. Together, these inventions form the core decision-making framework behind Waymo’s self-driving technology.
Tracking unexpected environmental changes
U.S. Patent No. 12,565,233 describes a methodology for computing a backward-looking surprise metric in autonomous driving environments, specifying the calculation of a surprise score based on the statistical variance between a predicted trajectory and a subsequent observed state.
To execute this, the system extracts real-world log data from a manually driven vehicle and divides the recorded route into a series of short path segments. These segments serve as a baseline constraint, ensuring that when the autonomous software is tested, the simulated vehicle is forced to follow the exact geographic path taken by the human driver.
Next, the testing platform runs a simulation using the autonomous control software under the same environmental conditions and duration as the original real-world drive. Throughout the simulation, the autonomous software calculates its own operational metrics, including localized speed, acceleration, and vehicle orientation. The system then evaluates the driving behavior by comparing the total distance traveled or the time required to complete each segment against the human baseline data.
To isolate the exact cause of any performance differences, the system tracks external objects, such as other vehicles or pedestrians, that cross the simulated path. If the time an object intersects the route falls between the time metrics of the manual vehicle and the autonomous vehicle, that object is flagged as the cause of the discrepancy. These isolated performance differences and their identified causes are then used within a machine learning system to automatically adjust and optimize the software’s driving parameters.
This patent, titled “Using simulations to identify differences between behaviors of manually-driven and autonomous vehicles” was filed on November 8, 2023, and was granted on March 3, 2026. The listed inventors are Brian Cohen, Guillaume Dupre, Jared Russell, Eric Schoenfeld, and Jack Joseph Wanderman. Legal representation was provided by Botos Churchill IP Law.
Simulating and Ranking Driving Paths
U.S. Patent No. 10,254,759 describes a method for predicting how a driving environment will change after a vehicle performs a proposed maneuver.
The system uses a behavior model to generate predicted environmental observations, assigns quality scores based on the driving context, and ranks candidate driving paths to select the most appropriate one. To accomplish this, the framework takes the vehicle’s current environmental observations and a candidate action as inputs to a neural network, which predicts the surrounding environment after the maneuver is executed.
The system then feeds the predicted environmental observation and candidate action into a context-specific quality model, which evaluates factors such as safety and driving progress to generate an immediate quality score. It combines the initial observation, candidate action, and quality score into an experience tuple, a structured data record used to train the autonomous driving system.
The reinforcement learning framework uses these experience tuples to improve its decision-making across a wide range of simulated driving scenarios. By learning from these experiences, the system can evaluate, rank, and select the most appropriate driving path when navigating complex road environments.
U.S. Patent No. 10,254,759, titled “Interactive autonomous vehicle agent”, was filed on September 14, 2017 and was granted on April 9, 2019, with Aleksandra Faust, Matthieu Devin, Yu-hsin Joyce Chen, Franklin Morley, Vadim Furman, and Carlos Alberto Fuertes Pascual as the listed inventors. Michael Shepherd from Fish & Richardson provided legal representation for this patent.