Safety

Waymo’s approach to building a reliable case for safety that can serve as a toolkit that any AV developer can use to inform their own safety case.

Grouping collisions with similar characteristics and causes is how traffic safety researchers measure effectiveness of different technologies, like autonomous vehicles. This paper introduces a comprehensive conflict typology that can be applied to both human and AV crash and near-crash data.

The potential severity in a collision is affected by braking and steering. The metric developed in this paper predicts the worst case probability of injury as if there was no avoidance maneuvers. This metric is important when comparing different crash datasets.

Our proposed Fatigue Risk Management Framework addresses prevention, monitoring, and mitigation of fatigue-induced risks that affect autonomous specialists during testing of the Waymo Driver.

Our framework to analyze and model response timing in a crash-imminent situation on the road.

Our Safety Framework – the careful and multilayered approach to safety that has made it possible for Waymo to deploy fully autonomous driving technology on public roads.

A study examining how well the Waymo Driver avoids collisions by using a human behavior reference model — the response time and evasive action of a human driver that is non-impaired, with eyes always on the conflict (NIEON).

Predicting injury risk in a vehicle collision is important for autonomous vehicle design, when most collisions happen in virtual simulations. This paper provides an injury risk function that is continuous in every direction that is more suitable for simulation use than previous injury risk curves.

Autonomous driving technology holds the promise to improve road safety and offer new mobility options to millions of people. The Waymo Safety Report, which was originally published in 2017, provides a high-level, general overview of our system safety program, how our fleet of fully autonomous vehicles work, our approach to testing and validation, and how we interact with the public.

This study compares all Waymo crashes reported under NHTSA’s Standing General Order (SGO) over 7+ million rider-only miles driven through the end of October 2023 in Phoenix, San Francisco, and Los Angeles to comparable human benchmarks.

An assessment of all contact events experienced by the Waymo Driver during its first 1M miles without a human behind the wheel supporting that the Waymo Driver is successful at reducing injuries and fatalities.

Our scenario-based testing methodology to evaluate the Waymo Driver's behavior in conflict situations initiated by other road users.

A study that explores the safety of the Waymo Driver using simulated versions of real-world events. Our latest paper looks at data on fatal human-driven crashes that occurred within our operating domain in Chandler, Arizona between 2008 - 2017.

Our Public Road Safety Performance Data whitepaper includes details about the miles we’ve driven on public roads in Arizona to provide data about our safe operations in practice.

A real-world case study of the Waymo One service that compares human drivers against the Waymo Driver, using Swiss Re’s property damage liability and bodily injury claims data from 2016-2021 and Waymo’s rider-only data collected through August 1, 2023. The analysis shows that the Waymo Driver is significantly and consistently safer than human drivers when it comes to crash causation.

This paper aims to ensure a fair comparison between autonomous and human driving by addressing the most common errors and biases and establishing valid benchmarks from the cities in which Waymo operates.

This paper presents a simplified framework for abstracting and organizing the current landscape of ADS safety standards into high-level, long-term themes. This framework is then utilized to develop and organize associated research questions that have not yet reached widely adopted industry positions, along with identifying potential gaps where further research and standardization is needed.

This paper calls for a diversified approach to safety determination beyond the analysis of safety outcomes. It highlights a “credibility paradox” within the comparison between ADS data and human-derived baselines, and speaks to continuous confidence growth in ADS performance estimates

This paper presents, for the first time, how computational models of surprise rooted in cognitive science and neuroscience combined with state-of-the-art machine learned generative models can be used to detect surprising human behavior in complex, dynamic environments like road traffic. We also present novel approaches to quantify surprise and use naturalistic driving scenarios to demonstrate a number of advantages over existing surprise measures from the literature.

The papers explore a novel way to model human road user behavior by means of active inference, an approach adopted from contemporary cognitive neuroscience. Specifically, the papers show how active inference can provide a middle ground between glass-box mechanistic models and black-box machine-learned models, generating versatile, yet interpretable behavioral outputs.

To gain practical insight into the types of challenges and limitations arising from executing traditional consumer-focused testing protocols to an ADSs, the Waymo Driver was the subject of a testing campaign that leveraged several of the most difficult currently available ADAS and active safety test procedures. The Waymo Driver was able to pass all of the test procedures and the main challenges discovered were that most procedures that are designed to evaluate collision avoidance behavior could not be evaluated as designed due to the increased capabilities of the Waymo Driver that prevented the vehicle from even entering into a conflict.

Assessing injury severity for collisions involving vulnerable road users (VRUs) is highly impactful for the continued development of traffic safety, including ADAS, ADS, and roadway design. Using naturalistic VRU collision data collected from dashboard cameras of third-party vehicles, a methodology for assessing event severity by pairing accelerometer and GPS data with video to compute impact speed was presented.

The papers explore a novel way to model human road user behavior by means of active inference, an approach adopted from contemporary cognitive neuroscience. Specifically, the papers show how active inference can provide a middle ground between glass-box mechanistic models and black-box machine-learned models, generating versatile, yet interpretable behavioral outputs.

Understanding how drivers manage uncertainty is key for developing simulated human driver models for autonomous vehicles. A generalizable, interpretable, computational model for adaptive human driving behavior is needed to accomplish this, but existing models lack computational rigor or only address specific scenarios. This paper proposes such a model based on active inference, a behavioral modeling framework originating in computational neuroscience.

There is an increasing need to comprehensively characterize the kinematic performances of different Micromobility Vehicles (MMVs). Using a variety of test track experiments, this study (1) characterizes the kinematic behaviors of different MMVs during emergency maneuvers; (2) explores the influences of different MMV power sources on the device performances; and (3) investigates if piecewise linear models are suitable for modeling MMV trajectories.