Washington University: Building Safer Autonomous Vehicles

In response to the growing scrutiny of autonomous vehicles, researchers at Washington University in St. Louis actively seek ways to enhance their safety. Computer engineer Yevgeniy Vorobeychik and architect Constance Vale have embarked on a project to create a scaled model of an urban neighbourhood for testing self-driving cars.

This model will offer an environment that can be customised to replicate scenarios that pose challenges for autonomous vehicles while ensuring people’s and property’s safety. The project, known as the “Architectural Design of an Experimental Platform for Autonomous Driving,” has led to the development of the “Washington University Miniature City” (WashU Mini-City).

The objective of the WashU Mini-City is to establish a tangible testing ground for autonomous driving that merges computer science and architectural principles, offering an affordable and secure setting for innovative experiments.

This miniature city aspires to function as a testing facility for conducting extensive assessments and surveillance of autonomous vehicles (AVs) to enhance their safety and dependability.

WashU Mini-City, occupying approximately 3000 square feet, boasts crafted building facades and streets with precise traffic markings, street lamps, stoplights, traffic cones, and billboards. A key feature is the inclusion of 3D-scaled pedestrians, some accompanied by their 3D canine companions, rendering it a faithful representation of the natural urban setting that autonomous vehicles (AVs) must navigate.

What sets WashU Mini-City apart is the high degree of modularity incorporated into nearly all its elements, allowing for easy relocation, replacement, or customisation based on specific experiments.

The attention to detail in design and construction is a notable achievement credited to Vale and Vorobeychik. For instance, the creation of lifelike trees was made possible through a sophisticated fabrication technique developed by Vale and her students, utilising digital and manual methods. These trees are crafted from vellum painted with translucent acrylic-painted strings, skillfully wrapped around laser-cut tree branches and slotted into 3D-printed tree trunks.

Enhancing the safety of autonomous vehicles (AVs) presents a challenge, particularly concerning what’s known as the “long-tail problem.” This problem revolves around rare and unconventional situations that test the capabilities of the driving algorithms governing these vehicles.

While AVs can be trained for routine tasks like stopping at traffic lights, they may struggle with unexpected scenarios, such as an escaped giraffe blocking the road. In such instances, human drivers instinctively pull over, stop the vehicle, and often share the experience on social media. Still, artificial intelligence (AI) tends to exhibit less predictability when confronted with extraordinary circumstances.

Using WashU Mini-City, Vorobeychik and Vale are addressing perceptual vulnerabilities, particularly those related to the long-tail problem. They consider factors like colour, sheen, and reflection crucial in machine sensing, making them challenging in modelling a miniature city for AV training.

Additionally, they’re exploring potential hacks, like tampering with road signs, which can deceive AV sensors. This controlled environment and 3D-printed pedestrians enable testing for scenarios involving a hostile actor’s attempt to manipulate the car’s behaviour, potentially causing accidents or endangering pedestrians.

WashU Mini-City facilitates cost-effective stress testing and malfunction engineering for AVs. Conducting experiments within this miniature environment is significantly more affordable and faster than full-scale tests, mitigating potential risks.

Furthermore, WashU Mini-City addresses the real-to-sim gap in AV research by providing a physical testbed. It offers insights that simulations cannot replicate, such as shading, lighting, and real-world traffic signals, contributing to a more comprehensive understanding of AV behaviour.

This model city also allows for easy modification of architectural and environmental factors, enabling researchers to evaluate their impact on AV sensors. For instance, adjustments to building features or adding simulated snow conditions inspired by techniques from the performing arts broaden the range of driving scenarios.

In the future, as WashU Mini-City grows, Vale and Vorobeychik plan to collaborate with academic and industry partners to enhance the project’s research potential. They aim to broaden the initiative and establish the Centre for Trustworthy AI in Cyber-physical Systems, with a vision of expanding beyond WashU. However, they acknowledge they are still in the early stages.