The State of Self-Driving Cars on the Roads Today: Safety and Challenges

Today, we explore the current landscape of self-driving cars and their safety concerns. This article delves into the progress, limitations, and hurdles faced by companies like Google in their pursuit of fully autonomous vehicles.

Current Fleet and Testing Environments

As of now, the largest fleet of self-driving cars is operated by Waymo, with around 600 Level 4 (L4) cars running on public roads. Other companies such as Tesla pursuing Level 2 (L2) autonomy keep their vehicles on test tracks. While there are various advanced driver assistance systems (ADAS) running on public roads, they do not qualify as true self-driving cars according to most definitions.

The Complexity of Edge Conditions

Designing self-driving cars to handle everyday driving scenarios is one thing, but dealing with unusual circumstances or 'edge conditions' presents a significant challenge. Unusual driving situations, such as navigating a rainy night with a pothole marked by a non-standard, makeshift barrier, are much harder to address.

To illustrate, Tesla and other companies are actively training their AI systems to handle these complex scenarios. Despite their progress, the job remains far more difficult than initially anticipated. The ability to handle edge conditions is crucial for the safety and reliability of self-driving cars.

Observations in Mountain View, California

In my neighborhood in Mountain View, California, one can witness numerous autonomous vehicles on the streets. The most prominent are the Waymo Google cars, and newer additions like Toyota Research’s vehicles. Unmarked cars with sensors, believed to be stealth research projects, are also spotted. Much of this activity involves a human driver collecting data for machine learning algorithms.

Self-Driving Car Safety and Capabilities

From available public evidence, self-driving cars are not yet safe. The core challenge lies in achieving accurate and reliable computer vision. While machines can see better than humans, using multi-spectral sensors and combining inputs from a wide range of optical, thermal, and RF sensors, they still lack intelligence.

To highlight the complexity, imagine a two-year-old child learning to recognize a 'laser shop' after seeing one in a video store. The child learns based on a single, highly detailed example. Current AI systems, in contrast, need millions of labeled examples with significantly fewer features to achieve similar accuracy. This gap suggests that significant advancements are still necessary before machines can match human-like recognition ability.

Challenges in Machine Learning and AI

Machine learning and AI capabilities are far from perfect. Google Translate, though backed by massive data, often produces poor translations. Similarly, despite impressive feat in word recognition, virtual assistants like Alexa or Google Home struggle with understanding sentences. This highlights the difficulty in machines comprehending situational context, akin to computer vision's challenge in identifying a road's endpoint based on its visual elements.

While attempting to search for photos using object names on Google Photos often results in frustratingly incorrect matches, this underscores the limitations in current AI systems. Even Google's own terms and conditions caution users about the unreliability of their software.

Given these challenges, it is advisable to exercise caution when considering self-driving car technology. While the future may hold more sophisticated and reliable autonomous vehicles, the current state of the technology warrants careful consideration and continued vigilance.