Why are driverless cars still so bad at driving? | Jennifer Dukarski | TEDxDetroit

Quick Overview

Driverless cars still struggle with complex driving scenarios, particularly those involving pedestrians of different skin tones or the unpredictable behavior of large vehicles like taco trucks, because the AI vision systems are predominantly trained on data sets that lack diversity, leading to a 10% difference in ability to detect dark-skinned people versus light-skinned people, highlighting the critical need for inclusive design in AI training data to achieve a truly safe and accessible autonomous future.

Key Points: Autonomous vehicle (AV) vision systems show a 10% poorer ability to detect dark-skinned people compared to light-skinned people in training data sets. The speaker, Jennifer Dukarski, an attorney and engineer, previously worked on steering columns and seating programs for automotive technology. A 2017 study by Georgia Tech found that AV systems detected pedestrians 20% more accurately when they were light-skinned versus dark-skinned. An incident involving a Waymo driverless vehicle in San Francisco showed the car stopped completely when a pedestrian in a wheelchair and a duck were present, demonstrating system hesitation in complex, unexpected scenarios. The speaker advocates for 'inclusion by design,' stressing that inclusivity must be foundational, not an afterthought, in developing AV technology. The data suggests AV systems are more likely to detect children than adults accurately, and the data sets collected are overwhelmingly white (over 80%).

Context: Jennifer Dukarski, who has an engineering degree and now works as a lawyer advising automotive companies on technology integration, presents at TEDxDetroit on the critical failures in artificial intelligence (AI) vision systems used in autonomous vehicles (AVs). She frames the discussion around the ethical and practical challenges of ensuring these systems can safely navigate real-world environments, particularly concerning diverse populations.

Detailed Analysis

Jennifer Dukarski argues that the promise of a perfect driving future via autonomous vehicles is hampered by inherent biases in the AI vision systems. She recounts her background in automotive engineering and her shift to law, focusing on technology. Dukarski highlights a critical finding from a 2017 Georgia Tech study: AV systems are significantly less effective at detecting pedestrians with darker skin tones compared to lighter skin tones, showing a nearly 10% difference in ability, and are better at detecting children than adults. She illustrates real-world issues with examples, such as a Waymo vehicle stopping completely in San Francisco when it encountered a pedestrian in a wheelchair and a duck, indicating system failure when faced with novel or complex inputs. The core message is the urgent necessity for 'inclusion by design'—meaning data sets used to train AI must be diverse across gender, age, race, and location (urban vs. suburban) to ensure equitable safety outcomes for all potential road users, rather than treating inclusivity as an afterthought.