Automated Driving - Why is it so hard to introduce

Why can a self-driving car mistake the moon for a traffic light? An Audi engineer explains the surprising sensor challenges holding back automated driving.

#1about 3 minutes

Defining the core challenge of automated driving

The fundamental question of why fully automated cars are not yet available is explored, setting the stage for a discussion on technical and organizational challenges.

#2about 2 minutes

Audi's history of automated driving concepts and studies

Audi's journey with automated driving includes early concepts like the A8, which introduced laser scanners but revealed perception system limitations for true Level 3 responsibility.

#3about 2 minutes

The crucial difference between assisted and automated driving

Assisted driving (Level 2) is a mature technology where the driver is responsible, while automated driving (Level 3+) is a disruptive shift that transfers responsibility to the car.

#4about 2 minutes

Understanding the automation dilemma in vehicle control

As automation capability increases, the driver's ability to manage system errors decreases, making the leap from Level 2 to Level 3 a fundamental challenge rather than an incremental improvement.

#5about 2 minutes

Why evolving assisted driving systems fails for automation

The automotive industry's attempt to incrementally improve assisted driving (Level 2) into automated driving (Level 3) is flawed because they are fundamentally different problems requiring separate development approaches.

#6about 2 minutes

The challenge of proving a positive risk balance

To release an automated system, carmakers must prove it's safer than a human driver by satisfying numerous criteria including safety KPIs, functional safety, and regional traffic laws.

#7about 3 minutes

Managing the endless complexity of real-world corner cases

Automated systems must handle a long tail of rare and complex scenarios, suggesting a need for shared, multi-layered knowledge bases and marketplaces for scenario data.

#8about 1 minute

Using prediction models to measure system performance

Prediction models help track the 'corner case rate' over time, showing that finding the most difficult and rare system errors becomes exponentially harder as development progresses.

#9about 3 minutes

Why environment perception is the weakest link

Environment perception is a major challenge because sensors react to the entire complex world, making it difficult to create an analytical model of their behavior and reliably detect anomalies.

#10about 3 minutes

Real-world examples of sensor interpretation failures

Practical examples, such as misinterpreting the moon as a traffic light or camera blindness from sudden light changes, highlight the profound difficulty of reliable sensor data interpretation.

#11about 2 minutes

The challenge of closed-loop system behavior and feedback

A system's reaction to sensor data can create complex feedback loops, where the car's movement alters sensor input and confirms false positives, a problem only discoverable in closed-loop testing.

#12about 2 minutes

Adopting agile systems engineering for complex development

The complexity of automated driving requires a shift to agile systems engineering and product-oriented organizational structures, such as cross-functional 'event chain teams' with end-to-end responsibility.

#13about 2 minutes

The multidisciplinary future of automated driving development

The future of automated driving lies in the successful combination of diverse fields like systems engineering, safety, software development, and data science to solve these complex challenges.