Why does Chris Urmson believe Level 2 autonomy systems are problematic?

Urmson points to the human factor of over-trust and complacency, where drivers become less vigilant over time. He also highlights the problematic marketing and divergent technological paths that Level 2 systems take compared to the development of true, full autonomy.

What is Aurora's approach to demonstrating the safety of their autonomous vehicles?

Aurora focuses on diligent and thorough engineering, incorporating functional safety processes. They aim to demonstrate capabilities through simulations, testing, and on-road data, and will engage deeply with regulatory bodies like NHTSA to gain their confidence.

What are the biggest challenges in developing fully autonomous vehicles?

The hardest part is dealing with the unpredictability of real-world actors, including pedestrians, cyclists, and other vehicles whose behavior is not always intended. The scale of operation in complex urban environments is also a significant challenge.

How does Chris Urmson view Elon Musk's stance on lidar?

Urmson acknowledges that driving is possible with passive vision (cameras), but argues that lidar is a valuable tool in the pursuit of robust safety. He believes any technology that accelerates the arrival of life-saving autonomous vehicles should be used, rather than setting arbitrary limitations.

When does Chris Urmson expect to see large-scale deployment of fully driverless vehicles?

He is confident that significant scale, with thousands of driverless vehicles operating continuously without safety drivers, will be achieved within the next ten years, likely starting in complex urban and suburban environments.

What is the most critical breakthrough needed to accelerate autonomous vehicle deployment?

Urmson identifies the car's perception and forecasting capability as the key. A perfect model of what is happening and what will happen in the next five seconds around the vehicle would dramatically speed up development.

What are the main competitors to Aurora, and how do they aim to out-compete them?

Aurora focuses on its mission and leverages the significant experience of its leadership team. They attract top talent by offering a compelling mission and a culture that allows focus on problem-solving, supported by investments in infrastructure and machine learning architectures.

Key Moments

Chris Urmson: Self-Driving Cars at Aurora, Google, CMU, and DARPA | Lex Fridman Podcast #28

Lex Fridman

Science & Technology4 min read45 min video

Jul 22, 2019|38,317 views|898|49

Save to Pod

Key Moments

TL;DR

Aurora's CEO discusses self-driving car tech, from DARPA to AI, safety, and industry challenges.

Key Insights

The DARPA Grand and Urban Challenges proved that autonomous driving was possible, despite initial skepticism.

Technological advancements like HD mapping and multi-beam lidar were crucial for early progress in self-driving systems.

Perception in autonomous vehicles has evolved significantly, now needing to handle unpredictable human behavior and diverse road users.

Lidar is considered essential by some experts, despite Elon Musk's view of it as a 'crutch,' as it contributes to sensor fusion for robustness.

Level 2 autonomous systems present significant human factor challenges due to over-trust and complacency, potentially necessitating a divergence from full autonomy development paths.

Demonstrating the safety of autonomous vehicles requires rigorous functional safety processes, a combination of testing (simulation, unit, on-road), and regulatory approval.

Large-scale deployment of driverless vehicles is anticipated within 10 years, likely starting in urban/suburban environments due to more frequent learning opportunities with lower risk.

Predicting the future behavior of other road users, especially vulnerable ones like pedestrians and cyclists, remains a complex technical challenge for autonomous systems.

FROM DARPA CHALLENGES TO REAL-WORLD POSSIBILITY

Chris Urmson, a pivotal figure in autonomous vehicle development, reflects on his early involvement in the DARPA Grand and Urban Challenges. These events, while seemingly impossible at the time, served as crucial proving grounds. The key takeaway was the demonstration that autonomous driving could, in fact, be achieved. This success was fueled by a combination of technical ingenuity and a willingness to tackle daunting challenges, inspiring confidence that the problem, though complex, was solvable.

KEY TECHNOLOGICAL LEAPS IN AUTONOMY

The evolution of self-driving technology can be marked by significant technical innovations. The Grand Challenge benefited greatly from High-Definition (HD) mapping, which provided detailed environmental models, allowing vehicles to navigate at speed by reducing the inherent complexity of the driving task. For the Urban Challenge, the advent of multi-beam lidar was a game-changer, enabling high-resolution 3D environmental modeling for better perception and localization, moving beyond reliance solely on GPS.

THE GROWING COMPLEXITY OF PERCEPTION AND PREDICTION

As autonomous systems advance, the perception and prediction capabilities required have become exponentially more complex. While early challenges involved static environments or predictable actors, today's systems must contend with the unpredictability of human drivers, cyclists, and pedestrians. Understanding and forecasting the behavior of these diverse road users, especially in dynamic urban settings, is a paramount and evolving challenge for ensuring safety and robust operation.

THE ROLE OF SENSORS AND THE LIDAR DEBATE

The debate surrounding sensor suites, particularly the role of lidar, was discussed in light of Elon Musk's critique. Urmson emphasizes that while cameras are essential, a fusion of data from lidar, cameras, and radar is critical for achieving true robustness. He frames lidar not as a crutch but as a valuable tool among others, arguing that any technology that accelerates the deployment of safer autonomous vehicles and reduces road fatalities should be embraced, regardless of its form.

CHALLENGES WITH AUTONOMOUS DRIVING LEVELS AND HUMAN FACTORS

Urmson expresses significant concern regarding human factors, particularly the over-trust and complacency associated with Level 2 and Level 3 autonomous systems. He believes that the marketing and public understanding of these systems often lead to dangerous misconceptions. The economic incentives for developing driver-assistance systems diverge from those required for truly driverless vehicles, potentially creating a schism in technological development and leading to safety concerns if not managed thoughtfully.

DEMONSTRATING SAFETY AND REGULATORY APPROVAL

Establishing the safety of autonomous vehicles is a multifaceted endeavor. Urmson highlights the necessity of a robust functional safety process, rigorous testing across simulations and real-world data, and transparent communication. Collaboration with regulatory bodies like NHTSA is key, as their approval, based on thorough evidence of capability and safety, will be critical for public acceptance and large-scale deployment of autonomous systems.

THE FUTURE OF AUTONOMOUS VEHICLE DEPLOYMENT

Urmson is confident that large-scale deployment of driverless vehicles, operating without safety drivers, will occur within the next decade. He anticipates this will likely begin in urban and suburban environments. While freeways might seem simpler, the lower speeds and frequent interactions in cities offer more opportunities for learning with reduced consequences, making them an ideal proving ground for establishing robust autonomous driving capabilities before scaling to higher-speed highway environments.

CRITICAL TECHNICAL HURDLES AND VULNERABLE ROAD USERS

Looking ahead, Urmson identifies the perceptual forecasting capability—accurately predicting the immediate future actions of all surrounding road users—as the most critical technical hurdle. He also expresses particular concern for vulnerable road users like pedestrians and cyclists, who lack the protection afforded to occupants of vehicles. The game-theoretic interactions and ensuring safety around these users, without creating undue risk or overly cautious behavior, represents a significant algorithmic and experiential challenge.

Mentioned in This Episode

●Products

●Software & Apps

●Companies

●Organizations

●Concepts

●People Referenced

Autonomous Vehicle Development and Deployment Guide

Practical takeaways from this episode

Do This

Embrace ambitious, hard challenges as opportunities for innovation.

Empower individuals by trusting their potential, not just their current abilities.

Utilize a comprehensive sensor suite (lidar, cameras, radar) for robust perception.

Focus on delivering full autonomy, as Level 2 systems have inherent human factor challenges.

Prioritize safety through rigorous engineering, validation, and functional safety processes.

Engage with regulators and trusted bodies to demonstrate safety.

Win public trust by allowing people to experience safe autonomous vehicles firsthand.

Develop systems that are capable of accelerating deployment and delivering value.

Focus on the perception and forecasting capabilities of the vehicle itself.

Avoid This

Underestimate the complexity of autonomous driving.

Rely solely on naive trust in GPS or single sensor modalities.

Market or design systems that encourage over-trust or complacency (e.g., ambiguously marketed Level 2 systems).

Solely focus on the cheapest sensor suite; prioritize a robust and economically viable one.

Use simplistic marketing metrics like disengagement numbers that can be easily gamed.

Ignore the human factors and potential for over-trust in driver assistance systems.

Assume freeways are inherently easier than urban environments for autonomous driving.

Neglect the challenges posed by vulnerable road users (pedestrians, cyclists).

Common Questions

The Grand Challenge's key innovation was HD mapping, enabling vehicles to operate at speed by providing detailed environmental models. The Urban Challenge saw the significant impact of multi-beam lidar for 3D world modeling and advancements in Bayesian estimation for vehicle tracking and prediction.

Topics

Ai-Ethics Human Performance AI & Machine Learning Technology & Innovation Public Perception Autonomous Driving Machine Learning Future Of Transportation Sensor Fusion Safety Engineering

Mentioned in this video

Companies

Uber

Mentioned in relation to Drew Bagnell's previous role.

Tesla

Mentioned in relation to Sterling Anderson's previous role as director of Autopilot.

Aurora Innovation

Autonomous vehicle software company founded by Chris Urmson, Sterling Anderson, and Drew Bagnell.

Google

Former employer of Chris Urmson, where he led the self-driving car team.

People

Red Whittaker

Leader of the CMU autonomous vehicle teams in the DARPA Challenges and a significant influence on Chris Urmson.

Drew Bagnell

Co-founder of Aurora Innovation and former autonomy and perception lead at Uber.

Elon Musk

Mentioned for his provocative statement that lidar is a crutch and that perception tasks can be done with cameras.

Chris Urmson

CEO of Aurora Innovation, former CTO of Google's self-driving car team, and a key leader in CMU's DARPA Grand and Urban Challenge entries.

Sterling Anderson

Co-founder of Aurora Innovation and former director of Tesla Autopilot.

Lex Fridman

Host of the podcast.

Trevor Noah

Mentioned for a segment where he discussed Tesla's self-driving capabilities, highlighting public misconception.

Products

multi-beam lidar

A key technological innovation for the Urban Challenge, enabling high-resolution 3D models of the world for environmental understanding.

GPS

Global Positioning System, a technology that was previously relied upon but was superseded by more robust localization methods.

Tesla Autopilot

A driver assistance system that has been subject to public over-trust and marketing challenges, according to Urmson.

Concepts

HD mapping

High-definition mapping was a key technology that unlocked performance in the Grand Challenge by bounding the complexity of the driving problem.

Bayesian estimation

A statistical technique used in robotics, including SLAM and state estimation for tracking vehicles.

trolley problem

A philosophical thought experiment about ethics and decision-making in unavoidable accident scenarios, often misapplied to autonomous vehicle dilemmas.

Organizations

Carnegie Mellon University

Where Chris Urmson was a graduate student and led autonomous vehicle entries in the DARPA Challenges.

NHTSA

The National Highway Traffic Safety Administration, a US regulatory body that Chris Urmson expects to engage with regarding safety demonstrations.

Found this useful? Build your knowledge library

Get AI-powered summaries of any YouTube video, podcast, or article in seconds. Save them to your personal pods and access them anytime.

Try Summify free