How did Waymo and Tesla perform in autonomous driving milestones in 2018?

Waymo reached 10 million autonomous miles, a significant step for fully autonomous vehicles. Tesla's semi-autonomous Autopilot system surpassed 1 billion driven miles, primarily using camera-based computer vision.

What are the biggest challenges facing the widespread adoption of autonomous vehicles?

Beyond technological hurdles, the adoption will depend on creating a better human experience, not just superior safety or speed. Autonomous vehicles are currently slower and more cautious, and the economic viability in the near term is still being established.

What are the different levels of autonomy and the key distinctions?

Autonomy ranges from Level 0 (no automation) to Level 5 (full automation). The key distinction is between human-centered autonomy (driver is responsible) and full autonomy (car is responsible legally, experientially, and algorithmically), which does not allow for teleoperation or simple ten-second rule takeovers.

What are the two main technological approaches for autonomous vehicle perception?

The two primary approaches are vision-based systems using cameras and deep learning, which offer rich data but struggle with accuracy and extreme conditions, and Lidar-based systems, which are more accurate and reliable but significantly more expensive and less data-intensive for training.

How do radar and ultrasonic sensors fit into autonomous vehicle systems?

Radar is crucial for obstacle detection and avoidance in various weather conditions due to its range, but has low resolution. Ultrasonic sensors excel at detecting proximity for actions like parking and complement radar well within sensor fusion packages.

What is the significance of the DARPA Grand Challenges for autonomous vehicles?

The DARPA Challenges, particularly the 2005 Race to the Desert and the 2007 Urban Grand Challenge, were pivotal in showcasing the potential of autonomous vehicles and spurred significant investment and research, demonstrating that the fundamental problems of autonomy could be tackled.

How are autonomous vehicles expected to be deployed in society?

Deployment strategies include last-mile delivery of goods (zero-occupancy vehicles), highway trucking with platooning, personalized public transport like autonomous taxis, and testing in closed communities. Semi-autonomous systems with teleoperation also play a role.

Key Moments

Self-Driving Cars: State of the Art (2019)

Lex Fridman

Science & Technology5 min read55 min video

Feb 1, 2019|290,615 views|4,593|352

self-driving cars deep learning lex fridman mit lex autonomous cars deep learning mit self-driving cars 2019 self-driving cars tesla mit tesla waymo mit waymo mit lex fridman

Save to Pod

Key Moments

On this page

TL;DR

Autonomous vehicle development in 2019: industry progress, challenges in human-AI interaction, sensor tech, and future outlook.

Key Insights

The primary goal of autonomous vehicles is to save lives and improve mobility, not just profit.

Waymo has achieved 10 million autonomous miles, while Tesla's Autopilot has covered 1 billion semi-autonomous miles.

Fatalities in autonomous vehicle incidents garner disproportionate public and media attention, creating a higher bar for success.

The debate continues on whether progress will be driven by vision-based systems (like Tesla) or lidar-based systems.

The human experience and interaction are critical for widespread adoption, not just safety or speed.

Truly autonomous vehicles (Level 4/5) require perfect problem-solving, unlike semi-autonomous systems (Level 2/3) where humans act as a failsafe.

THE OVERARCHING MISSION OF AUTONOMOUS VEHICLES

The core mission behind the development of autonomous vehicles extends beyond entrepreneurial aims; it's fundamentally about enhancing mobility for those with limitations, increasing transportation efficiency, and most importantly, saving lives. With a staggering statistic of one global car crash fatality every 23 seconds, the urgency to prevent injuries and fatalities is a primary driver for engineers and researchers in the field.

2018: KEY MILESTONES AND SOERING INCIDENTS

In 2018, Waymo made significant strides, accumulating 10 million autonomous miles, a notable achievement in public deployment. Correspondingly, Tesla's Autopilot system reached 1 billion miles driven semi-autonomously, primarily relying on camera-based computer vision and neural networks. However, the year also saw tragic fatalities, including an Uber incident in Tempe, Arizona, and a Tesla Autopilot-related crash in Mountain View, California, highlighting the critical need for public understanding and stringent safety standards.

THE CHALLENGE OF PUBLIC PERCEPTION AND SAFETY METRICS

Fatalities involving autonomous vehicles, whether fully or semi-autonomous, attract intense public and media scrutiny, setting an exceptionally high bar for performance and safety. While statistical comparisons can be complex, initiatives are underway to accurately compare safety metrics between autonomous and human-driven vehicles. The perception of safety is crucial, as public acceptance is vital for the successful proliferation of this technology.

CURRENT DEPLOYMENTS AND THE QUESTION OF SCALE

Public testing of autonomous taxi services expanded in 2018, with companies like Waymo, Cruise, and others operating in constrained environments, often with safety drivers. The true measure of success, however, lies in achieving scale—meaningful deployment beyond prototype stages. Reaching a threshold like 10,000 operational vehicles is considered essential for these systems to fundamentally impact the general population.

PREDICTIONS AND THE SPECTRUM OF OPTIMISM VS. PESSIMISM

Industry predictions for widespread autonomous vehicle adoption vary significantly, ranging from optimistic forecasts by companies like Tesla to more cautious estimates from leading engineers. While some foresee widespread adoption within years, others believe humans will remain in the loop, albeit with advanced assistance, for decades. This spectrum highlights the inherent difficulty in removing the human element entirely from the driving equation.

HUMAN EXPERIENCE AS THE KEY ADOPTION DRIVER

Beyond technical safety and efficiency, the widespread adoption of autonomous vehicles hinges on creating a superior human experience. This includes intuitive natural language interaction, personalized features, and a seamless transfer of control, making the journey enjoyable and enriching. Focusing on the human experience from the outset, rather than as an afterthought, is considered paramount by many in the field.

DEFINING AUTONOMY: LEVELS AND APPROACHES

Autonomous driving is categorized into levels, with Level 2 representing driver assistance and Level 5 signifying full autonomy where the vehicle is entirely responsible. The distinction between 'human-centered autonomy' and 'full autonomy' is critical, as full autonomy implies no teleoperation or reliance on the driver to take over in critical situations. The goal is for the vehicle to ensure safety independently, even finding a 'safe harbor' if necessary.

STRATEGIES FOR DEPLOYING FULLY AUTONOMOUS VEHICLES

Several deployment strategies for fully autonomous vehicles are being explored. These include last-mile delivery services for goods and people using zero-occupancy vehicles, autonomous trucking on highways with platooning, personalized public transport that mimics train-like services but with fewer passengers, and operations within closed communities. Each approach addresses specific use cases with varying levels of complexity and environmental constraints.

THE VISION VS. LIDAR DEBATE IN SENSOR TECHNOLOGY

A central debate in sensor technology revolves around a vision-centric approach, championed by companies like Tesla using cameras and deep learning, versus a lidar-based approach focused on highly mapped environments. Vision systems offer rich data but struggle with accuracy and adverse weather. Lidar provides precise depth information but is expensive and less reliant on deep learning, leading to a divergence in development philosophies.

THE ROLE AND LIMITATIONS OF VARIOUS SENSORS

Autonomous vehicles rely on a suite of sensors, each with strengths and weaknesses. Cameras offer high resolution but are susceptible to weather and depth estimation issues. Lidar provides accurate 3D mapping but is costly. Radar excels in obstacle detection and all-weather performance but lacks resolution. Ultrasonic sensors are effective for short-range proximity detection. Sensor fusion, combining data from multiple sensors, is crucial for creating a comprehensive environmental model.

SEMI-AUTONOMOUS VS. FULLY AUTONOMOUS SYSTEMS: FAILSAVE MECHANISMS

In semi-autonomous systems, the human driver often serves as the failsafe, stepping in when the system encounters difficulties. Conversely, fully autonomous systems, like those Waymo is developing, rely on failsafes such as highly accurate maps and robust sensor suites (potentially including lidar) to ensure safety without human intervention. This fundamental difference dictates the design and reliability requirements for each approach.

EMERGING CONCEPTS: CONNECTED VEHICLES AND TUNNELS

Beyond traditional sensor-based navigation, innovative concepts like connected vehicles are emerging. Vehicle-to-vehicle and vehicle-to-infrastructure communication could optimize intersections and traffic flow, potentially eliminating traffic lights. Another concept involves extensive underground tunnel networks, which by constraining the operational environment, could dramatically simplify the challenges of autonomous driving and enable higher speeds.

THE FUTURE OF AUTONOMY AND AI'S ROLE

The development of autonomous vehicles represents a defining challenge for artificial intelligence in the 21st century, pushing the boundaries of AI beyond theoretical benchmarks. It requires integrating AI algorithms into real-world applications where they directly impact human lives and civilization. This interdisciplinary challenge involves addressing aspects of psychology, philosophy, sociology, robotics, and perception, making it a continuously fascinating and evolving field.

Mentioned in This Episode

●Products

●Software & Apps

●Companies

●Organizations

●People Referenced

Autonomous Vehicle Deployment and Technology Comparison

Practical takeaways from this episode

Do This

Focus on creating a superior human experience to drive adoption.

Utilize sensor fusion (camera, radar, lidar, ultrasonic) for comprehensive perception.

Address the psychological and sociological aspects of human-robot interaction.

Consider the 'human in the loop' for semi-autonomous systems.

Develop explainable and reliable systems, especially for safety-critical functions.

Avoid This

Assume safety alone will drive adoption; focus on the overall experience.

Underestimate the complexity of non-verbal communication and subtle driving cues.

Rely solely on vision systems without considering limitations in adverse weather or depth estimation.

Treat human factors and the user experience as an afterthought.

Implement purely autonomous systems without a robust failsafe mechanism.

Autonomous Vehicle Sensor Comparison

Data extracted from this episode

Sensor Type	Pros	Cons
Vision (Cameras)	High resolution, rich texture, cheap, abundant data, world designed for human eyes	Noisy, poor depth estimation, bad in extreme weather, requires mass data for accuracy
Lidar	Consistent, reliable, explainable, high accuracy, 360-degree visibility (some sensors)	Expensive, not primarily deep learning-based, limited data due to low adoption
Radar	Cheap, good for obstacle detection and avoidance, works in extreme weather, good range	Lower resolution, limited information
Ultrasonic Sensors	Good for proximity detection, high resolution for close objects (used for parking)	Limited range, similar limitations to radar

Autonomous Vehicle Prediction Timeline (2019 Perspective)

Data extracted from this episode

Company/Group	Prediction	Type of Deployment
Tesla	2019	Fully Autonomous Vehicles
Major Automakers	Varying (2019-2021)	Deployment
Nissan, Honda, Toyota	2020	Highway/Urban Constraints
Hyundai, Volvo	2021	Scale
BMW, Ford	2021	Scale
Daimler	Early 2020s	Scale
Elon Musk	2019	Fully Autonomous Vehicles
Rodney Brooks	Beyond 2050	Fully Autonomous Vehicles (potential city bans in 2030s, US cities in 2040s)

Common Questions

The primary goal is to improve mobility access for those unable to drive due to age or location, increase transportation efficiency, and most importantly, save lives by preventing crashes, injuries, and fatalities.

Topics

Technology & Innovation AI & Machine Learning Mindset & Self-Improvement Autonomous Driving Self-driving Cars Computer Vision Sensor Fusion Machine Learning Human-computer Interaction Future Of Transportation

Mentioned in this video

Companies

Tesla

Reached 1 billion miles in semi-autonomous driving with its Autopilot system. The company's approach uses cameras as the primary sensor and is pushing towards full autonomy.

Waymo

A leader in autonomous vehicle deployment, having reached 10 million autonomous miles by October 2018 and operating the 'Waymo One' service in Phoenix, Arizona.

Drive.ai

Conducting autonomous vehicle testing in Texas.

Excite

A former leader in search engines, disrupted by Google.

Google

Used as an example of a company that disrupted an existing market (search engines with Yahoo, Infoseek, Excite) and is now taking a GPU-to-GPU approach similar to Tesla's hardware development.

May Mobility

Expanding autonomous operations beyond Detroit.

Infoseek

A former leader in search engines, disrupted by Google.

Yahoo

A former leader in search engines, disrupted by Google.

Uber

Experienced a fatality in March 2018 with a pedestrian collision in Tempe, Arizona. Also resumed autonomous vehicle taxi service testing in Pittsburgh after acquiring Otto.

Cruise

Conducting testing in San Francisco, Arizona, and Michigan, with representation from GM.

Optimist Ride

Operating autonomous taxi services in Boston.

Cadillac

Mentioned in the context of 'Super Cruise,' an L2 advanced driver-assistance system.

Locations

Tempe, Arizona

Location of a fatal pedestrian accident involving an Uber autonomous vehicle in March 2018.

Mountain View, California

Location of a fatal accident in 2018 where a Tesla slammed into a divider, killing the driver.

Software & Apps

Moral Machine

A platform used to explore ethical dilemmas in autonomous vehicle programming, specifically regarding choices made in unavoidable accident scenarios.

TensorFlow

A machine learning library mentioned in the context of achieving high accuracy in image classification benchmarks like ImageNet.

Voyage

Working on autonomous taxi services in isolated communities and villages in Florida.

Autopilot

Tesla's semi-autonomous driving system that controls lane position and longitudinal movement, primarily using camera-based computer vision and neural networks.

Organizations

Aurora

Founded by the former head of Tesla Autopilot, conducting testing in San Francisco and Pittsburgh.

MIT Human Centered Autonomous Vehicle

A research group at MIT focused on human-centered autonomous vehicles, possessing a fully autonomous vehicle for testing.

CMU

Their vehicle 'Boss' won the 2007 DARPA Urban Grand Challenge, further demonstrating progress in autonomous driving.

New York Times

Mentioned as an example of media outlets that respond to single fatalities involving autonomous vehicles, sometimes disproportionately.

DARPA

Organized the 'Race to the Desert' in 2005 and the 'Urban Grand Challenge' in 2007, which significantly advanced the field of autonomous vehicles and robotics.

Stanford

Their vehicle 'Stanley' won the 2005 DARPA Grand Challenge, captivating imaginations about autonomous vehicle capabilities.

People

Gill Pratt

A roboticist from Toyota who emphasizes the importance of the 'human in the loop' for autonomous vehicle development, stating that removing the human is still far off.

Ernest Rutherford

A physicist quoted for his view that physics is the only real science, implying that other fields, like computer science, might be seen as dealing with 'stupid silly details'.

Chris Urmson

A key figure behind Aurora, previously involved with DARPA challenges.

Elon Musk

Represents the optimistic side of the autonomous vehicle prediction spectrum, with bold predictions for their arrival, including fully autonomous vehicles by 2019.

Rodney Brooks

Represents the more pessimistic side of the autonomous vehicle prediction spectrum, suggesting fully autonomous vehicles may not be prevalent until beyond 2050, with significant city bans on manual driving in the 2030s.

Products

Uber Elevate

Uber's initiative exploring aerial mobility and flying cars, suggesting future transportation might involve taking elevators to rooftops to meet vehicles.

Ford F-150

Mentioned as the most popular car in America, indicating the continued prevalence and data generation from manually driven vehicles.

Volvo S90

An example of a vehicle with an advanced driver-assistance system (L2) that can keep the car in its lane.