Sebastian Thrun (Stanford Professor) – Towards Self-Driving Cars, ICAPS (Sep 2009)
Chapters
00:00:17 DARPA Grand Challenge: From Race to Robot Competition
DARPA Grand Challenge: DARPA organized a competition to build self-driving cars capable of navigating the Mojave Desert without human intervention. The prize money attracted 106 and later 195 teams. The first race in 2004 ended disastrously, with no team completing the course.
Stanford’s Entry: Stanford professor Sebastian Thrun entered the competition with a class of students, CS 294, Artificial Intelligence. The students received course credit instead of pay for their work. Volkswagen donated a car, Intel donated a computer system, and a drive-by-wire system was implemented. Various sensors, including GPS, inertial measurements, lasers, and cameras, were installed on the vehicle for proprioception and environment perception. Obstacle detection was simplified using a SICK laser range finder that scans the road ahead and flags vertically aligned points as obstacles.
00:05:58 Autonomous Driving Techniques in the Early 2000s
Planning: Instead of formal planning, the vehicle reacted to obstacles by generating a set of possible steering nudges, checking for collisions and proximity to the desired path, and selecting the best option.
Controls: A simple controller adjusted the front wheels to follow a reference trajectory or steered in proportion to cross-track error to correct deviations.
Desert Driving and Perception: Experiments took place in the Mojave Desert, with varying terrain from smooth to rutted and sandy. Laser sensors detected obstacles, generating data visualized as colored displays, with white indicating drivable areas, red indicating non-drivable areas, and gray indicating unknown areas. GPS data was used to define a corridor, but it was not very accurate.
Long-Range Perception: To see further than the limited range of laser sensors, a camera was mounted on the vehicle, and a computer vision algorithm was developed to find the road ahead.
00:09:56 DARPA Grand Challenge: Autonomous Robots Race Through the Desert
Stanley’s Challenges: Finding drivable terrain: Color and smoothness not reliable indicators Adaptive vision using laser sensor and machine learning Projecting road surface back into map with flat ground assumption Controlling a motorcycle: Turning a motorcycle requires a controlled fall Anthony Lewandowski’s numerous falls and creative solutions
The DARPA Grand Challenge: Pre-qualification event near Los Angeles: Narrow gate, abandoned car, high-speed area, emulated tunnel Many competitors struggled, highlighting the challenge’s difficulty The race: Carnegie Mellon had an advantage with multiple robots Stanley started second, followed by its chase vehicle Intense emotions as the robot embarks on its own journey
Stanley’s Victory: Carnegie Mellon’s fuel tank switching issue slowed them down Stanley took the lead Passing Carnegie Mellon’s paused vehicle Video footage used to navigate the only man-made obstacle Descending Beer Bottle Pass: Challenging terrain with limited visibility Stanley’s sensors covered in dust, relying on ruggedness controller
00:18:43 DARPA Grand Challenge and Urban Driving Planning
DARPA Grand Challenge Victory: Stanley, a robot vehicle, successfully completed the DARPA Grand Challenge, becoming the first robot to finish a challenge of such magnitude. The team celebrated their victory, while Carnegie Mellon, who came in second place, did not. The win was attributed to a lucky incident rather than superior planning.
Planning for Urban Driving: DARPA issued a new challenge focused on urban driving, which required more complex planning than desert driving. The goal was to develop a system that could plan routes, handle unexpected obstacles, and navigate through traffic autonomously.
Sensor Technology: The vehicle, Junior, was equipped with a scanning laser sensor, Juno, which provided 64 scan lines at 10 hertz. The sensor enabled obstacle detection, including curbs, using machine learning technology. The sensor also facilitated localization with an error of about five centimeters, enabling precision driving in lanes. The system could track other cars using particle filters, displaying them on a screen along with camera images.
Global Mission Planning: Dynamic programming is used to determine the optimal route based on expected time arrival. The algorithm considers factors such as traffic conditions, lane turning constraints, and multiple lanes. This method provides a globally optimal route that is slightly more comprehensive than what a personal navigation device offers.
Tactical Normal Driving Planning: Combines dynamic programming with rollouts to analyze route choices. Rollouts are used to compute costs and connect to a dynamic programming table for a global perspective. This approach integrates momentary traffic situations with a global perspective to find the optimal route.
Freestyle Planning Mode: Used when eligible trajectories do not provide a solution, such as in situations where the vehicle might get stuck. This mode involves shedding all rules and driving the shortest path. It is useful for handling road blockages, traffic cones, and parking maneuvers that require backups.
00:27:39 Hybrid A Star Path Planning for Autonomous Driving
Planning and Motion Planning: Planning is crucial for autonomous vehicles to navigate in complex environments. Motion planning involves finding executable paths that adhere to vehicle kinematics and obstacle constraints. Critical R must be executable, continuous, and fast (in the order of milliseconds).
Challenges in Path Planning: Discretization of the world using A* algorithm results in discrete plans that are not executable by real-world vehicles. Classical planning paradigms like A* and D* have limitations in handling continuous motion.
Hybrid A* Algorithm: Continuous A* algorithm was developed to address the limitations of discrete planning. It memorizes continuous coordinates along the path and executes real robot actions to generate a drivable trajectory. Hybrid A* provides an executable and feasible path, but it may not be optimal and can miss some plans.
Heuristics for Faster Planning: Offline heuristics factor in vehicle-specific geometry to favor paths that require minimal maneuvers. Online heuristics use a 2D search to provide a lower bound on the distance to the goal, reducing the search space.
Precision at the Target Location: Achieving high precision at the target location can be challenging with discretized planning. Reach shape curves offer a fast closed-form solution to find admissible curves to goal locations, ignoring obstacles. By incorporating reach shape curves, the planner can quickly find paths with perfect precision.
Path Smoothing: The generated paths may not be smooth and can result in jerky motions. A smoothing formula is used to create smoother paths that provide a more natural driving experience.
00:36:52 Path Planning with Voronoi Fields and Potential Fields
Path Smoothing: Path smoothing is a post-processing step to optimize the smoothness of a path, resulting in a locally optimized trajectory. It improves the trajectory by straightening out topological paths without finding topologically different paths. The process is efficient, typically taking only five milliseconds.
Voronoi Field for Obstacle Avoidance: Potential fields are commonly used for obstacle avoidance in autonomous navigation. A modification was made to rescale the obstacle repellent force based on the distance to the nearest clearing space. A Voronoi decomposition of the data is used to measure the clearance for any given point. The clearance is determined by finding the nearest Voronoi line and measuring the distance to the nearest obstacle. The obstacle repellent force in the potential field is rescaled according to the clearance, making the cost of going through a narrow opening the same as the cost of going through a wide opening.
Combining Voronoi Field with Path Smoothing: The Voronoi field is incorporated into the path smoothing process to effectively push the vehicle away from obstacles. This approach is particularly useful when dealing with openings of varying widths, ensuring that the vehicle can navigate through narrow openings as efficiently as wide openings.
Benefits of Voronoi Field: The Voronoi field modification is a minor improvement to the extensive literature on potential field navigation. It enables efficient and effective obstacle avoidance in autonomous navigation, particularly in scenarios with openings of varying widths.
00:39:07 Planning and Execution for Autonomous Vehicles
Autonomous Vehicle Planning: The planning algorithm used by Stanford’s autonomous car in the DARPA Grand Challenge was extremely fast, taking no more than 0.2 seconds for replanning. The algorithm utilized heuristics and smoothing techniques, with backups easily incorporated. Smoothing itself took approximately 10 milliseconds, while A-star replanning in the 3D space took no more than 200 milliseconds. The planning algorithm only activated when the most promising trajectory could not be smoothed further.
Real-World Examples: In a staged example, the car successfully backed up into a narrow parking spot, demonstrating the precision required for such maneuvers. The car navigated a parking lot filled with traffic cones, displaying its ability to handle complex environments.
The DARPA Grand Challenge: Stanford’s car came in second place, with Carnegie Mellon winning the race. The competition showcased impressive achievements, with six cars successfully completing the course. Several accidents occurred, including one involving MIT and Cornell’s vehicles, and a larger traffic jam involving four robots and five humans. The event highlighted the challenges of autonomous driving in complex real-world scenarios.
Unsolved Problems: The presentation acknowledged problems that autonomous vehicles still faced, such as navigating traffic without clear rules, like in Delhi, India, where the mass of the vehicle determines priority. The presentation also discussed the challenge of fully utilizing a vehicle’s potential, which was later solved through reinforcement learning techniques.
00:46:01 The Future of Transportation: Reimagining Roads and Vehicles
Inefficiencies of Current Transportation Systems: Sebastian Thrun draws a comparison between the inefficiency of current transportation systems and the inefficiencies of broadcast and print media before the internet. He highlights the lack of innovation in the automotive industry, with car usage remaining largely unchanged in the past 50 years.
Car Accidents and Human Error: Car accidents are the leading cause of death for young people, and 98% of them are due to human error. Distractions while driving, such as texting, are a major contributing factor to accidents.
Lost Time and Commuting: Commuting takes up a significant portion of time for working individuals, with an average of over an hour per day in the U.S. This time could be utilized more productively if cars were able to behave like trains on a highway, allowing for activities during the commute.
Accessibility for People with Disabilities: Thrun emphasizes the importance of autonomous cars in providing independence and a higher quality of life for individuals who cannot drive due to disabilities or age-related factors.
Energy Efficiency: Road traffic consumes a substantial amount of energy, with a significant portion being wasted. Driving closely together, like a train, can increase energy efficiency by reducing wind drag. This could lead to increased highway capacity and a reduction in traffic congestion.
Underutilized Cars and Parking Spaces: Cars are often parked for extended periods, resulting in wasted natural resources and real estate. Self-driving cars could increase car usage and reduce the number of vehicles needed. This would free up parking spaces and potentially eliminate the need for individual car ownership.
00:53:36 Transportation Innovations for a Safer, More Convenient, and Efficient Future
Benefits of Autonomous Vehicles: Increased safety: Reduced accidents and fatalities. Cost savings: Cars and car-based transportation are the second largest expense for American households. Convenience: Passengers can focus on other activities while the car drives. Efficiency: Vehicles can travel faster and more efficiently. Sharing: Cars can be shared among multiple users, reducing the number of vehicles on the road.
Challenges in Developing Autonomous Vehicles: Lack of funding: The US automotive industry is struggling and there is little funding available for autonomous vehicle research. Technical challenges: Developing the necessary computer science and engineering solutions to enable autonomous vehicles is a complex task.
Call to Action: Encourages collaboration and innovation to advance the development of autonomous vehicles. Highlights the potential impact of autonomous vehicles on society, including increased safety, cost savings, convenience, efficiency, and environmental sustainability.
Car-to-Car and Car-to-Environment Communication: Acknowledges ongoing work in this area but chooses not to emphasize it due to the chicken-and-egg problem. Believes that focusing on developing the core technology of autonomous vehicles is more important at this stage.
00:56:14 Autonomous Cars: Issues, Challenges, and Potential Benefits
Co-evolution of Environment and Cars: Thrun believes that autonomous vehicles (AVs) should adapt to existing infrastructure rather than waiting for major changes, just as early cars did on horse trails.
Computational Complexity of Global Traffic Optimization: Thrun acknowledges the computational challenges of optimizing traffic on a global scale, requiring more advanced scheduling and planning algorithms.
Safety and Reliability Concerns: He emphasizes the need for proper engineering methodology to guarantee safety and reliability, including extensive error analysis and component testing.
Cost-effectiveness and Marketability: Thrun mentions the high cost of current AV systems and the importance of cost reduction to make them marketable.
Liability and Insurance: Liability currently lies with the operators of AVs, but he believes that insurance companies will eventually benefit from reduced costs due to fewer accidents.
Energy Efficiency: Thrun suggests that AVs could become more energy-efficient by forming closely linked trains on highways, exploiting wind drag, and using special HOV lanes.
Government Interest: He notes that the Obama administration has expressed interest in pursuing AV technology if it becomes available.
Challenges and Time Constraints: Thrun acknowledges that there are many unanswered questions and challenges related to AVs, and his answers may evolve over time.
Conclusion: The presentation ends with a thank you to the panel, indicating that the discussion on autonomous vehicles has come to an end.
Abstract
The Evolution and Impact of Autonomous Vehicles: A Journey from DARPA Challenges to Modern Day
Abstract:
This article delves into the revolutionary development of autonomous vehicles, tracing their evolution from the DARPA Grand Challenge to modern-day advancements. We investigate the technical milestones, challenges, and societal impacts of self-driving cars, presenting a comprehensive understanding of this transformative technology.
—
Introduction: The DARPA Grand Challenge and its Legacy
The DARPA Grand Challenge, launched in 2004, marked a pivotal moment in automotive history. It aimed to foster the development of self-driving cars, and the million-dollar prize competition attracted over 100 teams. Stanford University, led by Sebastian Thrun and Mike Montemello, emerged as formidable contenders. Despite initial setbacks, including the disastrous first race where no team could complete more than 5% of the course, the challenge set the stage for groundbreaking innovations in vehicle sensors, obstacle detection, and planning algorithms.
Overview of Autonomous Vehicle Competitions and Stanford’s Entry:
The first DARPA Grand Challenge in 2004 saw no team able to complete the course. Professor Sebastian Thrun of Stanford University, inspired by the competition, entered a class of students who worked on the project for course credit instead of pay. Stanford secured a car donation from Volkswagen, a computer system from Intel, and the students built a drive-by-wire system to enable proprioception and environmental perception. Sensors, including GPS, inertial measurements, lasers, and cameras, were strategically installed on the vehicle to facilitate these functions.
Stanford’s Journey: From Concept to Victory
Stanford’s entry into the competition, with limited resources, showcased ingenuity and collaboration. The team, armed with the donated Volkswagen car, Intel’s computer system, and the dedicated efforts of students enrolled in CS 294, developed “Stanley,” a robotic vehicle equipped with advanced sensors for proprioception and environmental perception. This innovative approach included GPS, inertial measurements, and a fusion of lasers and cameras for precise navigation.
Technical Breakthroughs: Obstacle Detection and Planning
The team’s obstacle detection strategy, utilizing a SICK laser range finder, proved pivotal in navigating challenging terrains like the Mojave Desert. Their planning approach, which included reactive steering adjustments and a straightforward controller for aligning with the trajectory, played a critical role in Stanley’s victory in the 2005 DARPA Grand Challenge.
Early Self-Driving Cars: Reactive Planning, Simple Controls, and Basic Perception:
Instead of formal planning, the vehicle responded to obstacles by generating potential steering adjustments, checking for collisions and proximity to the desired path, and selecting the most suitable option. A rudimentary controller adjusted the front wheels to follow a reference trajectory or steered in proportion to cross-track error to rectify deviations. Experiments took place in the Mojave Desert, with varying terrain ranging from smooth to rutted and sandy. Laser sensors detected obstacles, generating data visualized as colored displays, where white represented drivable areas, red indicated non-drivable areas, and gray signified unknown areas. GPS data was used to define a corridor, but its accuracy was limited. To enhance visibility beyond the limited range of laser sensors, a camera was mounted on the vehicle, and a computer vision algorithm was developed to identify the road ahead.
Stanley: The Autonomous Car that Won the DARPA Grand Challenge:
Stanley encountered several challenges, including finding drivable terrain, controlling a motorcycle, and navigating the race course. The pre-qualification event near Los Angeles highlighted the challenge’s complexity, with many competitors struggling. During the race, Stanley started second, followed by its chase vehicle, and overtook Carnegie Mellon’s paused vehicle by utilizing video footage to navigate the only man-made obstacle. Stanley’s victory was significant as it showcased the immense potential of autonomous vehicles.
DARPA Grand Challenge Victory:
Stanley, a robot vehicle, successfully completed the DARPA Grand Challenge, becoming the first robot to accomplish a challenge of such magnitude. The team celebrated their triumph, while Carnegie Mellon, who came in second place, did not. The victory was initially attributed to a fortunate incident rather than superior planning.
From Desert to Urban: Escalating Challenges
Following their desert triumph, DARPA escalated the challenge to urban driving, necessitating more intricate route planning and navigation. Junior, Stanford’s subsequent entry, boasted advanced sensors and algorithms for dynamic programming, tactical driving planning, and freestyle maneuvers. This transition underscored the escalating complexity and potential of autonomous driving technology.
Planning for Urban Driving:
DARPA issued a new challenge focused on urban driving, which posed greater planning complexities than desert driving. The goal was to create a system capable of autonomously planning routes, handling unexpected obstacles, and navigating through traffic.
Sensor Technology:
The vehicle, Junior, was equipped with a scanning laser sensor, Juno, which provided 64 scan lines at 10 hertz. The sensor facilitated obstacle detection, including curbs, using machine learning technology. Moreover, it enabled localization with an accuracy of approximately five centimeters, allowing for precise driving in lanes. The system could track other cars using particle filters, displaying them on a screen alongside camera images.
Path Smoothing and Voronoi Field for Autonomous Navigation:
Path smoothing is a post-processing step that enhances the smoothness of a trajectory, resulting in a locally optimized path. It rectifies topological paths without finding different topological paths. This process is efficient, typically taking only five milliseconds.
A modification was made to rescale the obstacle repellent force based on the proximity to the nearest clearing space. A Voronoi decomposition of the data is used to measure the clearance for any given point. The clearance is determined by finding the nearest Voronoi line and measuring the distance to the nearest obstacle. The obstacle repellent force in the potential field is rescaled according to the clearance, making the cost of navigating through a narrow opening equivalent to the cost of going through a wide opening.
The Voronoi field is incorporated into the path smoothing process to effectively repel the vehicle from obstacles. This approach is particularly useful when dealing with openings of varying widths, ensuring that the vehicle can traverse narrow openings as efficiently as wide openings. The Voronoi field modification offers a minor improvement to the extensive literature on potential field navigation. It enables efficient and effective obstacle avoidance in autonomous navigation, particularly in scenarios with openings of varying widths.
Sebastian Thrun’s Presentation on Autonomous Vehicle Planning:
The planning algorithm employed by Stanford’s autonomous car in the DARPA Grand Challenge was remarkably fast, taking no more than 0.2 seconds for replanning. The algorithm leveraged heuristics and smoothing techniques, effortlessly incorporating backups. Smoothing itself took approximately 10 milliseconds, while A-star replanning in the 3D space required no more than 200 milliseconds. The planning algorithm only activated when the most promising trajectory could not be further smoothed.
In a staged demonstration, the car successfully reversed into a tight parking spot, showcasing the precision required for such maneuvers. It adeptly navigated a parking lot filled with traffic cones, demonstrating its ability to handle intricate environments.
Sebastian Thrun’s Perspective on the Future of Transportation:
Sebastian Thrun draws parallels between the current inefficiencies of transportation systems and the inefficiencies of broadcast and print media before the internet. He underscores the lack of innovation in the automotive industry, with car usage remaining largely unchanged over the past 50 years.
Thrun emphasizes the paramount importance of autonomous cars in providing independence and enhancing the quality of life for individuals who cannot drive due to disabilities or age-related factors. He also highlights the potential for energy efficiency, with road traffic consuming a substantial amount of energy, a significant portion of which is wasted. Driving closely together, akin to a train, can enhance energy efficiency by reducing wind drag, potentially increasing highway capacity and alleviating traffic congestion.
Cars are often parked for extended periods, resulting in wasted natural resources and real estate. Self-driving cars could potentially increase car usage and reduce the number of vehicles needed, freeing up parking spaces and possibly eliminating the need for individual car ownership.
A Call for Collaboration and Innovation
The journey of autonomous vehicles from the DARPA challenges to their impending commercial success underscores the need for sustained innovation, collaboration, and adaptation. While significant progress has been made, the road ahead demands a concerted effort from all stakeholders to fully realize the transformative potential of self-driving cars.
—
Additional Insights from Sebastian Thrun’s Presentation:
Benefits of Autonomous Vehicles:
– Increased safety: Reduced accidents and fatalities.
– Cost savings: Cars and car-based transportation are the second largest expense for American households.
– Convenience: Passengers can focus on other activities while the car drives.
– Efficiency: Vehicles can travel faster and more efficiently.
– Sharing: Cars can be shared among multiple users, reducing the number of vehicles on the road.
Challenges in Developing Autonomous Vehicles:
– Lack of funding: The US automotive industry is struggling and there is little funding available for autonomous vehicle research.
– Technical challenges: Developing the necessary computer science and engineering solutions to enable autonomous vehicles is a complex task.
Call to Action:
– Encourages collaboration and innovation to advance the development of autonomous vehicles.
– Highlights the potential impact of autonomous vehicles on society, including increased safety, cost savings, convenience, efficiency, and environmental sustainability.
Car-to-Car and Car-to-Environment Communication:
– Acknowledges ongoing work in this area but chooses not to emphasize it due to the chicken-and-egg problem.
– Believes that focusing on developing the core technology of autonomous vehicles is more important at this stage.
Co-evolution of Environment and Cars:
– Thrun believes that autonomous vehicles (AVs) should adapt to existing infrastructure rather than waiting for major changes, just as early cars did on horse trails.
Computational Complexity of Global Traffic Optimization:
– Thrun acknowledges the computational challenges of optimizing traffic on a global scale, requiring more advanced scheduling and planning algorithms.
Safety and Reliability Concerns:
– He emphasizes the need for proper engineering methodology to guarantee safety and reliability, including extensive error analysis and component testing.
Cost-effectiveness and Marketability:
– Thrun mentions the high cost of current AV systems and the importance of cost reduction to make them marketable.
Liability and Insurance:
– Liability currently lies with the operators of AVs, but he believes that insurance companies will eventually benefit from reduced costs due to fewer accidents.
Energy Efficiency:
– Thrun suggests that AVs could become more energy-efficient by forming closely linked trains on highways, exploiting wind drag, and using special HOV lanes.
Government Interest:
– He notes that the Obama administration has expressed interest in pursuing AV technology if it becomes available.
Challenges and Time Constraints:
– Thrun acknowledges that there are many unanswered questions and challenges related to AVs, and his answers may evolve over time.
DARPA's Grand Challenge sparked a technological revolution in autonomous vehicles, demonstrating their potential for safer, more efficient, and environmentally friendly transportation. However, challenges remain in addressing technological limitations, standardization issues, and public acceptance to realize this transformative vision fully....
Sebastian Thrun's work spans AI, self-driving cars, wearable tech, and revolutionizing online education with Udacity's nanodegree programs, impacting technology and education. Thrun's vision for the future involves outsourcing personal experiences through technology and leveraging AI to empower humans rather than replace them....
Sebastian Thrun's focus on AI and machine learning has revolutionized autonomous vehicles and urban air mobility, while his emphasis on education and soft skills aims to empower individuals for the future workplace....
Online education has the potential to democratize knowledge and provide personalized learning experiences, but challenges remain in ensuring rigor, quality, and support for diverse learners. Sebastian Thrun's work highlights the importance of engaging and interactive learning, continuous improvement, and adapting education to meet evolving societal needs....
Sebastian Thrun's passion for solving societal problems led him to focus on self-driving cars and education, using logical execution and rapid iteration to drive innovation and achieve societal impact. Thrun's methodical approach and unwavering dedication to improving society through technology serve as an inspiration for aspiring innovators....
Sebastian Thrun's views on AI and technology revolve around its potential to revolutionize various aspects of human life and augment human capabilities, leading to a "superhuman" future. He emphasizes the need for responsible usage and open dialogue to ensure technology benefits humanity and addresses global challenges....
Self-driving cars are anticipated to become mainstream in the next decade, presenting challenges in developing business models, legal frameworks, and building consumer trust. Udacity's Self-Driving Car Nanodegree program equips students with in-demand skills for a growing industry, addressing the global demand for talent in this field....