What are the limitations of the RRT algorithm?

The original RRT algorithm can struggle to converge to optimal solutions, as early trajectories can constrain the search space. This can lead to inefficient or looping paths, as seen with the autonomous forklift example.

How does RRT* improve upon RRT?

RRT* enhances RRT by performing local corrections on the newly found trajectories. This ensures asymptotic optimality, meaning the algorithm will eventually converge to the best possible path, with minimal increase in computational cost.

What was the objective of the DARPA Urban Challenge?

The DARPA Urban Challenge was a competition where teams used street-legal vehicles instrumented with sensors and computers to navigate a 60-mile course in under six hours within an urban traffic environment, facing other vehicles.

What were the key challenges in developing the DARPA Urban Challenge vehicle?

Key challenges included building complicated hardware, developing essential infrastructure like communication middleware (LCM), and extensive testing, which was limited by time constraints.

How did the MIT team's vehicle perform in the DARPA Urban Challenge?

MIT's team was one of six finishers, placing fourth. They experienced a collision with Cornell's vehicle due to an inference error by the collision checker, highlighting the complexities of real-world autonomous driving.

How have affordable cars changed cities and transportation?

Affordable cars led to suburban sprawl, altered urban landscapes, and created issues like congestion and pollution. They also changed how people interact with their environment, prioritizing cars over pedestrians.

What is the future potential of autonomous vehicles in urban environments?

Autonomous vehicles, combined with sharing, electrification, and new business models, could significantly reduce transportation costs and increase mobility, transforming cities and generating new economic activity.

What are the main challenges for widespread adoption of autonomous vehicles?

Beyond technology, challenges include regulations, insurance, societal acceptance, ethical considerations, urban architecture, and the need for new business models to support widespread deployment.

Can a car safely navigate using only cameras (vision-only)?

It's believed that vision-only navigation is achievable within the next 3-5 years, using techniques like deep learning and geometric approaches to build 3D maps from camera data. Cameras are good sensors, but require significant computational power.

What role does communication play in autonomous vehicle systems?

Communication between vehicles (V2V) and with infrastructure (V2I) is crucial for advanced scenarios like coordinated intersection management and lane following. However, cybersecurity is a major concern.

Key Moments

Sertac Karaman (MIT) on Motion Planning in a Complex World - MIT Self-Driving Cars

Lex Fridman

Science & Technology3 min read63 min video

Dec 13, 2017|43,578 views|801|26

mit deep learning self-driving cars driving

Save to Pod

Key Moments

TL;DR

MIT professor Sertac Karaman discusses motion planning, autonomous vehicles, and the future of transportation.

Key Insights

Motion planning is a computationally challenging problem, especially with increasing dimensions, leading to algorithms like RRT and RRT* to find efficient paths.

Autonomous vehicles require robust perception, high-performance computing, and sophisticated planning and control systems, often involving significant hardware integration.

The development of autonomous vehicles has evolved from academic projects like the DARPA Urban Challenge to industry-level solutions, influencing urban infrastructure and societal norms.

Future transportation systems could leverage autonomous vehicles, sharing, and electrification to significantly reduce costs and increase mobility, but face hurdles in regulation and societal integration.

Deep learning is increasingly important, particularly for vision-based perception, but geometric and hybrid approaches remain relevant for autonomous navigation.

Communication between vehicles (V2V) and infrastructure (V2I) can enhance autonomous operations, though cybersecurity remains a critical concern.

BACKGROUND IN AUTONOMOUS VEHICLE RESEARCH

Sertac Karaman, a professor at MIT in the Aero-Astro department, specializes in autonomous vehicles for land and air. His background includes participation in the DARPA Urban Challenge, focusing on motion planning. Early work involved creating autonomous vehicles like a Land Rover for the challenge and an autonomous forklift that responded to voice commands, predating modern personal assistants. This foundational experience in robotics and autonomous systems shaped his research trajectory.

MOTION PLANNING ALGORITHMS: FROM RRT TO RRT*

A core area of Karaman's research is motion planning, the process of finding a path for a robot from a start to a goal while avoiding obstacles. He discusses the Rapidly-exploring Random Tree (RRT) algorithm, a widely used method that explores the state space by incrementally building a tree of possible trajectories. However, RRT has limitations, including a potential to get stuck in suboptimal paths. To address this, Karaman developed RRT*, which guarantees asymptotic optimality by performing local corrections, ensuring convergence to the best possible trajectory over time with increased computational effort.

VEHICLE-LEVEL AND SYSTEM-LEVEL CHALLENGES

Karaman's research group tackles both individual vehicle capabilities and the integration of autonomous vehicles into larger systems. Vehicle-level challenges include developing fast, agile robots capable of complex maneuvers, often requiring high-performance computing and advanced perception systems with high-rate cameras. System-level work involves orchestrating fleets of autonomous vehicles, such as those used in warehouses (like Kiva systems) or future delivery networks, and considering their interactions in complex environments like busy ports or ride-sharing scenarios.

DARPA URBAN CHALLENGE AND ITS LEGACY

The DARPA Urban Challenge was a pivotal event that significantly spurred the field of autonomous vehicles. Karaman details MIT's entry, an instrumented Land Rover LR3 equipped with extensive sensors and computational power, including multiple cameras, radars, and LiDARs, necessitating an onboard generator and air conditioner. Despite a collision during the race, the experience provided invaluable lessons in system integration, software development (including the foundational Lightweight Communications and Marshalling – LCM), and the critical importance of extensive testing and simulation, influencing subsequent industry developments like Google's self-driving car project.

THE FUTURE OF TRANSPORTATION AND AUTONOMOUS SYSTEMS

The affordable car revolutionized cities in the 20th century, leading to urban sprawl, pollution, and congestion. Karaman posits that 21st-century technologies like robotics, AI, and connectivity offer an opportunity to reimagine urban mobility. He envisions a future with more accessible, affordable, and efficient transportation services, potentially driven by autonomous vehicles, ride-sharing, and electrification. This could decrease the cost of travel significantly, leading to increased mobility and economic activity, fundamentally changing how people live and interact with urban environments.

CHALLENGES AND OPPORTUNITIES IN AUTONOMY

While technological advancements are rapidly progressing, the widespread adoption of autonomous vehicles faces significant non-technical challenges. These include navigating complex legal frameworks, insurance policies, ethical considerations, regulatory approvals, and evolving societal acceptance. The field is a blend of high-speed, low-complexity environments like highways (easier to conquer) and slower, high-complexity environments like urban streets (much harder). Integrating these systems requires a multidisciplinary approach, combining engineering with urban planning, policy, and new business models to fully realize the potential benefits.

Mentioned in This Episode

●Products

●Software & Apps

●Companies

●Organizations

●People Referenced

Common Questions

Motion planning is the process of determining a path for an autonomous vehicle to navigate from its current position to a goal, while avoiding obstacles. Algorithms like RRT and RRT* are used to find these paths in complex environments.

Topics

Technology & Innovation AI & Machine Learning Science & Mathematics Autonomous Vehicles Motion Planning Deep Learning Computer Vision Urban Planning Transportation Systems

Mentioned in this video

Software & Apps

RRT*

An improved version of the RRT algorithm that guarantees asymptotic optimality, developed by the speaker.

Lightweight Communications and Marshaling

A communication middleware developed for the DARPA Urban Challenge, now utilized in the industry for autonomous cars.

Rapidly-exploring Random Tree

An algorithm for motion planning that the speaker worked on, which explores state space by sampling.

GPU

Graphics Processing Units used for high-performance computing in autonomous vehicles, including on drones and for deep learning.

Robotic Operating System

A software framework for robotics, mentioned as infrastructure that was absent during the early days of autonomous vehicle development.

Companies

Kiva Systems

A system used by Amazon for packing orders with autonomous vehicles.

Amazon

Company referenced for its use of Kiva Systems for order fulfillment.

Tesla

A company known for its camera-based autonomous vehicles and deep learning approaches.

Nokia

A company that used to make phones and handheld devices, mentioned in the context of the autonomous forklift project.

Velodyne

A company that manufactures 3D laser scanners, a crucial sensor for the DARPA Urban Challenge.

Optimist Inc.

A new company the speaker is involved with, focusing on autonomous vehicles, which recently raised seed funding.

Organizations

Virginia Tech

University whose team came third in the DARPA Urban Challenge.

Draper Laboratory

Provided significant support for system and vehicle integration during the DARPA Urban Challenge.

Olin College

Provided vehicle engineering support, helping package the initial prototype for the DARPA Urban Challenge.

MIT

Massachusetts Institute of Technology, where the speaker is a professor and many of the discussed projects took place.

Cornell

A university whose autonomous vehicle was involved in a collision with the speaker's team's vehicle during the DARPA Urban Challenge.

CMU

Carnegie Mellon University, whose team won first place in the DARPA Urban Challenge.

Stanford

University whose team came second in the DARPA Urban Challenge.

People

Ryan Chin

Co-founder of Optimist Inc., his doctoral thesis involved the MIT City Car project.

Sertac Karaman

Professor at MIT in the aero-astro department, specializing in autonomous vehicles and motion planning.

John Boehner

An individual the speaker shook hands with exactly a decade prior to the talk, marking the start of his graduate student work on the DARPA Urban Challenge team.

Jenny Larios Berlin

Part of the founding team of Optimist Inc., with an MBA and Master's from MIT, and former MD of University of Campus Operations at Zipcar.

Albert Wong

Part of the founding team of Optimist Inc., with prior experience at Rethink Robotics, Google X (Project Wing), and the Urban Challenge.

Ramiro Romania

Designer and founding team member of Optimist Inc., with a background from Harvard Graduate School of Design and experience building Quito's subway system.

Products

Land Rover LR3

The vehicle used by the speaker's team for the DARPA Urban Challenge, which was made autonomous.

DGX-1

A high-performance computing system used by the speaker's group for computing controllers.

Google car

An autonomous vehicle developed by Google, seen as a spin-off from the DARPA Urban Challenge, featuring similar sensing packages.

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