What are the three main tasks involved in self-driving?

The three core tasks are perception (understanding the environment from sensor data), prediction (anticipating the future behavior of other agents), and planning (making safe, comfortable, and progressive driving decisions).

How does Waymo's 'ML factory' process work?

Waymo uses a continuous cycle: deploying software, driving and collecting data, selecting interesting data for labeling, training ML models, testing and validating them, and then redeploying the improved models to collect more data.

What are the key ingredients for Waymo's machine learning infrastructure?

The key ingredients are: 1) Computing software infrastructure (leveraging Google's resources like TensorFlow and specialized hardware), 2) High-quality labeled data (involving smart selection and auto-labeling), and 3) High-quality models (benefiting from Alphabet's AI leadership).

How does Waymo ensure robustness when ML models are uncertain?

Waymo uses redundant and complementary sensors (camera, lidar, radar) and a hybrid system approach. If ML models are not confident, the system drives more conservatively and leverages expert domain knowledge or fixed rules to ensure safety.

Why is simulation so critical for self-driving development?

Simulation allows for testing billions of miles in diverse and rare scenarios that are unsafe or impractical to test in the real world. It helps validate changes to perception systems and ensures robustness across a vast range of situations.

What are the challenges in simulating realistic non-player character behavior?

Simple models like 'break and swerve' are insufficient. Realistic simulation requires learning agent behavior from real-world demonstrations to capture complex interactive scenarios, which is an active area of research.

Does Waymo use a single model for all driving conditions, or multiple models?

Ideally, a single model that adapts to most scenarios is preferred. However, complementary approaches and potentially specialized models are still needed, especially for complex or rare situations.

How important is capturing human-like reasoning in AI for self-driving?

Reasoning, especially understanding human intent and subtle cues like attention, is crucial. While current deep learning excels at pattern recognition, exploring more complex reasoning capabilities in AI is seen as a fruitful long-term direction.

How does Waymo balance machine learning with traditional expert-designed algorithms?

Currently, a hybrid approach is vital. While ML is improving, it's not error-free. Expert-designed algorithms and knowledge complement ML, especially in complex scenarios where collecting sufficient data is difficult, ensuring safety and reliable behavior.

What makes scaling self-driving technology to new cities challenging?

Each new environment presents unique challenges, including new intersections, local driving customs (like the 'Pittsburgh left'), and varied road layouts. The system must adapt to these 'long tail' variations to scale effectively.

Key Moments

Drago Anguelov (Waymo) - MIT Self-Driving Cars

Lex Fridman

Science & Technology3 min read66 min video

Feb 12, 2019|167,560 views|2,460|137

self-driving cars artificial intelligence deep learning machine learning self driving cars waymo google lex fridman autonomous cars computer vision waymo one deep learning mit

Save to Pod

Key Moments

TL;DR

Waymo's Drago Anguelov discusses taming the long tail of autonomous driving through ML, simulation, and hybrid systems.

Key Insights

The "long tail" of rare and unpredictable events is the primary challenge in achieving fully autonomous driving.

Machine learning, especially with a "factory" approach to data collection, labeling, training, and validation, is crucial for handling complexity.

Simulation plays a vital role in testing and validating self-driving systems on a massive scale, covering billions of miles.

Hybrid systems, combining machine learning with expert-designed algorithms and complementary sensors, are essential for robustness, especially in uncertain scenarios.

Developing realistic agent behaviors for simulation, particularly for pedestrians and other drivers, is critical for effective testing.

Automated machine learning (AutoML) is used to optimize neural network architectures for performance and efficiency.

THE CHALLENGE OF THE LONG TAIL

Drago Anguelov from Waymo highlights that achieving fully autonomous driving requires addressing the 'long tail' of rare, unpredictable events. While common driving scenarios are manageable, the vast number of infrequent situations, like bizarre objects on the road or rule-breaking drivers, pose the greatest hurdle. Successfully taming this long tail is essential for enabling safe and scalable self-driving capabilities.

CORE COMPONENTS: PERCEPTION, PREDICTION, AND PLANNING

Autonomous driving relies on three core AI tasks: perception, prediction, and planning. Perception involves interpreting sensor data to understand the environment, identifying objects, and mapping scenes. Prediction focuses on anticipating the future behavior of other road users, considering past actions, semantic context, and subtle cues. Planning then generates safe, comfortable, and efficient vehicle actions based on these inputs.

MACHINE LEARNING AS A SCALABLE SOLUTION

Modern machine learning is presented as a powerful tool for tackling the complexity of autonomous driving, akin to a 'factory' process. This involves building robust infrastructure for data collection, labeling, training, and validation. By feeding large datasets into this factory, Waymo can iteratively develop and improve machine learning models that handle intricate mappings and diverse scenarios, essential for addressing the long tail.

THE ML FACTORY: INFRASTRUCTURE, DATA, AND MODELS

Waymo's 'ML factory' comprises key ingredients: computing infrastructure (leveraging TensorFlow, data centers, and specialized hardware), high-quality labeled data, and advanced models. Data selection is critical, focusing on rare and interesting cases through techniques like active learning and data mining. Collaboration with Google and DeepMind provides access to cutting-edge AI research and model architectures, enhancing perception and decision-making capabilities.

AUTOMATED MACHINE LEARNING AND HYBRID SYSTEMS

Automated machine learning (AutoML) is employed to optimize neural network architectures, finding efficient and high-performing models for tasks like lidar segmentation and lane detection. Complementing this, hybrid systems integrate machine learning with expert-designed algorithms and redundant sensors (camera, lidar, radar). This approach enhances robustness, allowing safe operation even when ML models are uncertain or encounter novel situations.

IMMERSIVE SIMULATION FOR LARGE-SCALE TESTING

To test and validate self-driving systems rigorously, Waymo utilizes extensive simulation, equivalent to billions of miles driven virtually each day. This simulation generates vast numbers of scenarios, including those derived from real-world logs and custom-designed situations. Simulating realistic agent behaviors, such as those of pedestrians and other drivers, is crucial for creating a believable and effective testing environment.

MODELING AGENT BEHAVIOR AND THE LONG TAIL OF TESTING

Modeling realistic driver and pedestrian behavior is key to effective simulation. Techniques range from simple 'break and swerve' models to complex learned agents that mimic real-world interactions. End-to-end driving models, trained on extensive data, show promise but still struggle with the very edge cases in testing. Trajectory optimization, informed by observed behaviors and learned potentials, offers a more constrained yet robust approach to simulating diverse agent interactions.

SCALABLE DEPLOYMENT AND CONTINUOUS IMPROVEMENT

Scaling self-driving capabilities to numerous cities requires a systematic approach. This involves driving extensively in new environments to collect data, enabling the system to quantify its uncertainty and identify areas for improvement. The goal is a virtuous cycle where data collection, retraining, and deployment lead to continuous enhancement, supported by scalable training and testing infrastructure and the ability for systems to reason and self-update.

Mentioned in This Episode

●Software & Apps

●Companies

●Organizations

●Studies Cited

●People Referenced

Common Questions

The 'long tail' refers to the vast number of rare, unusual, and challenging situations that autonomous vehicles must be able to handle safely, beyond the common driving scenarios. Taming this 'long tail' is crucial for achieving truly driverless operation at scale.

Topics

AI & Machine Learning Technology & Innovation Data Collection Autonomous Driving Machine Learning Perception Systems Prediction Models Model Testing Long Tail Problem

Mentioned in this video

People

Stefan Ross

Mentioned as the developer of the 'dagger' problem in imitation learning at Waymo.

Daniel Kahneman

Mentioned in relation to 'Type 1' and 'Type 2' reasoning in humans, contrasting it with current AI capabilities.

Andrew Ng

Mentioned in the context of papers on weakly supervised learning and imitation learning.

Daphne Koller

Drago Anguelov's PhD advisor at Stanford, a prominent figure in AI and machine learning.

Sebastian Thrun

A pioneer in autonomous driving and robotics, who founded Waymo (formerly Google's self-driving car project).

Products

SSD

A state-of-the-art object detection model developed at Google, known for its speed and accuracy.

Software & Apps

TensorFlow

A deep learning framework developed by Google, leveraged by Waymo for its machine learning infrastructure.

Inception

A neural network architecture invented at Google, which became popular and improved object detection.

AutoML

Automatic machine learning, a system developed at Google that allows machines to search for and optimize neural network architectures.

ImageNet

A large-scale visual database used for training deep neural network models, where Google and DeepMind have achieved state-of-the-art results.

Companies

Waymo

A self-driving technology company, part of Alphabet, celebrating its 10-year anniversary. They aim to lead the world in autonomous driving.

DeepMind

An AI research laboratory owned by Google, collaborating with Waymo to improve its models and advance AI in areas like perception and reinforcement learning.

Google

The parent company of Waymo and DeepMind, providing infrastructure and AI expertise, including TensorFlow and specialized hardware for model training.