After this course, you will be able to program your Duckiebots to navigate (without accidents!) in road lanes of a model city with rubber-duckie-pedestrian-obstacles using predominantly computer vision-based techniques.

Moreover, you will:

• recognize essential robot subsystems (sensing, actuation, computation, memory, mechanical) and describe their functions

• make your Duckiebot drive in user-specified paths

• understand how to command a robot to reach a goal position

• make your Duckiebot take driving decisions autonomously according to "traditional approaches", i.e., following the estimation, planning, control architecture

• make your Duckiebot take driving decisions autonomously according to "modern approaches" (reinforcement learning)

• process streams of images

• be able to set up an efficient software environment for robotics with state-of-the-art tools (Docker, ROS, Python)

• program your Duckiebot and make it safely drive in empty roads lanes

• program your Duckiebot and make it recognize and avoid rubber duckie obstacles

• submit your robot agents (a.k.a. "robot minds") to public challenges, and test your skills against your peers

Additional goals (require hardware)

• independently assemble a Duckiebot and a Duckietown

• remotely operate your Duckiebot and see with its eye(s)

• be able to discuss differences between theory, simulation, and real word implementation for different approaches

• experience the challenges of deploying complex autonomous robots in the real world, and reap the rewards of getting it to work

Module 0: Welcome to the course

• Welcome to the course, by Prof. Emilio Frazzoli

• You will familiarize yourself with the logistics and navigation interface of the course resources

• You will start a learning journey in the world of robot autonomy with Duckietown

Module 1: Introduction to self-driving cars

• The potentials and challenges

• Levels of autonomy

• The vision for autonomous vehicles (AVs)

• Activities: You will set up your learning environment, and your Duckiebot, and make your first challenge submission

Module 2: Towards autonomy

• Making a robot

• Sensorimotor architectures

• Stateful architectures

• Logical and physical architectures

• Application: You will create a reactive "Braitenberg" agent to avoid duckies and see how your agent compares to other submissions

Module 3: Modeling and Control

• Introduction to control systems

• Representations and models

• PID control

• Application: You will design an odometry function and PID controller to command your Duckiebot's angular velocity

Module 4: Robot Vision

• Introduction to projective geometry

• Camera modeling and calibration

• Image processing

• Application: You will develop image processing techniques necessary for visual lane servoing - controlling your Duckiebot to drive within markings

Module 5: Object Detection

• Introduction to neural networks

• Convolutional neural networks

• One and two-stage object detection

• Application: You will train a convolutional neural network (CNN) to detect duckies and integrate your model with ROS to run onboard your Duckiebot and avoid duckies

Module 6: State Estimation and Localization

• Bayes filtering framework

• Parameterized methods (Kalman filter)

• Sampling-based methods (particle and histogram filter)

• Application: You will build a state estimation algorithm combining the dynamics and sensor data of your Duckiebot in order to predict its pose as it travels through the world

Module 7: Planning I

• Formalization of the planning problem

• Application: You will create a collision checker to determine if your Duckiebot is crashing into an obstacle

Module 8: Planning II

• Graphs

• Graph search algorithms

• Application: You will tackle a variety of path-planning challenges and leverage all the skills you've built thus far to navigate your Duckiebot in a variety of simulated environments

Module 9: Learning by Reinforcement

• Markov decision processes

• Value functions

• Policy gradients

• Domain randomization

• Application: You will explore the capabilities and limitations of reinforcement learning models when applied to real-world robotics tasks such as lane following

I learned a lot. The theory was good but I mostly enjoyed gaining practical skills: I had never before used docker, ROS, GitHub, Slack, Stack Overflow, Jupyter notebooks, Colab, to name a few tools, and only very marginally Linux command line and Python. So for me, there was really a lot to take in. It was challenging and therefore all the more rewarding to see progress materialize.