COMP417: Intro to Robotics and Intelligent Systems, Winter 2017
OverviewThis course provides an introduction to robotic systems from a computational perspective. A robot is regarded as an intelligent computer that can use sensors and act on the world. We will consider the definitional problems in robotics and look at how they are being solved in practice and by the research community. The emphasis is on algorithms, probabilistic reasoning, optimization, inference mechanisms, and behavior strategies, as opposed to electromechanical systems design. This course aims to help students improve their probabilistic modeling skills and instill the idea that a robot that explicitly accounts for its uncertainty works better than a robot that does not.
Office: McConnell 403
Office Hours: Thu 10:30-11:30am
Office: McConnell 403
Office Hours: Wed 3-4pm
Office: McConnell 111
Office Hours: Mon 1-2pm
X = cim.mcgill.ca, Y = mail.mcgill.ca
Course DescriptionThis course will broadly cover the following areas:
- Kinematics and Dynamics: how can we model robotic systems using approximate physical models that enable us to make predictions about how robots move in response to given commands?
- Feedback Control and Planning: how can we compute the state-(in)dependent commands that can bring a robotic system from its current state to a desired state?
- Mapping: how can we combine noisy measurements from sensors with the robot’s pose to build a map of the environment?
- State Estimation: the state of the robot is not always directly measurable/observable. How can we determine the relative weighs of multiple sensor measurements in order to form an accurate estimate of the (hidden) state?
- The Geometry of Computer Vision: how can modeling pixel projections on an RGB camera help us infer the 3D structure of the world? How can we triangulate points seen from two cameras? How can we estimate the camera’s pose (and therefore the robot’s) while it is moving in the environment?
- (Intro to) Reinforcement Learning: how can we learn the parameters of a state-dependent controller without having a prior physical model of the robot’s dynamics? This is an enormous research field, with many exciting results. In this course we will only have time to see two algorithms.
Motivation, logistics, rough description of assignments, sense-plan-act paradigm.
Syllabus, Quiz 0 (Introduction, Background, Expectations)
Dynamics (lecture given by Martin Gerdzhev)
Dynamical systems and control. Examples: Dubins car, differential drive car, unicycle, pendulum, cartpole, quadcopter. Holonomic vs. non-holonomic systems.
|pdf, pptx||Lavalle 13.1
Dudek & Jenkin 3.1.5,6
Kinematics (lecture given by Martin Gerdzhev)
Frames of reference. Rotation representations. Homogeneous coordinates and transformations. Rigid body motion.
|Intro to ROS||pdf, pptx||Paul Furgale: robot pose
Sensors and Actuators
Observation models for the following sensors: cameras, lasers, tactile, IMU, depth, GPS, Hall-effect, encoders, RGBD. Pulse-Width Modulation.
|pdf, pptx||Dudek & Jenkin 3.1.1,4, 3.2-3, 4.1-8, 4.10, 5.1.1
Optional: Mike Langer's notes
Tuning, cascading PID, advantages and drawbacks.
|Linear algebra refresher||pdf, pptx||Optional: Astrom and Hagglund, Ch. 2|
Artificial Potential Fields and Obstacle Avoidance
Implementation issues, navigation functions. Vector-field histogram (VFH), dynamic window approach (DWA).
|pdf, pptx||Lavalle Ch. 8.4
Dudek & Jenkin 6.3.4
Optional: Howie Choset's notes
Linear Quadratic Regulator (LQR)
Computing optimal actions for linear dynamical systems with quadratic cost-to-go functions.
Quiz 1 (Transformations and PID)
|Probability refresher||pdf, pptx, code||Optional: Stephen Boyd's LQR notes and examples.|
|pdf, pptx||Blog post on A*
Udacity Lesson 4
Rapidly-exploring Random Trees (RRT), Probabilistic RoadMaps (PRM)
|pdf, pptx||Lavalle 5.5, 5.6
Map Representations and Map Alignment
Occupancy grids, Octrees, Voronoi Graph, Homotopy Classes. Map alignment with known or unknown correspondences. Iterative Closest Point (ICP).
Quiz 2 (Potential fields and LQR)
|pdf, pptx||Pieter Abbeel's notes|
Occupancy Grid Mapping With Known Robot Poses
Log-odds ratio, Probabilistic dynamics and measurement models, Bayesian estimation.
|Intro to numpy||pdf, pptx||Pieter Abbeel's notes
Probabilistic Robotics Ch. 2 and Ch. 9
Maximum Likelihood, Least Squares Estimation, Maximum A Posteriori Estimation
Least squares as a special case of maximum likelihood estimation on Gaussian models.
|pdf, pptx||Optional: Simon Prince Ch.2 and Ch. 4|
Expectation and Covariance. Geometric interpretation of the covariance matrix. Nonlinear Least Squares formulation of the Simultaneous Localization And Mapping (SLAM) problem.
Quiz 3 (Map representations and Bayes' rule)
|pdf, pptx||Udacity Lesson 6
Probabilistic Robotics Ch. 11
||Midterm review session|
Bayes' rule on Gaussian distributions. Example of 1D Kalman Filter.
|pdf, pptx||Udacity Lesson 2
Kalman Filter, Illustrated
Bayes' Filter and Kalman Filter
Kalman Filter as a special case of Bayes' Filter. Examples of 2D and 4D Kalman Filter. General prediction and update equations.
|pdf, pptx||Probabilistic Robotics Ch. 2,3|
Extended Kalman Filter (EKF)
Bayes' Filter and nonlinear transformations. Monte Carlo sampling vs. Linearization. EKF prediction and update equations. Examples: EKF Localization and EKF SLAM.
Quiz 4 (GraphSLAM and Gaussians)
|Basic Kalman Filter implementation||pdf, pptx||Cyrill Stachniss' intro to EKF
Cyrill Stachniss' intro to EKF-SLAM
Probabilistic Robotics Ch. 2,3
Representing multimodal distributions. Particle propagation and resampling. Pathologies of particle filter.
|pdf, pptx||Udacity Lesson 3
Importance Sampling. Examples: Markov localization in a known map. FastSLAM.
|pdf, pptx||Optional: Thrun's paper on PF|
Camera Optics and Multi-view Geometry
Pinhole cameras, lenses, perspective projection. Aperture, focal length, exposure time, depth-of-field. Structure from Motion.
|pdf, pptx||Optional: James Tompkin's notes
Visual odometry and Visual SLAM
Epipolar constraints. Depth from stereo disparity for parallel cameras. Triangulation as a least-squares problem. Scale issues in visual odometry with a single camera. Visual SLAM vs. structure from motion.
Quiz 5 (KF/EKF)
|pdf, pptx||Optional: James Tompkin's notes on stereo and
Sanja Fidler's notes on depth from stereo
Intro to Reinforcement Learning
Research highlights (non-examinable material). Model-free RL: policy gradient estimation and the cross-entropy method.
|Markov Localization||pdf, pptx
|Optional: Pieter Abbeel's policy optimization notes|
Intro to Reinforcement Learning (invited talk by Juan Camilo Gamboa Higuera)
Research highlights (non-examinable material). Model-based reinforcement learning. Learning to swim on the Aqua robot.
Intro to Human-Robot Interaction (invited talk by Anqi Xu)
Research highlights (non-examinable material). Modeling human trust, and trust-aware control.
|Assignment 4 Discussion||pdf, pptx|
Review session for final exam
- Wall following and intro to ROS with starter code here. Due Feb 4.
- A*, RRT, and LQR with starter code here. Due Feb 18.
- Occupancy grid mapping, least squares localization, and EKF with starter code here. Bonus question: GraphSLAM implementation. Due Mar 24.
- Depth from stereo disparity, Markov localization. Starter code is here. Bonus question: GPU implementation of depth from disparity. Due Apr 10.
- 4 assignments worth 12.5% each = 50%
- 5 quizzes worth 1% each = 5%
- 1 midterm exam worth 15%
- 1 final exam worth 30%
- The final exam grade can replace the midterm grade if it improves the student's final mark.
Recommended, but optional, textbooks
- Computational Principles of Mobile Robotics, 2nd edition, by Dudek and Jenkin
- Probabilistic Robotics, by Thrun, Fox, and Burgard
- Planning Algorithms, by Lavalle
- Robotics, Vision, and Control, by Corke
- Introduction to Autonomous Mobile Robots, by Siegwart, Nourbakhsh, Scaramuzza
- (Chapters 2 and 4 from) Computer Vision: Models, Learning, and Inference, by Prince