COMP 765: Intelligent Robotics, Fall 2017
Office: McConnell 112N
Office Hours: Tu 3-4pm
X = mcgill.ca
NewsSept 14, 2017 The list of assigned papers for student-lead presentations can be found in this Google Doc. Please sign up exactly once. Each presentations will be 15 minutes plus 5 minutes for questions. You are encouraged to explore the code, demos, videos, data etc that go with many of the papers and show those live during your presentation. Your main goal is to teach the rest of the class about the important parts of the method: what are the key steps in the math, what data structures are used and why, and what are the pros and cons? Everyone should try to read the papers at least briefly before the presentations so you come ready with some good questions.
OverviewCOMP 765 is a research seminar on Intelligent Robotics: the intersection of rocket science and the latest machine learning on-board systems that can sense the world and act upon it. The class will begin with lectures on definitional problems and algorithms in robotics. We will then transition to mixed student-lead and instructor-lead discussions of recent developments in research and in practice. The emphasis is on algorithms, probabilistic reasoning, learning to improve behaviors using data, and decision making under uncertainty, as opposed to electromechanical systems design.
We will broadly cover the following areas:
- Probabilistic Estimation: how can we represent uncertainty in our robot's sensing and motion and what are efficient algorithms that exploit these representations to accurately estimate quantities in the world?
- Classical 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, when models of the robot and its environment are available?
- Decision Making Under Uncertainty: how can the same problems be solved when our model knowledge is missing or imprecise? In particular, what motion strategies best complement our probabilistic estimators, for example by gathering useful information or acting conservatively?
- Learning for Robotics: how can we use data to learn models for a robot's continuous time series data in case these are not known in advance?
- Heirarchical Decision Making: how can tasks with longer time-scale or for larger target systems be factored productively into multiple levels of reasoning?
Schedule(Note: this will be populated as we go with the slides, additional links, and information)
Motivation, logistics, rough description of assignments, sense-plan-act paradigm, introduction to modeling perception and action.
1 - Introduction
2 - Spatial Representation
|PR Chapters 1, 5 and 6.|
Recursive Bayesian Filter derivation, Kalman filter, particle filter.
3 - Bayesian Estimation
4 - Particle and Kalman Filters
|PR Chapters 7 and 8.|
Modern Localization and Mapping
Rao-Blackwellized Particle Filtering, large-scale sparse graph optimizers, mapping an entire city, multi-sensor fusion techniques.
5 - EKF and Mapping
||PR Section 3|
Basic analysis of control system, ODE background, stability, controllability, PID controllers and tuning.
Guest Lecture on Control Intro
Daves slides on PID
Relationship between control and Reinforcement Learning, MDP formulation, dynamic programming, trajectory optimization, LQR and DDP.
6 - Optimal Control
7 - Trajectory Optimization
|PR Chapter 14.|
|Oct 10||Geometric Planning
Planning in 2D and 3D, for simple and complex systems. Solutions to the piano-movers problem Spatial decompositions, probabilistic roadmaps, RRT.
8 - Geometric Planning
||Planning Algorithms book by Lavalle|
|Oct 17||Planning Under Unceratinty
POMDP formulation, information-theoretic planning heuristics.
9 - POMDPs
||PR Chapters 16 and 17|
|Oct 24||Learning for Robotics
Machine Learning for robotics problems. Linear regression, Gaussian Processes. Active learning , black-box optimization applied to probabilistic search.
|10 - GPs for Robotics||GPforML Chapters 1-3
Iain Murray GP slides
Wolfram Burgard GP slides
|Oct 31||Learning to Control
Model-based policy search, PILCO, GPS, PDDP. Adapting behavior to changing models, MPC.
Student-lead Research Paper Presentations
Each student will present one paper they have selected.
Paradigms for imitation including LfD, IRL and AfP.
Collaborative and Cloud Robotics
Controlling a multi-robot team, centrallized vs decentralized reasoning, planning and control, communication constraints.
Invited talk from HRI expert. HRI for home robots, self-driving and robots in the workplace. Safety assessment of intelligent robots.
|Dec 5||Project Demonstrations and Wrap-Up|
AssignmentsAssignments will be short implementations of the core algorithms seen in lecture for the first 2 months. Starter code and simulation environments are provided, students must implement only the core algorithmic components and prepare a brief report describing their findings.
- 3 assignments worth 10% each = 30%
- 2 in-class research paper presentations worth 15% each = 30%
- 1 final project demo or video 10%
- 1 final project written report worth 30%
Recommended, but optional, textbooks
- Probabilistic Robotics, by Thrun, Fox, and Burgard (PR in the outline)
- Gaussian Processes for Machine Learning, by Rasmussen and Williams (GPforML in the outline)
- Computational Principles of Mobile Robotics, 2nd edition, by Dudek and Jenkin
- Planning Algorithms, by Lavalle
- Robotics, Vision, and Control, by Corke
- Introduction to Autonomous Mobile Robots, by Siegwart, Nourbakhsh, Scaramuzza