CS 6756: Advanced Topics in Robotics: Learning for Perception and Planning

CS 6756: Advanced Topics in Robotics: Learning for Perception and Planning (Fall 2013)

Instructor: Ashutosh Saxena.
Tue, Thu: 1:30-2:30pm.
Upson 315.

Class webpage: http://www.cs.cornell.edu/~asaxena/cs6756/
Discussion, announcements: https://piazza.com/class#fall2013/cs6756.

Description

The format of the course will be short lectures and paper presentations, followed by discussion and case studies on the applications.

This course focuses on the recent developments in learning algorithms, with applications to perception and planning tasks. Currently, most works in robotics take a piecemeal approach, where they address these two tasks in isolation. In this course, we will study how to take a unified approach to perception and planning. Specifically, we will study: (a) figuring out the right representation (or learning one), (b) how to handle data, or in other words, convert perception and planning problems into data-driven problems, (c) what should be the goal or criterion of learning (e.g., for planning, just a geometric objective may not be relevant), and (d) how to learn, i.e., useful machine learning algorithms.

Our motivating applications will be 3D Perception and mobile manipulation, but the ideas and machine learning techniques discussed in the course should be applicable to more general AI agents that need to perceive and act.

In each of the following (tentative) course topics, we will start from some background material, and quickly move on to discussing recent papers in respective fields:

Graphical Models (e.g., conditional random fields), learning and inference methods. Applications: Modeling environments (context).
Unsupervised representation learning (e.g., topic modeling, Infinite Latent CRFs, deep learning). Applications: Learning hidden human context, learning 3D features.
Learning human preferences: co-active learning, Inverse reinforcement learning, imitation learning. Applications: robot manipulation, path planning.
Intersection of perception and planning: belief-space planning (e.g., POMDP), planning with uncertaintity.
Importance of representation in other tasks: HRI, language, etc. for robots.

Paper reading list: Please join Piazza for the updated list and discussion.

Pre-requisities

You should be a PhD student. Seniors and Masters students could take the course only with prior instructor approval.

Necessary: A Machine learning course (e.g., CS 4758/6758, CS 4780 or CS 6780).
Necessary: Good knowledge of statistics, linear algebra, and programming.
Nice to have: Knowledge of basic computer vision is nice to have, but not necessary.
Nice to have: Knowledge of a intro to AI course (e.g., CS 4700) is recommended, but not necessary.
Nice to have: Knowledge of some robotics (e.g., CS 4758) is recommended, but not necessary.

CS 6756 satisfies the breadth, applied, and project requirement for CS PhD students, and can also be used for minor for non-CS PhD students.

Grading

This is a research class with focus on paper reading, discussion and a small project. (The project could be combined with research credits or graduate research with instructor permission, but not with another class.)

Project: 40%.
Paper reading/discussion/presentation: 40%.
Class participation and others (TBD): 20%.
No homeworks!

Useful links

Main conferences: NIPS 2012 proceedings, CVPR 2013 proceedings, RSS 2013 proceedings.
RGB-D workshop at RSS
Fall 2011 offering of CS6756 (You're allowed to re-take the course.)
PCL
OpenCV