COL 864: Learning/AI for Cognitive Robot Intelligence

Description

This course will introduce students to the area of learning/AI-based robotics. The course studies the computational aspects of intelligent robotic systems that can sense, reason and act in the physical world often interacting with human partners. This course discusses topics such as probabilistic state estimation, planning and acting under uncertainty, human-robot interaction, learning from reinforcement & demonstration, world modeling etc. The course will include a brief review of basic tools, cover foundational models and overview contemporary techniques. This course will include student paper presentation/review and hands-on exercises. Prerequisites include an introductory machine learning or an AI course.

Prerequisites

Prerequisites include an introductory machine learning or an AI course. Strong background in topics such as Bayesian networks, factor graphs, basic deep learning models, classical search, Markov models, basic RL models etc. The course would require programming skills and ability to work with standard tools.

Course Information:

• Instructor: Rohan Paul
• Classes: Monday and Thursday. Time: 3:30-5:00PM. Location: IIA-201.
• Teaching Assistant: Mr. Vikas Upadhyay (vikas.upadhyay@cse.iitd.ac.in).

Announcements:

• The course has been completed for AGP students.
• The course will resume for non-graduating students once instructions are received from the Dean's office.
• Mid-term exam: 14/03/20. Location: Bharti Seminar Room IIA - 501. Time: 3:00pm - 4:30pm. You may bring your hand written class notes to the examination.
• Homework II is up. Shreshth Tuli will be the TA for this homework exercise.

Topics (covered)

• Introduction
  • Topics: Embodied systems. Challenges. Role of AI/Learning.
  • References: Slides


• Physical Agent Representation
  • Topics: Sensor. World model. Agent architectures.
  • References: Slides, AIMA (Ch. 2, 25)


• Estimating the World State
  • Topics: Recursive Bayes estimation. Bayes Filter. Kalman Filter.
  • Readings: PR (Ch. 1, 2), AIMA (Ch. 15, 25).


• Planning in Symbolic Worlds
  • Topics: Planning representations. Planning as search (heuristic search). Graph Plan.
  • Readings: Relevant sections from AIMA (Ch. 3 and 10). PA (Ch. 2). Behaviour Trees Ch 1.


• Planning under Uncertainty
  • Topics: MDP review. Value/Policy Iteration.
  • Readings: PR (Ch. 14)
  • Additional references: Slides.


• Learning from Reinforcement (A Review)
  • Topics: Reinforcement learning review. Q-learning/Policy learning. Bandit models.
  • References: SB (Ch. 2, 3 and 4)


• Imitation Learning
  • Topics: Behaviour cloning. Data Aggregation (DAGGER). Inverse Reinforcement Learning (IRL).
  • References: Slides


• Understanding Human Intent
  • Topics: Language to specifications. Grounding graphs. Modeling intent.


• Semantic Mapping and Perception
  • Topics: Place recognition and Loop Closing

References

• [AIMA] Artificial intelligence: a modern approach. Russell, Stuart J., and Peter Norvig. Link.
• [PR] Probabilistic robotics. Thrun, Sebastian, Wolfram Burgard, and Dieter Fox. Link. Online.
• [PA] Planning Algorithms. LaValle, S.M., 2006. Cambridge University Press. Online.
• [SB] Reinforcement Learning (Second Edition). Richard Sutton and Andrew Barto. MIT Press. 2018. Online.

Submission Information

• Honor code. Cases of copying a homework/assignment will be awarded zero on an assignment. Institute guidelines will apply in handling cases of severe copying. Department guidelines for checking plagiarism will apply.
• Buffer days may be allowed for an assignment during which there is no penalty is levied. However, any delays beyond the buffer days will result in a 20% penalty for each day.

Course Components and Tentative Weightage (AGP Track)

• Homework I (20%).
• Homework II (20%).
• Paper Survey and Presentation (20%).
• Exam (40%)