EG554W: INDUSTRIAL ROBOT PROGRAMMING AND LEARNING (2021-2022)

The course spans from manual teach-pendent programming of industrial manipulators to reinforcement learning for autonomous robots. It covers the path planning problem and its (un-)informed graph search solutions as well as the machine learning problem and the concept of overfitting. The course begins with a classification of programming approaches for industrial robots. Later, the characteristics of maximum likelihood and maximum a posteriori estimation are compared just as those of (Gaussian) naïve Bayes and logistic regression classifiers. An overview on the units of multi-layer neural networks and backpropagation is given. Finally, the differences and similarities between value-based, policy-based and actor-critic reinforcement learning techniques are conveyed.

Course Content

1. Introduction – Nomenclature, history, state of the art, challenges, course logistics
2. Robot-oriented Programming – Teach pendant programming, hand guiding, text-based programming, graphical offline programming
3. Task-Oriented Programming – Roadmap generation, uninformed graph search, informed graph search
4. Decision Trees – Definition, decision tree learning, overfitting, receiver-operating characteristic
5. Generative Probabilistic Classifiers – MLE, MAP, naïve Bayes, Gaussian naïve Bayes
6. Discriminative Probabilistic Classifiers – Logistic regression, gradient ascent, generative vs. discriminative classifiers
7. Artificial Neural Networks – Parameterized units, multilayer networks, backpropagation
8. Task-oriented Programming under Uncertainty – Markov Decision Process, utility, value iteration, policy iteration
9. Reinforcement Learning – Monte-Carlo and temporal-difference learning, Q-learning, DQN, policy-based, actor-critic

Information on contact teaching time is available from the course guide.

More Information about Week Numbers

There are no assessments for this course.