ECE 531/COS 531: Robot Planning Meets Machine Learning
Princeton University, Spring 2027
Instructor: Tom Silver
Time: Tuesday, Thursday 10:40 AM-12:00 PM
Location: To be announced
Previous offerings: Fall 2025
Course description: Planning and learning are essential for intelligent robots. In this graduate course, we consider how planning and learning can be combined to accomplish difficult tasks over long time horizons with sparse feedback, complicated constraints, and multiple forms of uncertainty. The course is organized into units. Each unit pairs a lecture on one kind of planning with a student-led paper session on machine learning for that kind of planning. The units are grouped into three parts according to the role that learning plays:
- Learning to Make Planning Possible. The robot learns the models that planning requires.
- Learning to Make Planning Fast. The robot has models and learns to plan with them more efficiently.
- Planning to Learn. The robot plans its own data collection so that it learns more from its experience.
Key questions: Throughout the course, we revisit the following questions:
- What learning and planning should happen in the factory vs. the wild?
- Where on the learning-planning spectrum should robots be?
- What planning models should be learned?
- How should we taxonomize learning-for-planning methods?
- What objectives should we use to train and evaluate these methods?
- To what extent are learning and planning necessary for general-purpose robots?
Prerequisites: This course is designed for graduate students and advanced undergraduates. Students should be very comfortable with Python. Machine learning experience at the level of ECE 364 or COS 324 ("Introduction to Machine Learning") is expected. The following prerequisites are recommended but not required: ECE 346 / COS 348 / MAE 346 ("Intelligent Robotic Systems") and MAE 345 / ECE 345 / COS 346 / MAE 549 ("Introduction to Robotics").
Course structure:
- Lectures. The instructor lectures on a planning problem and its standard algorithms.
- Paper sessions. Each session is organized around one anchor paper. A team of students leads the session, and each team member takes a different role, such as a historian who places the paper in the context of earlier work or an investor who assesses whether the approach can scale. Each student presents multiple papers in different roles during the semester; the number depends on course enrollment.
- Problem sets. Three problem sets, primarily coding-based (Python), cover the planning material from lectures.
- Final project. Students work in teams on a research project at the intersection of planning and learning. Teams submit a proposal, two written updates, and a final report, and they present in class during the last week.