Mini-Lecture 4 | MIT 6.832 (Underactuated Robotics), Spring 2021

This mini-lecture provides a clear and insightful comparison between discrete-time and continuous-time LQR, a fundamental topic in optimal control. The instructor, presumably Russ Tedrake, demonstrates deep expertise and pedagogical skill. The content is mathematically rigorous, with derivations performed on a virtual whiteboard, and numerical examples are used to reinforce the theoretical findings. The key contribution is the explanation of why the discrete-time cost-to-go is always higher than the continuous-time one, attributing it to the constraint of holding control actions constant over time steps. This is a subtle point that is often glossed over in textbooks, and the lecture does an excellent job of making it intuitive. The argumentation is solid, and the use of the double integrator as a simple yet illustrative example is effective. The sources are not explicitly cited in the video, but the content is based on standard control theory and the instructor’s own course notes, which are available online. The video is well-structured, with a clear progression from recap to new material. The only minor weakness is that the video is a mini-lecture, so it does not cover all aspects of LQR, but it serves its purpose as a focused discussion. The title accurately reflects the content. Overall, this is a high-quality educational resource for students of robotics and control theory.

The title accurately describes the content: a mini-lecture from MIT's Underactuated Robotics course, covering material from Lecture 4.

8/10

Lecture from MIT OpenCourseWare, presented by an expert in robotics, with clear mathematical derivations and numerical demonstrations. The content is rigorous and well-structured, though it is a mini-lecture and not a peer-reviewed publication.

Markers derived by PSI from the transcript: the creator did not define chapters.

• 00:00Introduction and recap of previous lecture topics
• 02:00Discussion of the Hamilton-Jacobi-Bellman sufficiency theorem and technical conditions
• 04:00Introduction of the linear quadratic regulator (LQR) and its relevance in reinforcement learning
• 06:00Setup of the double integrator example and continuous-time LQR formulation
• 08:00Derivation of the discrete-time LQR using Euler integration
• 10:00Key insight: discrete-time cost-to-go is always greater than continuous-time for any positive time step
• 12:00Explanation of why discrete-time costs more: constraint of holding control actions constant
• 14:00Numerical demonstration of the cost difference using MATLAB-like code
• 16:00Plotting eigenvalues and eigenvectors of the cost-to-go matrices
• 18:00Discussion of the limit as time step goes to zero and conclusion

• MIT 6.832 Underactuated Robotics Course Notes — The lecture refers to the course notes for technical details and derivations.

• MIT 6.832 Underactuated Robotics Course Notes — The lecture is based on the course notes, which provide a more detailed treatment of the topics.