Lecture 8 | MIT 6.832 (Underactuated Robotics), Spring 2018

This lecture is a masterclass in rigorous mathematical exposition. The instructor excels at motivating the need for global stability certificates, clearly articulating the limitations of sampling-based methods. The transition from linear systems to nonlinear systems is handled with care, building intuition before diving into formalities. The explanation of why quadratic Lyapunov functions are natural for linear systems is particularly illuminating, connecting back to the cost-to-go in LQR. The proof of convexity of the positive definite matrices is concise and accessible. The discussion of SDP solvers’ numerical challenges is honest and practical, providing valuable guidance for students. The lecture is well-structured, with clear signposting of the logical flow. The only minor criticism is that the lecture assumes familiarity with convex optimization and matrix inequalities, which might be challenging for some viewers. However, this is appropriate for an advanced graduate course. The content is highly reliable, coming from a leading researcher in the field, and the mathematical derivations are sound. The title accurately reflects the content, and the lecture delivers on its promise to address a major limitation of previous methods.

The title accurately reflects the content: a lecture on sums of squares optimization within the context of underactuated robotics.

9/10

Lecture from MIT's Underactuated Robotics course, presented by a leading expert in the field. The content is rigorous, mathematically grounded, and based on established convex optimization theory. The course materials are publicly available and widely used in academia.

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

• 00:00Introduction and overview of the lecture's goal: to replace sampling-based Lyapunov verification with global certificates.
• 02:00Review of previous lecture's approach: sampling points and formulating a linear program to find Lyapunov functions.
• 05:00Discussion of limitations: sampling scales poorly with dimension and does not provide a formal proof.
• 10:00Introduction of quadratic Lyapunov functions for linear systems and the condition for positive definiteness.
• 15:00Explanation of why quadratic forms are natural for linear systems, linking to the cost-to-go in LQR.
• 20:00Derivation of the Lyapunov derivative condition and its equivalence to a matrix inequality.
• 25:00Proof that the set of positive definite matrices is convex, enabling semidefinite programming.
• 30:00Introduction of semidefinite programming (SDP) as a convex optimization tool for these problems.
• 35:00Practical considerations: numerical conditioning and the maturity of SDP solvers compared to LP solvers.
• 40:00Transition to sums of squares optimization for polynomial Lyapunov functions.

• Underactuated Robotics Course Website — Course materials, including lecture notes and assignments, referenced in the video description.

• Underactuated Robotics Course Website — The course website provides lecture notes and additional resources that align with the content of this lecture.