Lecture 22: MIT 6.832 Underactuated Robotics (Spring 2022) | " Stochastic /Robust Control"

The lecture provides a rigorous and comprehensive introduction to stochastic and robust control, building on previous material. The instructor clearly explains the motivations and trade-offs between different performance metrics, such as expected cost versus worst-case guarantees. The derivation of stochastic LQR is well-structured, highlighting the key result that the optimal cost-to-go retains a quadratic form plus a constant term that grows unboundedly, which is a crucial insight for understanding infinite-horizon stochastic control. The discussion of dissipation inequalities is particularly valuable, as it offers a framework for analyzing robustness without requiring absolute guarantees. The lecture is well-paced and includes intuitive examples, such as the windy path planning scenario, to illustrate abstract concepts. However, the presentation is dense and assumes a strong background in optimal control and dynamic programming. The lack of explicit citations in the transcript is a minor weakness, but the content aligns with established literature in the field. Overall, the lecture is of high quality, offering deep insights into stochastic and robust control, and is suitable for advanced students or researchers.

The title accurately reflects the content, which focuses on stochastic and robust control methods for underactuated robotics.

8/10

The lecture is part of MIT OpenCourseWare, presented by a recognized expert in the field. The content is mathematically rigorous, with derivations and references to established methods. The presentation is clear and well-structured, though it lacks formal citations in the transcript.

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

• 00:00Introduction to stochastic and robust control, motivation with windy path example.
• 05:00Roadmap of performance metrics: expected cost, worst-case, chance constraints, gain bounds, regret.
• 10:00Why expected cost is dominant: compatibility with dynamic programming and Bellman recursion.
• 15:00Stochastic LQR formulation: assumptions on noise, cost function, and infinite horizon cost issue.
• 20:00Derivation of optimal cost-to-go for stochastic LQR: quadratic form plus constant term.
• 25:00Discussion of unbounded cost in infinite horizon and possible remedies (discounting, average cost).
• 30:00Introduction to dissipation inequalities as a tool for robust analysis.
• 35:00Examples and applications of gain bounds and robustness metrics.
• 40:00Further discussion on chance constraints and risk metrics.
• 45:00Conclusion and summary of key concepts.