6.4210 Fall 2023 Lecture 21: Multibody Parameter Estimation

The lecture is a masterclass in connecting classical system identification with modern machine learning, delivered by a leading expert in the field. Tedrake’s pedagogical approach is clear and rigorous, starting with a high-level motivation and progressively diving into mathematical details. He effectively uses historical and contemporary examples (WHAM, TossingBot) to illustrate the practical relevance of the concepts. The argumentation is solid: he justifies the use of physics-based models by highlighting their advantages for control, such as the availability of powerful tools like LQR and trajectory optimization, and the ability to generalize with few data points. The lecture is well-structured, with a logical flow from problem formulation to solution methods, including the derivation of the regressor form and the concept of base parameters. Tedrake also addresses important nuances, such as the need for exciting trajectories and the impact of model mismatch. The sources cited are credible, including classic papers by Slotine and Atkeson, and the lecture is consistent with established literature. The only minor weakness is that the lecture is part of a course and assumes prior knowledge, but this does not detract from its quality. Overall, this is an excellent resource for anyone interested in system identification for robotics.

The title accurately reflects the lecture's focus on multibody parameter estimation, with a clear connection to control and learning.

9/10

Lecture by a leading MIT professor, based on established multibody dynamics and system identification theory, with clear derivations and references to classic and modern works.

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

• 00:00Introduction and connection between machine learning and physics-based modeling.
• 05:00Discussion on behavior cloning and the importance of plant models for control.
• 10:00Motivation with TossingBot and WHAM examples.
• 15:00Formulation of the parameter estimation problem using manipulator equations.
• 20:00Introduction of the regressor form and linearity in parameters.
• 25:00Discussion on base parameters and identifiability.
• 30:00Least squares estimation and recursive methods.
• 35:00Practical considerations: sensor noise, model mismatch, and excitation.
• 40:00Trade-offs between model complexity and control performance.
• 45:00Conclusion and future directions.

• TossingBot: Learning to Throw Arbitrary Objects with Residual Physics — Mentioned as a modern example of learning-based throwing with residual physics.
• The Whole-Arm Manipulator (WHAM) — Historical example of a robot that estimates inertial parameters of unknown objects.

• TossingBot: Learning to Throw Arbitrary Objects with Residual Physics — Supports the idea that physics-based priors improve learning and generalization.