Robotics 2: PD Control and Stability Analysis (Lecture 14b) | Prof. De Luca

Added:

Model-Free Reinforcement Learning and Robust Adaptive Control

While classical PD control with gravity compensation and Lyapunov stability analysis offers elegant mathematical guarantees, it relies heavily on having an accurate dynamic model of the robot. In practice, obtaining precise models for complex, flexible, or soft robots—or those operating in highly unstructured environments—is extremely difficult. Model-free approaches, such as Deep Reinforcement Learning (DRL) and Robust Adaptive Control (e.g., Sliding Mode Control), present a major alternative. DRL allows robots to learn optimal control policies directly through trial-and-error interaction with the environment, bypassing the need for explicit gravity compensation or analytical model formulation. Critics of traditional Lyapunov-based PD control argue that its strict stability proofs can be fragile under real-world uncertainties, unmodeled friction, and payload changes, making data-driven and adaptive schemes more practical and resilient for modern, highly dynamic robotic tasks.

Euler-Lagrange formulation of robot dynamics, specifically understanding joint inertia matrices, Coriolis/centrifugal terms, and gravity vectors.
Fundamentals of classical control theory, particularly linear Proportional-Derivative (PD) control and the concepts of closed-loop systems.
Basic Lyapunov stability theory, including positive/negative definite functions and the definition of asymptotic stability.
State-space representation of multi-input multi-output (MIMO) physical systems.
Computed Torque Control (Feedback Linearization) for high-performance trajectory tracking.
Adaptive and Robust control strategies to deal with payload uncertainties and unmodeled robot dynamics.
LaSalle's Invariance Principle for proving asymptotic stability in robot regulation tasks when the Lyapunov derivative is only negative semi-definite.
Force and interaction control schemes, such as Impedance and Hybrid Force/Motion Control, for contact-heavy robotic tasks.
2.5K views0likes1:58:16@videodiag-sapienzauniversi4757Original Release: 2020-04-23

This lecture covers the theoretical foundations and practical implementation of PD (Proportional-Derivative) and PID (Proportional-Integral-Derivative) control strategies for robotic manipulators. The PD controller u = KP × e - KD × q̇ achieves global asymptotic stability for desired equilibrium states when gravitational effects are absent, using Lyapunov-based analysis with a candidate function combining kinetic energy and quadratic position error terms. When gravity is present, constant gravity compensation at the desired configuration (u = KP × e - KD × q̇ + G(qd)) ensures global asymptotic stability if the minimum eigenvalue of KP exceeds a bound α related to the gravity gradient norm. The lecture addresses practical implementation challenges including decentralized control with diagonal gain matrices, velocity estimation through numerical differentiation, and realization of non-proper transfer functions. PID control extends PD with an integral term to eliminate steady-state error from gravity mismatches, with stability conditions derived using the Routh-Hurwitz criterion. Saturated PID control prevents integral windup while maintaining robustness against modeling uncertainties.