Robot Learning 2026 – Lecture 2: Robot Control & Markov Decision Processes | ETH Zürich

Robot control relies on representing diverse hardware as collections of rigid bodies or articulated chains, utilizing homogeneous transformation matrices to manage coordinate frame hierarchies. Effective motion planning requires navigating configuration space (C-space) to avoid obstacles while mapping tasks to the physical workspace through forward and inverse kinematics. To ensure smooth, hardware-safe movement, controllers employ polynomial-based trajectory generation and PID feedback loops to compensate for environmental noise and physical constraints. These mechanical foundations integrate into the Markov Decision Process (MDP) framework, which models decision-making as a sequence of states, actions, and rewards. By defining policies as probability distributions over actions, robots learn to optimize long-term goals in stochastic environments, bridging the gap between low-level motor torques and high-level autonomous behavior.