Machine learning (ML) is a fundamental bio-inspired approach towards building AI systems optimized for various tasks, including robot learning. In the talk, we will present several ML methods based on artificial neural networks, useful for this domain. We will cover two paradigms: first, the supervised learning that can be exploited for building world models (latent representations) important for robots in order to represent the world. Second, the reinforcement learning paradigm used for robots in order to learn desired behavior (policy).

1. Self-Supervised Learning (SSL) for Robot Perception and Representation Methods that allow robots to learn from raw sensory data without manual labeling.
(a) Representation Learning: Contrastive learning = Learn state embeddings by distinguishing positive (e.g., time-adjacent frames) from negative samples.
(b) Predictive modeling (forward models, inverse models), Multimodal alignment.

2. Reinforcement Learning (RL) for Robot Control and Decision-Making = Methods that optimize behavior through trial and error.
(a) Model-Free RL – Value-based (e.g., DQN), Policy gradient (e.g., PPO, TRPO, A2C), Actor–critic (e.g., SAC, TD3), common for continuous-control robotic tasks (grasping, locomotion).
(b) Model-Based RL – Learned dynamics models (e.g., Dreamer) = Use predicted future trajectories for planning or policy learning. Hybrid planning + learning (combine model-based planning (MPC) with learned controllers).