Robot Learning 2026 – Lecture 6: Generative Models | ETH Zürich

Robotic action prediction requires modeling multimodal distributions rather than deterministic functions, as tasks often involve stochasticity and expert inconsistency. Generative models address this by mapping simple noise distributions to complex data distributions. Autoencoders establish the foundation by compressing data into latent representations, while Variational Autoencoders (VAEs) introduce probabilistic structure, enabling sampling through the reparameterization trick. Vector Quantized VAEs further refine this by using discrete codebooks to eliminate posterior collapse. Diffusion models improve performance by iteratively denoising data, with techniques like classifier-free guidance allowing for controlled generation. Finally, flow matching simplifies the process by learning velocity fields that transport noise to data along straight paths. These frameworks, particularly diffusion and flow matching, have become essential for modern robot learning, enabling more robust and scalable policies for complex manipulation tasks.