World Modelling
Module aims
World Modelling covers the world-model problem and its taxonomies, planning and Model Predictive Control, latent-variable and autoregressive generative models, self-supervised and representation learning, deep world models for embodied and multi-agent systems, reinforcement learning and model-based RL, and frontier topics such as vision-language-action models and latent actions. Aimed at students comfortable with deep learning, it combines lectures and invited talks with theoretical and hands-on coursework. No prior RL or generative modelling knowledge is required.
Learning outcomes
On successful completion of this module, students will be able to:
1. Design self-supervised learning principles to learn predictive representations, including approaches such as contrastive learning and the JEPA family methods.
2. Develop generative world models using approaches such as autoregressive, latent-variable, and diffusion-based modelling.
3. Design and implement planning and control methods using hand-crafted and learned world models.
4. Assess world-model approaches, identifying key limitations, trade-offs, and open research challenges.
Module syllabus
Indicative weekly content:
Week 1 – Introduction to World Models: What world models are, why they matter, and the main approaches to learning predictive models of the world.
Week 2 – Planning and Model Predictive Control: Learning and using dynamics models to predict future states, plan towards objectives, and manage uncertainty.
Week 3 – Generative Models: Approaches for modelling and generating complex data, including images, video, and sequential observations.
Week 4 – Self-Supervised Learning: Learning useful representations through prediction, including masked and latent-prediction approaches.
Week 5 – Deep World Models: Designing and evaluating world models for embodied and multi-agent settings under partial observability.
Week 6 – Model-Based Reinforcement Learning: Learning models of the environment to support decision-making, exploration, and planning through imagined experience.
Week 7 – Emerging Directions: Recent developments and open challenges in world modelling.
Teaching methods
The module runs over seven weeks at four hours per week, comprising lectures, invited talks, and practical sessions with GTAs. Lectures develop the core theory and methods; invited talks from researchers and practitioners expose students to current work and open problems. Two coursework assignments, with focus on theory and coding. The module assumes familiarity with deep learning but introduces reinforcement learning and generative modelling from first principles.
Assessments
The module is assessed by a final exam (70%) and two coursework assignments (30% combined), combining theoretical and practical exercises. The final exam assesses conceptual understanding across generative modelling, planning and control, self-supervised representation learning, and (model-based) reinforcement learning.
Marked homework will be returned with written feedback within the College's standard turnaround. Feedback is given on both assignments and test, alongside individual feedback on submitted code and results, so that students can improve before subsequent assessments.