Subject

While modern robots have made impressive progress, they still fall short of human capabilities—particularly when it comes to learning efficiently and adapting to unfamiliar situations. State-of-the-art data-driven methods, such as deep reinforcement learning, require vast amounts of training data and often struggle to generalize beyond narrowly defined scenarios. In contrast, humans—especially children—learn about the world through direct interaction, exploration, and qualitative reasoning. Inspired by this, our research explores an alternative approach to robotic learning: building world models using qualitative relationships derived from interaction and experimentation. Instead of relying solely on numeric representations, qualitative models express causal relations in a more abstract and general form. Examples include: - “Turning the steering wheel clockwise makes the car turn right.” - “Pressing the gas pedal increases the car’s speed.” - “Picking up an object allows it to be moved elsewhere.” These models can often be learned with fewer interactions than traditional reinforcement learning techniques and are typically more transferable across tasks and domains. Once acquired, the models are used to guide the robot in executing complex tasks in a goal-directed manner.

Kind of work

In this master's thesis, the student will:
- Study and implement our qualitative-model-based learning approach,
- Work with our existing codebase and adapt it to a new setting,
- Apply the method to a selected challenge—either a simulation task (e.g., OpenAI Gym) or a real robotic setup, depending on the student's preference,
- Compare the performance and generalizability of the qualitative approach against reinforcement learning baselines.

Framework of the Thesis

For more information, consult http://parallel.vub.ac.be/learningrobots/

Number of Students

2

Expected Student Profile

This thesis is suitable for students interested in robotics, AI, machine learning, and cognitive-inspired models. A solid programming background in Python and a basic understanding of machine learning are recommended.