A fundamental aspect of human cognition is interpreting their surrounding environment as a collection of objects and the relations among them. What an object means to us is not only characterised by its physical properties, but crucially entails its affordances, informing us what kind of actions can be executed on and with it, and how it responds to such actions. During my DPhil, I will work on learning such meaningful object-centric representations to get closer to human-level generalisation in autonomous agents. Current approaches for learning object-centric representations are often entirely trained from visual information and consequently only encode visual features. As a result, even though two objects might share affordances and semantics, they could be very distant in latent space – prohibiting generalisation from one object to the other. In contrast, semantically sound representations support abstraction from specific problem instances towards more general task structures by enabling agents to leverage semantic similarities between scenarios for action selection, planning, and counterfactual reasoning. To this end, I will be researching methods that more explicitly incorporate action information during training, as well as methods that leverage the large amount of implicit semantic knowledge stored in current large language models.