Object-Centric Representation Learning in Embodied Reinforcement Learning

Agents which understand their environment in terms of objects, and their properties, dynamics and interactions, may learn more powerful and generalisable representations than contemporary Deep Reinforcement Learning systems. This ontological commitment underpins the formal logic utilised in Logic-based AI methods; whose synthesis (either through inductive bias or differentiable combination) with modern Deep Learning systems forms the primary focus of our investigation.

Research area: Artificial intelligence technologies