AXIOM — Active eXpanding Inference with Object-centric Models — is a model-based agent that learns to play Atari-like games in minutes by combining:

• Online structure learning. The model grows new mixture components as it encounters novel objects and dynamics, rather than fixing capacity ahead of time.
• Object-centric perception. A mixture model decomposes pixel observations into a small set of interpretable object slots.
• Model-based planning. Action selection uses model-predictive path integral control (MPPI) over learned dynamics, with planning driven by expected free energy.

I co-lead development of AXIOM at VERSES AI together with Tim Verbelen, Christopher Buckley, Toon Van de Maele, and Alexander Tschantz. Read the paper or check out the code on GitHub.