We also tested CPM's usefulness as a world model for a model-based RL agent. The agent's task was to move an object to a target location. Our CPM-based agent (red) broadly outperformed baselines (especially in the challenging "Unobserved" setting), achieving higher mean rewards.

We tested our model in a simple physics environment with 'Observed' & 'Unobserved' settings. The plots show CPM (red) has higher prediction accuracy (H@1) than GNN & Modular (separate transition MLP per slot) baselines. The performance gap widens over longer prediction horizons.

How does the CPM build its causal graph? We treat causal discovery as a multi-agent RL problem. As shown in the Causal MDP, controller agents make sequential decisions to add edges to the graph, determining which objects interact.

Our model (see diagram) has an object-centric vision encoder to instantiate object representations and an action encoder, for force representations. The core part of the architecture is CPM. It acts as a dynamic transition function using a causal graph to predict object dynamics.