COMET Enhances RL Planning with Causal Object-Centric Models.

Model-based reinforcement learning (RL) algorithms that use Monte Carlo Tree Search (MCTS) are powerful, but their efficiency can be limited in complex environments. This research introduces COMET (Causal Object-centric Model for Efficient Tree search), a novel algorithm that enhances MCTS by operating within a slot-structured latent space, leveraging object-centric representations. COMET integrates a frozen unsupervised object-centric encoder with a transformer-based world model. A key innovation is its action-slot fusion mechanism, which binds actions directly to objects for more accurate slot transition prediction. Furthermore, the policy and value heads employ object-causal attention, modulating token interactions based on learned per-slot relevance scores. This ensures that the decision-making process focuses on the most task-relevant entities. By adding an explicit object-level inductive bias to MuZero-style latent planning, COMET demonstrates superior performance. Across eight visually and dynamically diverse tasks from benchmarks like Object-Centric Visual RL, ManiSkill, Robosuite, and VizDoom, COMET achieved higher mean normalized scores during the early stages of training compared to both object-centric and monolithic baselines.