Causal Object-Centric Models for Planning with Monte Carlo Tree Search
Researchers have introduced COMET, a novel model-based reinforcement learning algorithm designed for planning. COMET utilizes Monte Carlo Tree Search within a slot-structured latent space, pairing a frozen unsupervised object-centric encoder with a transformer-based world model. It incorporates a unique action-slot fusion mechanism and object-causal attention to focus decision-making on relevant entities. In early training stages across diverse tasks, COMET demonstrated superior performance compared to existing object-centric and monolithic baselines. AI
IMPACT This research introduces a new method for AI planning that improves early-stage training performance by focusing on object-centric reasoning.