Concept-Guided Spatial Regularization for World Models in Atari Pong
Artificial IntelligenceMachine Learning
Summary
The authors studied five world models used for playing Atari Pong, focusing on the models by themselves rather than within full reinforcement learning systems. They found that when these models were frozen and tested alone, their predictions often had errors like the ball disappearing or moving incorrectly. Training new policies inside the frozen models resulted in much worse game performance compared to original training. To improve this, the authors introduced a method called Concept-Guided Spatial Regularization (CGSReg) that helps the model better understand important parts of the game screen, like the ball. Applying this method improved some of the models but did not solve all issues.
Authors
Yukuan Lu, Zaishuo Xia, Weyl Lu, Yubei Chen
Abstract
World models are usually evaluated as components of model-based reinforcement learning (MBRL) systems, while the world models themselves are rarely studied in isolation. We examine five representative visual world-model agents in Atari Pong: DreamerV3, DIAMOND, TWISTER, Simulus, and STORM. After reproducing their training pipelines and matching the reported agent performance, we freeze the learned world models and evaluate them with a closed-loop rollout diagnostic: a policy trained separately from the corresponding MBRL agent interacts with each frozen model, and the generated video trajectories are inspected for visual and dynamical errors. Across all five models, the rollouts contain clear failures, including ball disappearance, incorrect ball motion, and invalid ball-paddle interactions. Beyond visual trajectories, we further evaluate them with pixel-space zero-shot MBRL, where a new policy is trained entirely inside a frozen world model and then evaluated in the real environment. Across all five models, the resulting policies substantially underperform those produced by the corresponding original MBRL training pipelines. The gap is particularly large for DreamerV3, whose mean return drops from -5.5 to -20.9, near the minimum Pong return of -21. We hypothesize that insufficient modeling of task-critical concepts, such as the ball in Pong, may contribute to these failures. We therefore propose Concept-Guided Spatial Regularization (CGSReg), an auxiliary pixel reconstruction loss applied to segmented concept regions. Experiments show that CGSReg improves both closed-loop rollouts and pixel-space zero-shot MBRL in DreamerV3, DIAMOND, and TWISTER. Its effects vary across the remaining models and evaluation metrics, indicating that CGSReg alone does not address all world-model bottlenecks.