International Scientific Journals
of Scientific Technical Union of Mechanical Engineering "Industry 4.0"

World-model reinforcement learning agents plan by imagining future trajectories through a learned latent dynamics model. In safety-critical driving tasks, the quality of these imagined rollouts — particularly their representation of vulnerable road users — directly determines whether the agent can anticipate and avoid dangerous situations. We analyze the imagined rollouts of a DreamerV3 agent trained to navigate urban environments in CARLA using semantic segmentation and depth observations. By decoding the agent’s latent imagination into observation space, we qualitatively and quantitatively examine how pedestrians are represented in imagined futures, comparing episodes that result in collision against those where the agent successfully avoids. We investigate whether the world model accurately predicts pedestrian presence and motion in its imagined horizon, and whether reconstruction fidelity of pedestrian regions correlates with avoidance outcomes. Our analysis provides insight into the internal representations that underlie emergent safety behaviors in model-based driving agents and highlights limitations of finite imagination horizons for pedestrian safety.