LIDAR-AD Enhances Autonomous Driving with Latent World Models

Autonomous driving demands sophisticated decision-making capabilities over extended time horizons within complex, dynamic traffic environments. Latent world models offer a promising framework by enabling imagination-based planning in compact, abstract representations. However, traditional approaches often struggle with redundant multi-source observations and suboptimal absolute action modeling, which hinder the learning of decision-relevant latent dynamics. To address these challenges, researchers developed LIDAR-AD (Latent-Interaction Dreamer with Action-Residual Chains for Autonomous Driving). This innovative model replaces conventional observation reconstruction with a redundancy-reduced latent alignment strategy, which encourages the formation of compact representations focused on risk-relevant relationships within multi-source driving inputs. Furthermore, LIDAR-AD models vehicle control as residual action updates and employs residual-action sequence contrastive learning to align multi-step, residual-driven rollouts with future latent states. A deterministic analysis confirms that the latent-tanh residual parameterization maintains interior action reachability while effectively representing smooth, long-horizon control through compact local updates. These combined design choices significantly enhance risk-aware state abstraction, continuous-control modeling, and long-horizon dynamics prediction. Extensive experiments across diverse simulated driving scenarios demonstrate LIDAR-AD's superior performance, achieving higher rewards and success rates compared to other learning-based methods and world-model baselines. Its transferability was also validated on nuPlan-derived log-reconstructed scenarios, showcasing its potential in real-world traffic layouts.