Reinforcement learning (RL) offers a principled framework for learning driving policies from interaction with the environment. However, its application to urban autonomous
driving is hindered by long-horizon credit assignment, safety constraints, and ...
Reinforcement learning (RL) offers a principled framework for learning driving policies from interaction with the environment. However, its application to urban autonomous
driving is hindered by long-horizon credit assignment, safety constraints, and limited opportunities for online exploration. Hierarchical RL mitigates these issues by decomposing decision-making into high-level planning and low-level control. Existing approaches often rely on variational autoencoders (VAEs), whose latent skills are optimized for reconstruction rather than for capturing the temporal structure of driving. As a result, the learned skills may not align with meaningful maneuvers or support coherent long-term planning.
This work proposes a hierarchical RL framework for urban autonomous driving built on Hilbert representations rather than VAE-based latent skills. First, a Hilbert encoder is trained on offline driving data so that Euclidean distances in the latent space approximate temporal distances between states. On top of this temporally grounded representation, a low-level skill policy is learned using intrinsic rewards that measure directional progress in the Hilbert space, which allows skills to be interpreted as latent directions. Finally, a high-level policy operates over these skills at a coarser time scale, selecting skill directions to maximize the original environment reward.
The proposed method is trained purely offline using the LMDrive dataset and evaluated in the CARLA simulator. Comparative experiments against various baselines indicate that the Hilbert-based hierarchy improves long-horizon decision-making, achieving more reliable urban driving behavior and better overall performance.