Preference-Augmented Distributional Reinforcement Learning for Risk-Sensitive Navigation

원문 초록

Ensuring safe and reliable navigation under diverse sources of uncertainty remains a central challenge for autonomous systems. Within distributional reinforcement learning (DistRL), Implicit Quantile Networks (IQN) have attracted interest for navigation under uncertainty because they enable risk adaptation through flexible quantile sampling. Existing approaches, however, typically use a fixed or randomly sampled Conditional Value-at-Risk (CVaR) level during training and adjust the risk level during deployment using hand-tuned hyperparameters or external sensor signals. This scalar-CVaR formulation confines learning to a narrow family of tail-risk preferences, creating a mismatch between training-time preference exposure and deployment-time risk adjustment. We instead represent variations in risk preference as shifts in quantile-weight distributions and expose the agent to diverse preferences through preference-guided exploration. At deployment, we optimize against the worst-case preference distribution to derive a conservative policy that accounts for multiple risk profiles. This reduces hyperparameter tuning and reliance on limited sensor signals under partial observability. Experiments across simulated environments and vehicle platforms show that our method reduces collisions while improving task success.