Standard Reinforcement Learning (RL) algorithms are typically designed to maximize the expected accumulative reward, which may be inadequate in scenarios where risk sensitivity is critical. In this work, the problem of risk-sensitive RL with Iterated Conditional Value at Risk is studied, where the objective is to optimize outcomes under a specified risk level τ at each step. This work provides the first minimax optimal sample complexity analysis for this problem with a generative model. Specifically, the sample complexity is firstly characterized as a function of the number of statesS, actionsA, and effective horizon (1−γ)−1(resp. horizonHin the finite-horizon setting), and is further shown to be minimax optimal via a novel minimax lower bound analysis when the risk level 0Hin the finite horizon setting). For the case when the risk level is small, the limiting case of τ → 0, termed worst-path RL, is then studied, and the minimax optimal sample complexity is also theoretically characterized.

In this work, we study the sample complexity problem of risk-sensitive Reinforcement Learning (RL) with a generative model, where we aim to maximize the Conditional Value at Risk (CVaR) with risk tolerance level $\tau$ at each step, named Iterated CVaR. We develop nearly matching upper and lower bounds on the sample complexity for this problem. Specifically, we first prove that a value iteration-based algorithm, ICVaR-VI, achieves an $\epsilon$-optimal policy with at most $\tilde{\mathcal{O}}\left(\frac{SA}{(1-\gamma)^4\tau^2\epsilon^2}\right)$ samples, where $\gamma$ is the discount factor, and $S, A$ are the sizes of the state and action spaces. Furthermore, if $\tau \geq \gamma$, then the sample complexity can be further improved to $\tilde{\mathcal{O}}\left( \frac{SA}{(1-\gamma)^3\epsilon^2} \right)$. We further show a minimax lower bound of ${\tilde{\mathcal{O}}}\left(\frac{(1-\gamma \tau)SA}{(1-\gamma)^4\tau\epsilon^2}\right)$. For a constant risk level $0<\tau\leq 1$, our upper and lower bounds match with each other, demonstrating the tightness and optimality of our analyses. We also investigate a limiting case with a small risk level $\tau$, called Worst-Path RL, where the objective is to maximize the minimum possible cumulative reward. We develop matching upper and lower bounds of $\tilde{\mathcal{O}}\left(\frac{SA}{p_{\min}}\right)$, where $p_{\min}$ denotes the minimum non-zero reaching probability of the transition kernel.

Drone-based vehicle detection is crucial for intelligent traffic management. However, current methods relying solely on single visible or infrared modalities struggle with precision and robustness, especially in adverse weather conditions. The effective integration of cross-modal information to enhance vehicle detection still poses significant challenges. In this letter, we propose a masked-guided cross-modality fusion method, called MCMF, for robust and accurate visible-infrared vehicle detection. Firstly, we construct a framework consisting of three branches, with two dedicated to the visible and infrared modalities respectively, and another tailored for the fused multi-modal. Secondly, we introduce a Location-Sensitive Masked AutoEncoder (LMAE) for intermediate-level feature fusion. Specifically, our LMAE utilizes masks to cover intermediate-level features of one modality based on the prediction hierarchy of another modality, and then distills cross-modality guidance information through regularization constraints. This strategy, through a self-learning paradigm, effectively preserves the useful information from both modalities while eliminating redundant information from each. Finally, the fused features are input into an uncertainty-based detection head to generate predictions for bounding boxes of vehicles. When evaluated on the DroneVehicle dataset, our MCIF reaches 71.42% w.r..t. mAP, outperforming an established baseline method by 7.42%. Ablation studies further demonstrate the effectiveness of our LMAE for visible-infrared fusion.