Follow-the-regularized-leader framework has shown effectiveness and flexibility in online learning problems, where the choice of learning rates are known to be crucial. Recently, adaptive learning rates defined in terms of the arm-selection probabilities, obtained by solving convex optimization, have achieved improved best-of-both-worlds (BOBW) guarantees in various bandit problems. In contrast, BOBW guarantees for its computationally efficient alternative, follow-the-perturbed-leader (FTPL), remain relatively limited since its optimization-free nature ironically makes the design of adaptive, probability-dependent learning rates non-trivial. To address this challenge, we propose an adaptive learning rate for FTPL by introducing surrogate probability functions that can be computed only from the available quantities, without requiring the exact probabilities. Based on these learning rates with surrogate functions, we provide the BOBW guarantee for FTPL with Pareto perturbations for any shape parameter $\alpha >1$, generalizing prior results restricted to specific choices of $\alpha=2$. We further show the BOBW guarantees for FTPL with adaptive learning rates in the bandit problem with expert advices. Our approach preserves the computational simplicity of FTPL while enabling probability-dependent adaptivity, and the surrogate-based methodology may be of independent interest in other algorithmic frameworks beyond FTPL and learning rate designs.

Consistency models imitate the multi-step sampling of score-based diffusion in a single forward pass of a neural network. They can be learned in two ways: consistency distillation and consistency training. The former relies on the true velocity field of the corresponding differential equation, approximated by a pre-trained neural network. In contrast, the latter uses a single-sample Monte Carlo estimate of this velocity field. The related estimation error induces a discrepancy between consistency distillation and training that, we show, still holds in the continuous-time limit. To alleviate this issue, we propose a novel flow that transports noisy data towards their corresponding outputs derived from a consistency model. We prove that this flow reduces the previously identified discrepancy and the noise-data transport cost. Consequently, our method not only accelerates consistency training convergence but also enhances its overall performance. The code is available at https://github.com/thibautissenhuth/consistency_GC.

Precise localization in map-changing environments is a major challenge faced in the case of autonomous vehicles. In such environments, map-matching-based localization can result in an incorrect vehicle pose because the High-Definition map (HD map) and the sensor measurements from the real environments are different. In order to solve this problem, this paper proposes robust localization in map-changing environments based on a hierarchical approach of sliding window optimization and filtering. The changing environments are explicitly modeled by a sub-map from the sliding-window-based graph optimization, and the generated sub-map is used for the map matching of the real-time filter. Since the optimized sub-map includes sensor measurement from both the changed and unchanged regions, it reduces the proportion of the changed region in the total matching region, thereby increasing the robustness of the map matching. The proposed hierarchical localization algorithm is verified and evaluated via simulation and experiments. The results show that the proposed algorithm provides a robust vehicle pose in map-changing environments.

An automated parking system allows drivers to park cars free of stress and collisions in parking lots. Most commercial systems use ultrasonic sensors and require a scanning process to search for a vacant parking space between two adjacent vehicles. If automated parking fails, drivers must return to the initial position for further scanning or take the steering wheel to complete the remaining parking procedure. To improve such limitations, we propose a re-plannable automated parking system with a standalone around view monitor. The proposed system can constantly reflect several errors and risks of perception, positioning, and control in real-life situations, and then re-generate the parking path to improve the parking precision and avoid any collisions. To realize this automated parking system, we propose a new system architecture and introduce four core modules: environmental perception, precise positioning, parking path planning, and vehicle control. In our experiments, we evaluated the system in narrow parking lots with a perpendicular parking space defined as 2.1 × 5.0 m. The proposed system successfully performed reverse automated parking with a lateral error of 0.049 m, a longitudinal error of 0.249 m, and a heading error of 0.468°.

This paper proposes a multiple exposure images based traffic light recognition method. For traffic light recognition, color segmentation is widely used to detect traffic light signals; however, the color in an image is easily affected by various illuminations and leads to incorrect recognition results. In order to overcome the problem, we propose the multiple exposure technique which enhances the robustness of the color segmentation and recognition accuracy by integrating both low and normal exposure images. The technique solves the color saturation problem and reduces false positives since the low exposure image is exposed for a short time. Based on candidate regions selected from the low exposure image, the status of six three and four bulb traffic lights in a normal image are classified utilizing a support vector machine with a histogram of oriented gradients. Our algorithm was finally evaluated in various urban scenarios and the results show that the proposed method works robustly for outdoor environments.