Veranstaltungsprogramm

This work presents a hybrid planning approach for autonomous car racing that integrates pre-trained Soft Actor-Critic (SAC) agents into a local iterative planning and control framework. Unlike end-to-end control, the proposed architecture embeds SAC agents with the environment simulation in a prediction-based loop to generate reference trajectories for the ego vehicle. These trajectories are then executed by a robust motion controller, which compensates for actuation deadtime. Training is organized through a three-stage curriculum: fast driving without collisions, overtaking stationary cars, and overtaking moving opponents. This progressive strategy enables safe and adaptive skill acquisition in competitive scenarios. The method is evaluated on a 1:76 scale racing setup, Mini-Auto-Drive MAD76, with up to four cars, where it is tested for lap time and its ability to perform efficient overtaking maneuvers. Safety analysis reveals a crash-free reliability of 98.74% across more than 60,000 maneuvers. Results demonstrate that the hybrid design combines the adaptability of reinforcement learning with the robustness of classical control, making it suitable not only for autonomous car racing but also for broader multi-agent robotics tasks such as warehouse navigation. It is indented to implement the system on an embedded platform by model-driven software engineering, to verify its real-time capabilities.

With rising product variety, shorter life cycles, and small-batch demands, companies—especially in high-wage countries—face significant economic challenges. The Smart Factory Grid (SFG) vision addresses these by decoupling production facilities from rigid material flow systems, enabling dynamic, service-based manufacturing. Within the German Research Foundation’s “Smart Factory Grids” initiative, Esslingen University investigates adaptive control of material flows using Unmanned Aerial Vehicles (UAV). These complex mechatronic systems, providing high agility and responsiveness compared to traditional ground-based intralogistics.

The paper proposes a three-subsystem solution for multi-criteria process optimization (balancing cost, time, and energy) and UAV-based transport. The "Order Assignment" subsystem decomposes production orders into steps and determines optimal resource allocation using evolutionary algorithms and AI methods like reinforcement learning. The "Route Allocation" subsystem manages UAV navigation by dividing airspace into layers for conflict-free flight planning. Finally, the "Path Planning" subsystem ensures efficient, safe, and adaptive flight routes using a combination of snap-based optimization and model predictive control, with digital twins providing real-time system feedback.

The approach highlights UAVs as a promising supplement to conventional logistics, enabling resilient, efficient, and highly flexible material flows in future smart factories.