| Tartalmi leírás: |
Autonomous mobile robots increasingly rely on learning-based methods instead of fixed, hand-crafted control strategies. In real deployments, robots often communicate over mobile networks (e.g., 5G), where bandwidth, latency, and packet loss vary over time and can disrupt both training and operation. At the same time, robotic sensor data such as camera images or lidar point clouds are often sensitive, so sending raw data to a central server is not desirable. Federated learning addresses this by allowing robots to share model updates while keeping data local. This thesis studies how to make reinforcement learning training more robust to network fluctuations. A simulation environment is used to model both robot learning and network behavior under different conditions. Each robot is trained locally with reinforcement learning, while a federated server aggregates the resulting model updates into a global model that is distributed back to the clients. The system is evaluated using metrics for convergence, task performance, and robustness under varying bandwidth. The aim is to assess how well this architecture maintains stable learning and reliable behavior in realistic network scenarios without transferring sensitive raw sensor data.
|
