您查看的活动已结束

Topic1: Tackling Safe and Efficient Multi-Agent Reinforcement Learning via Dynamic Shielding

Speaker：Wenli Xiao

Abstract: Multi-Agent Reinforcement Learning (MARL) has been increasingly used in safety-critical applications but has no safety guarantees, especially during training. In this paper, we propose dynamic shielding, a novel decentralized MARL framework to ensure safety in both training and deployment phases. Our framework leverages Shield, a reactive system running in parallel with the reinforcement learning algorithm to monitor and correct agents’ behavior. In our algorithm, shields dynamically split and merge according to the environment state in order to maintain decentralization and avoid conservative behaviors while enjoying formal safety guarantees. We demonstrate the effectiveness of MARL with dynamic shielding in the mobile navigation scenario.

Topic2: Less is more: A Robust Visual Inertial Odometry with Active Feature Extraction

Speaker: Muhan Lin

Abstract: Abstract— To achieve robust performance, it is common for visual odometry and SLAM to track more features, like several hundreds of points in real time. Although this strategy performs well on high-end desktop PCs, it is difficult to apply it to some mobile platforms with limited computation resources, such as VR, Micro UAV, and multi-camera systems. Additionally, noisy visual feature points may decrease the accuracy of visual odometry and SLAM. Therefore, fewer but more informative features can boost efficiency and accuracy compared to extracting more features. It means that less is more. To achieve this target, we propose a new criterion for the active feature selection based on singular values and then incorporate this method into an advanced VIO system, TP-TIO [1]. With the new system, using half of the features required by the original TP-TIO, the residuals can be reduced to 56.23% of the ones generated by the original TP-TIO without increasing the processing time to a large degree. The new system was verified with the mmpug datasets [2], which were extracted in a long and dark corridor.