Robots need an accurate 3D understanding of their direct vicinity to navigate freely. This two-fold talk will first focus on how to fuse sparse 3D data + dense RGB data with CNNs and improve vision tasks such as semantic segmentation, data completion, object detection, etc. Different from mediated perception, the second part will present an end-to-end driving algorithm which is trained in a deep reinforcement learning fashion in a realistic synthetic environment.

The seminar will be presented at 2:00 p.m. in room PLT-3370.