Hallucinating robots: Estimating the uncertainty over obstacle distances for safer robotic navigation

Friday, June 29, 2018

2.30 p.m.

ISI seminar room 1st floor

Dr. Francesco Verdoja University of Turin

Estimating the uncertainty of predictions is a crucial ability for robots in unstructured environments. The appearance of mobile robots in human-inhabited environments such as warehouses and industrial plants has highlighted the issue, since misbehaviors can cause expensive damages and also human harm. Most mobile robots for indoor use rely on 2D laser scanners for localization, mapping and navigation. These sensors, however, cannot detect transparent surfaces or measure the full occupancy of complex objects such as tables. We propose the use of a Deep Neural Network tailored to overcome this limitation by learning to estimate the true occupancy of objects. We demonstrate the ability of the network to provide estimates in real time and improve local navigation safety. However, these estimates are subject to noise, making the evaluation of their confidence an important issue.
For this reason we also studied the uncertainty estimation of these deep models, discussing why the widely adopted approach to deep model uncertainty estimation, MC Dropout, cannot be applied in this domain, and proposing an alternative solution based on a fully convolutional autoencoder.
We also present how to build a map using the estimates while taking into account the level of uncertainty in each estimate. We finally show how the constructed map can be used to increase global navigation safety by planning trajectories which avoid areas of high uncertainty.

Francesco Verdoja received his PhD in Computer Science from University of Turin, Italy in July 2017. During his PhD studies his research focused on graph-based image processing, in particular exploring the use of the Graph Fourier Transform in the context of image compression and anomaly detection. He has also conducted research on image and 3D point cloud segmentation, and his work on tumor segmentation awarded him the "Best student award" at the International Computer Vision Summer School 2014 (ICVSS14). Since September 2017 he joined the Intelligent Robotics group at Aalto University in Helsinki, Finland as a postdoctoral researcher. His research going forward will focus on robotic perception with particular interest in inter-sensor learning.