Discussion on Explainable AI for Robotic Applications

The current AI approaches based on Deep Learning were originally developed for fast data queries in large datasets for search engines, social media and advertising. The common property of these fields is that they are not used in critical decision loops (control) of a robotic system, but they serve as an index key to find previously searched information that is similar to the current situation.  This origin resulted in a strong development of the data labeling direction that is essential for fast data association. In my talk, I want to discuss the necessary extensions that need to be added to the current AI approaches to make them applicable for decisions on robotic systems. While the approaches become increasingly Better in answering the "what is there?" question, a robotic system requires in addition also information about the "confidence" of each query. A 95% accurate system running for 24 hours fails during 72min/day. It is essential for the control system to identify these periods to prevent damages to the system and the surrounding environment. Additionally, usually not a single sensor is used for control and for a robust data-fusion a (metric) error covariance is important. I show ways how to achieve this goal in DL context. The last step is a discussion of temporal extensions of the current AI approaches, which need to understand not only the current snapshot of the scene but its temporal evolution to grasp the current context and model dynamic events. I will present our initial work on temporal scene modeling and discuss the necessary updates to the benchmarking in current AI to make it applicable to robotics.

Author(s): Darius Burschka

Abstract | Full-Text | PDF

Share This Article