Abstract: In this talk, we present the latest research work from the Toyota Research Institute on robotic assistance in the home. Motivated by the growing problem of an ageing society, a mobile manipulator robot is demonstrated that can be easily taught by human demonstration in virtual reality to achieve complex tasks. We show our latest results of real tasks executed in various homes around the Bay Area and discuss the future of where Toyota is headed with this research.
Algorithms for Verifying Deep Neural Networks and Their Applications
Abstract: Deep neural networks are widely used for nonlinear function approximation with wide range of applications in robotics. Although these networks involve the composition of simple arithmetic operations, it can be very challenging to verify whether a particular network satisfies certain input-output properties for safety and robustness. In this talk, we will first provide an overview of existing methods for soundly verifying such properties. These methods borrow insights from reachability analysis, optimization, and search. Then we will discuss ways to scale existing methods and apply them on real applications. In particular, we will discuss an idea of composing different algorithms together to achieve greater applicability. And then we will talk about how to perform efficient sound runtime verification on problems that require real time adaptation. We will conclude the talk with a discussion of open problems and future directions.
The survey paper “Algorithm for Verifying Deep Neural Networks” can be found at https://arxiv.org/pdf/1903.06758.pdf. The NeuralVerification.jl toolbox can be found in https://github.com/sisl/NeuralVerification.jl.
Speaker Bio: Dr. Changliu Liu is an assistant professor in the Robotics Institute, School of Computer Science, Carnegie Mellon University (CMU), where she leads the Intelligent Control Lab. Prior to joining CMU, Dr. Liu was a postdoc at Stanford Intelligent Systems Laboratory. She received her Ph.D. from University of California at Berkeley and her bachelor degree from Tsinghua University. Her research interests lie in the design and verification of intelligent systems with applications to transportation and manufacturing. She published the book “Designing robot behavior in human-robot interactions” with CRC Press in 2019. She received Amazon Research Award and Ford URP Award.
Abstract: Automated systems especially automated driving promises the revolution of mobility, improved traffic efficiency, and safety. To guarantee a fault free automated system is one of today's most challenging engineering tasks. This lecture introduces scenario-based assessment on a system-level to derive a safety case for automated systems using the example of automated driving. Automated driving is being designed to work in a shared traffic space with human drivers introducing a high uncertainty environment the system is acting in. The lecture discusses the methodology behind system-level testing including different test domains for example simulation, agent modeling, perception modeling, performance indicators, and clustering of scenarios.
Speaker Bio: Dr. Thomas Kuehbeck is working as a leading architect at the Toyota Research Institute for automated vehicles of the future. Prior to joining TRI, Dr. Kuehbeck had various roles within the BMW Group designing the very first automated driving stack, building up the simulation efforts, deriving the verification and validation methods for L3/L4/L5 systems. The main work of Dr. Kuehbeck was integrated into PEGASUS a German research project paving the way to save autonomy. He received the Diploma Degree from the University of Applied Sciences Regensburg, the Bachelor of Honors in computing science from Staffordshire University, and his PhD in computing science from Staffordshire University. His research is focusing around verification and validation of automated systems as well as learning from fleet data.