We need to predict each and every critical situation on the road and change the course of actions leading to danger. Our cars can help us in this. They can become more intelligent. Only if we let them dream.
See the video below for an introduction to Dreams4cars. The second video is a comprehensive summary of the whole project, comparing the driving agent with the human brain, the composition of the driving agent and how it teaches itself through dreamlike simulations to cover as many driving possibilities as possible in the safe environment of a virtual simulator. The final video compares Dreams4cars with other autonomous driving projects to highlight its self-learning capability.
Dreams4Cars is about "The Driving Agent" - the complex system comprising of software and sensors that aims to take over the control of the automated vehicle. The project aims to enable the driving agent to deal with difficult and rare road circumstances.
The H2020 Dreams4Cars research project (a Research and Innovation Action funded under the EU Robotics banner) deals with the architecture, and the abilities, of agents that should be capable of learning reliable driving and natural human-robot interactions.
Dreams4Cars (D4C) should not be seen as yet another project for developing automated driving. Rather, it is a robotics initiative to develop cognitive abilities that may in turn be used for the development of automated driving. The goals of D4C are 1) automatic discovery of significant situations and 2) automatic learning from those situations. In other words: self-discovery and self-optimisation of behaviour at all levels of sensorimotor control, and including rare events. It also tackles the issue of verifiability of agent learned behaviours (by inspection of the agent’s interpretable motor cortex).
The purpose of the agenda is to give an overview of some of the key features and preliminary results of the project.
PLACE: SAFER LINDHOLMSPIREN 3A, DEMOSTUDION
DATE: June 17th 2019 13.00–17:00
|13:05||Dreams4Cars approach and tangible results||Mauro da Lio|
|13:35||Agent implementation with focus on application||Mauro da Lio|
|14:10||CarMaker and OpenDs test environments – test scenarios ready to try||Ricardo Dona|
|14:30||Discussion / Coffee break|
|15:00||The open source simulation environment OpenDS – free simulation environment||Rafael Math|
|15:20||Dream generation with Open DS||Henrik Svenson|
|15:40||Dream generation with deep neural networks – a look into the future||Alice Plebe|
|16:00||Learning from Episodes: reinforcement learning challenges for autonomous cars||Sara Mahmoud|
|16:20||Learning from episodic simulations: bootstrapping of sensorimotor abilities||Mauro da Lio|