Joint Workshop with SAFER
27/02/2018
Dreams4Cars joined forces with SAFER - Vehicle and Traffic Safety Centre at Chalmers. SAFER is also the open innovation arena where partners from the society, the academy and the industry can meet and share research and knowledge within safe mobility – a multi-disciplinary research hub that enables progress for its partners and for the society.
Dreams4Cars held a workshop around cognitive systems in autonomous vehicles on 26th January 2018 in Gothenborg.
The Agenda
08:15 - 08:30 | Giulio Piccinini Chalmers / Safer |
Introduction about QUADRAE Project |
08:30 - 09:00 | Jonas Bärgman Chalmers |
Predictive processing framework for modelling driver's behavior in automated driving |
09:00 - 09:30 | Giulio Piccinini & Esko Lehtonen Chalmers |
Preliminary results about modelling driver's behavior as a function of kinematic criticality and automation level |
09:30 - 10:00 | Coffee | |
10:00 - 10:30 | Fredrik Granum Volvo Cars |
Driver models in safety benefit simulations |
10:30 - 12:00 | Mauro da Lio University of Trento |
Artificial Cognitive System Architecture for long-term reliable automated driving |
12:00- 13:00 | Lunch | |
13:00 - 13:30 | Henrik Svensson University of Skövde |
Active discovery of threatening situations by dream-like simulations |
13:30 - 14:00 | Franceso Biral Univesity of Trento |
Behaviour optimisation and generation of training examples with offline Optimal Control |
14:00 - 14:30 | Rafael Math DFKI |
OpenDS environment for simulations and dreaming |
14:30 - 15:30 | Open discussion |
Abstracts
Mauro Da Lio, University of Trento
Artificial Cognitive System Architectures for long-term reliable automated driving
Automated driving will need unprecedented levels of autonomy, reliability and safety for market deployment. The average human-driver fatal accident rate is approximately 1 (fatality) every 100 million miles in the US and the EU. Hence, automated vehicles will have to provably, and significantly, best these figures. Unfortunately, according to many reports on the disengagements of prototype automated driving systems, today we are far from demonstrating these levels and, even worse, a large amount of resources appear to be still necessary towards achieving this objective.
This talk presents the position of the partners of the H2020 Dreams4Cars research project (a Research and Innovation Action funded under the EU Robotics banner) regarding the architecture, and consequently the abilities, of agents capable of long-term reliable driving.
Today almost all approaches to automated driving implicitly assume the sense-think-act paradigm (aka perception-decision-action). Several examples of such current approaches will be examined. We believe that the agent architecture, implicitly assumed, besides being very hungry in terms of resources for development, is also inadequate to achieve robust autonomy for the same reasons it was unsuccessful in other similar robotic applications. We review several critical aspects of this architecture, from scalability to maintainability and validation; including some considerations regarding recent examples of Deep Neural Network implementations that still retain many of the weaknesses of the paradigm.
Then, I will introduce and motivate a (biologically inspired) layered control architecture that, we believe, can scale much better to deal with the complexity of the real world. I will also describe a learning mechanism similar to human dreams, in which the agent itself can anticipate potential threats and prepare to act in threatening situations before they are even met. This way the agent may become expert (like senior drivers are compared to young drivers) by learning from discovered potential threats. Besides increased autonomy and robustness, this approach looks to be more economical in terms of resources needed for development.
The talk will give many details about implementation of the agent, which uses Deep Neural Networks as building block but writhing a network of networks that reproached the main functionalist of the human brain, including simultaneous affordance generation, episodic simulations, robust adaptive action-selection, sensory anticipation. I will also mention how the similarity of the function implementation may allow to trigger a “mirroring” mechanism with the humans, which makes the agent capable of “understanding” human intentions and being naturally understood by humans (this will be contrasted with the known causes of accidents occurred so far in AD testing).
In the end, recent developments in the Dreams4Cars project will be presented. In particular how to engineer artificial drivers that can learn by rehearsing their own experiences, in a way that is very similar to human dreams.
Henrik Svensson, University of Skövde
Active discovery of threatening situations by dream-like simulations
This talk outlines some of the aspects of the dream-like simulations to be developed within the H2020 Dreams4Cars research project (a Research and Innovation Action funded under the EU Robotics banner).
In this project one aim is to create dream-like simulations, which consists of reconstructions and recombinations of the cars previous experiences into novel situations from which the car gains new knowledge. While dreams in this project are not an exact equivalent of the dreams of biological agents, there are some crucial aspects of “biological dreams” that will be drawn upon in this project: (1) dreams and other kinds of mental simulations are off-line, i.e., they are active but are not interacting with the controlled entity, the extra-neural body in biological agents, in particular (2) dreams reactivate the control system as if it were interacting with the controlled entity. (3) Dreams enable the biological agent to think about previous and future situations to (4) increase its ability to handle situations in the waking state (Svensson, Thill & Ziemke, 2013; Svensson & Thill, 2016).
The talk presents some of the background on biological dreaming and simulation as well as outlines the type of “dreams” that will be used in the project and some more specific details on how we aim to implement the “dreams”.
Francesco Biral, University of Trento
Behaviour optimisation and generation of training examples with offline Optimal Control
The majority of modern and state of the art approaches for the development of automated driving heavily rely on artificial intelligence (AI) and on the use of Deep Neural Networks. The AI has to be trained and tested on a large amount of driving data to achieve the reliability and safety level necessary to drive 100 millions mile without a fatal accident.The H2020 Dreams4Cars research project aims at solving this problem using a dream mechanism to teach the agent to become an expert driver by learning from discovered potential critical situations.
This talk presents the approach used in Dreams4Cars to train the agent in the dream state via the use of optimal control techniques to model different driving behaviors. According to experimental evidence, the optimality of human sensorimotor control is constantly assumed in this project and it is used to implement the inverse models (i.e. affordances) that map sensory data with controls.
The layered control architecture of Dreams4Cars, which is biologically inspired by the human brain, foresees the use of three data streams. In particular, the first of stream (dorsal stream) creates an artificial “motor cortex” map, which is a two-dimensional map of the control space (i.e. lateral and longitudinal controls) from perceived sensory data. The value that is encoded in the motor cortex map is called salience (s) of the trajectory/trajectories. The salience originates at each couple of lateral and longitudinal control and expresses “how good” the particular choice of controls is. Points (i.e. controls) that are close in the motor cortex correspond to trajectories aiming at similar, but slightly different, directions in the space. Therefore, different regions in the cortex map correspond to different affordances (e.g., lane change, car follow etc.) which are produced with different (families of) trajectories. The height of each hump in the motor cortex is the urgency of that particular affordance and describes the active regions in the motor cortex.
The talk introduces the formulation of the inverse models as optimal control problems and the numerical approach used in Dreams4Cars to find the optimal solution that generates the optimized behavior in term of motor/salience map and affordances.
The talk also discusses computational challenges along with the approach adopted to incorporate the driving style and the characteristics dynamic response of vehicle to driver’s inputs.
In the end, recent examples from the Dreams4Cars project will be shown.
Raffael Math, DFKI
An Open-Source Driving Simulator for Automated Driving
This talk introduces OpenDS, a free driving simulator, which has first been released under open-source license in 2013. As full-fledged driving simulation software for the evaluation of automotive applications is high in price and low cost simulators often lack of extensibility, OpenDS was initiated to provide a basic simulation toolkit to the researcher community and counts today more than 1500 users from both academia and industry.
OpenDS is implemented in Java and based on the jMonkeyEngine, a high performance scene graph based graphics API which uses Bullet for physics simulation. Due to a high number of extensions and a variety of pre-defined driving tasks, many different scenarios can be simulated out-of-the-box or be created with little effort. Moreover, OpenDS is ready to connect to various hardware (eye tracker, CAN bus, motion seat, steering wheel, Oculus Rift) as well as software (traffic light simulation, multi-driver simulation, data and multimedia provider) and supports multi-screen output for surround projection.
The latest development is driven by the H2020 Dreams4Cars research project, where the simulation software will be used to train an artificial driver in a safe environment. By the means of simulation, almost any critical situation which might never appear under real conditions – even after millions of kilometers of driving – can be created, modified, and simulated multiple times.
The talk also gives insight in the recent simulation developments of the project, especially into the “dreaming mechanism” where critical scenarios will be generated automatically for subsequent simulation, the integration of an interface to operate the virtual vehicle by the artificial driver, and the integration of Chrono, a more realistic multibody physics engine.