How can smart territories shape the future of automated driving?
With the rapid evolution of automotive technology enhanced by advanced AI, having autonomous vehicles actively present in our life is getting closer to reality. However, there are still several issues to be addressed, in order to ensure that AI-enhanced self-driving cars could eventually handle all aspects of driving safely. One of the most complicated scenarios that automated cars have to face is successfully driving through complex traffic situations, such as not-signalised intersections, roundabouts or mixed traffic zones. In fact, safely crossing these unpredictable traffic situations requires complex decisions for which AI-enhanced automated cars would need to be trained very cautiously. Avoiding these complex traffic situations altogether for AI-enhanced self-driving vehicles would be an inefficient solution, since it would increase pollution and transportation cost, a contradictory concept to the purpose of automated driving.
EURECOM’s Professor Jérôme Härri, expert in connected cooperative automated mobility (CCAM), explains how smart territories could play a crucial role in addressing the automated vehicles safe circulation through realistic environments and complex traffic situations such as roundabouts. This work has been in collaboration with his PhD student Duncan Deveaux, who has recently successfully defended his thesis.
Q. What is the state of the art for automated driving cars driving in smart cities?
JH: Automated driving corresponds to autonomous vehicles equipped with technology like AI, sensors, cameras, and exchanging information over advanced communication technologies with other vehicles and road infrastructures. This technology enables vehicles to drive autonomously without or with minor human interventions. Currently, looking at automated driving, we observe that there is a big mismatch between vehicles, that are becoming smarter in autonomous driving and the road infrastructure, which is not designed to support the AI embedded in autonomous vehicles. We can see that by observing that SAE level 5 (full autonomy) vehicles are already capable of self-driving in well-defined and protected zones, whereas SAE level 3 (partial autonomy) are currently expected to drive in more general road infrastructures (for comparison, the Tesla ‘autopilot’ is considered SAE level 2). Reaching full automation is not going to be achieved by the sole efforts of the automotive industry and smart territories are expected to take a leading role.
Road infrastructure should not be limited to providing signalling, but should also be equipped with AI technologies supporting autonomous driving. Considering a roundabout for example, an autonomous vehicle needs to decide on priority by following a yield sign. Its decision is based on its assessment if inbound vehicles are going to exit the roundabout or not, which is a complex task even for a human, and which is specific to each roundabout. It is therefore an opportunity for smart territories, which built these roundabouts and have knowledge of traffic conditions to provide assistance with adequate and precise knowledge to automated vehicle on the best strategy to drive safely through such complex traffic situation.
Q. What are your collaborations and partnerships on this topic?
JH: We have a partnership with TOYOTA InfoTech Labs, USA for the concept of knowledge networking between vehicles and road infrastructures in vehicular networks. Also, this work is part of an EU project IntellIoT, one of which objective is to build a decentralised secured trustworthy IoT framework for developing knowledge networking as a service and potentially paving the way towards a decentralised knowledge marketplace.
There is an entire ecosystem related to both partnerships, spanning from automotive, manufacturing, farming or even health industries. However, we believe that smart territories should also be integrated into this ecosystem. Indeed, our territory-centric approach aims to remove the pressure from the automotive industry to provide autonomous cars that are equipped a priori with full AI knowledge about driving policies, in all potential complex traffic situations. Instead, smart territories’ road infrastructures would play a leading role in providing, potentially selling, such knowledge and orchestrating the AI learning process according to the specific contexts and data owned by smart territories. An aspect to keep in mind is that connected cooperative automated mobility (CCAM) is not a public service and providing the required knowledge for safely and efficiently driving autonomously may be part of a promising smart territories’ led marketplace.
Q. What is the contribution of EURECOM in these projects and what are the challenges you have yet to overcome ?
JH. We strongly believe that the involvement of smart territories infrastructure with guidelines for safe and efficient driving would be crucial for automated vehicles, particularly for unpredictable and episodic scenarios. Here, AI meets smart territories, where cars would receive knowledge created by smart territories providing instructions on how to safely and efficiently drive through complex traffic situations.
In our study, we take the particular case of roundabout representing a complex traffic situation. We define risk reasoning in roundabout in the form of exit probability as well as time to collision, and developed AI models capable of assessing such risk. We then analyse the characteristics of such AI models, to assess if they are specific to each roundabout or if various roundabouts with similar contexts (e.g. number of exits, size, traffic density, number of lanes) could use the same AI models. More important, we propose a semantic, uniquely-defining roundabout risk-reasoning AI models and their applicability context, such that models can be referenced, queried, shared and finally used independently by different actors. We observed that there is a significant need for standardisation in this domain, where the lack of globally agreed ontologies defining what a model is and does under which conditions forbids efficient knowledge networking.
Accordingly, EURECOM collaborated with Toyota ITC to propose a framework [3] to create, name, store and share knowledge between vehicles and smart road infrastructures.
In more technical details, EURECOM’s contributions are summarised as:
1. Defining an AI model capable of quantifying roundabout risks and hazards
2. Clustering roundabouts based on their common features, characterising them in terms of similarities in AI risk assessment
3. Extracting the semantics out of the features similarities in roundabouts, and defining them as parameters uniquely identifying a risk reasoning AI model and its applicability context
4. Implementing a prototype in EURECOM’s IoT platform with all the learning, storing and sharing mechanisms, using a vehicular/driving simulators capable of precisely modelling various roundabouts with vehicles equipped with controllable sensors.
Another point is that EURECOM did not only focus on how to identify and disseminate knowledge created by different stakeholders, but also developed an orchestration service framework for decentralised knowledge creation, where each vehicle could collaboratively participate to knowledge creation. Decentralised AI mechanisms such as Federated Machine Learning are promising technologies to benefit for the wisdom of the crowd to train AI models, but their efficiency strongly depends on the dynamic selection of the most appropriate ‘crowd’. Benefiting from the defined AI semantics for roundabouts, the proposed framework allows to identify which vehicles could have the most efficient roundabout environment to train a particular AI model.
However, a serious limitation yet to overcome is the system’s security. It is important to make sure that smart territories are attack proof, so nobody could potentially inject fake knowledge in the system. So the next step, is to secure knowledge and trace where it came from. To address this issue, we are now starting to look Blockchain-based Distributed Ledger Technologies, which may log every knowledge-related transaction made by any vehicle. However, adapting such technologies to vehicular networks is not straightforward.
Q. What do you think is the future direction of this work ?
JH: A knowledge marketplace. Once again, I think that we should not rely only in automotive industry to train autonomous cars to learn how to drive in all conditions. It is a very important aspect but it should not be the only one available. We need to engage territories to deploy models of risk assessment in their road networks and make such knowledge available to automated driving cars. Large AI stakeholders as well as the automotive industries will probably own knowledge related to autonomous driving. Whoever owns knowledge controls it eventually. We foresee unconstrained access to knowledge in future networks bearing similar societal challenges, as it was the case for information in the pre-Internet era.
In this context, we would like to provide an ecosystem that motivates the territories to take an active role in knowledge creation and sharing for connected automated mobility. After all, they already hold a key position, owning most of the required data and accordingly should not be kept away from this promising future revolution. Benefiting from next generation IoT networks integrating innovations such as private 5G, decentralised AI and distributed ledger technologies, smart territories could create knowledge related to road infrastructure potentially integrating it in a marketplace available to smart vehicles. In turn, the automotive industry could rely on smart territories to provide complementary knowledge supporting more efficient driving in specific contexts. It is a mutual benefit: smart territories could profit from innovative sources of income by providing high quality knowledge to a knowledge marketplace, whereas smart vehicles could have unconstrained access to various sources of applicable knowledge required for autonomous driving. For example, as a first step, we could imagine smart territories using traffic data from their large-scale traffic or security surveillance system, or equipping city vehicles and signalised intersections with AI and private 5G capabilities to generalise knowledge creation as we did on a reduced set of roundabouts.
by Dora Matzakou for EURECOM
References
2. D. Deveaux, J. Härri, “On the Traffic Impact of Risk-Aware Autonomous Vehicles Routing Strategies”, the 3rd Symposium on Management of Future Motorway and Urban Traffic Systems (MFTS 2020), Luxembourg, July 2020
3. D. Deveaux, T. Higuchi, S. Uçar, J. Härri and O. Altintas, “A definition and framework for vehicular knowledge networking”, IEEE Vehicular Technology Magazine, Volume 16, Issue 2, June 2021.
