Post-Doc

Connected and Autonomous Vehicles (CAVs) take advantage of the advancement of

communication and sensing technologies to offer a potential sustainable alternative to current

mobility services. Many ongoing projects are studying the effects of CAVs on the network, at

the same time trying to identify the best strategies to develop in order to design new and

dedicated traffic control strategies. The challenge here is to solve the medium-term situation

where both conventional and automated traffic will share the road network.

Very first real-world deployments have been carried out experimentally and tend to confirm

the expected benefits of cooperation between AVs (Stern et al., 2018), that have been

previously identified in simulation (Guériau et al., 2016). However, the high cost of running

such full scale tests (with relatively small amount of vehicles) is one of the reason CAV

technologies deployment takes more time than expected. Indeed, autonomous driving

algorithms, especially when built using Artificial Intelligence techniques and learning-based

approaches, require substantial quantities of data and experience that cannot be solely

provided by actual field-tests.

The deep-learning community is being more and more aware that machine learning models

are “costly to train and develop, both financially, due to the cost of hardware and electricity or

cloud compute time, and environmentally, due to the carbon footprint required to fuel modern

tensorprocessing hardware”1. This is led to the emergence of an interesting and novel research

effort, called Green AI in the litterature (Schwartz et al.,2019), that aims at reducing the amount

of data required for deep-learning based approaches to converge. One the the strategies

“which has been recently gaining importance to drastically reduce computational time and

energy consumed is to exploit the availability of different information sources”1, or

different models, environments.

Simulation and robotic models (such as in Hyldmar et al., 2019) appear to be the fastest

way to first develop, train and then test autonomous driving tasks, allowing the system

designer to investigate the behaviour of embedded navigation systems in a wide range of

situations and/or conditions and across different environments (including for instance

adverse weather conditions that affect vision-based perception, Blin et al. 2020). This process

however faces several challenges: reality gap, when a simulation/model fails to capture all

the particularities of a real system, and domain adaptation, when a model is

developped/trained in a particular context and has to adapt to a different one.

In this context, techniques akin to transfer learning (TL) have been developed to enable

knowledge acquired by one or multiple (Taylor et al., 2019) learner (source) agents to be

transferred to another or the same (target) system, helping the latter to learn a similar but

different task or to adapt an existing algorithm to a similar domain. Recently, TL was shown to

be particularly efficient when transferring autonomous driving tasks, between different

domains (Sharma et al., 2019) and from simulation to a real system (Balaji et al., 2019).

1 https://news.mit.edu/2020/artificial-intelligence-ai-carbon-footprint-0423

The work intended within this Post-doc position is part of a research project named

MultiTrans, that focuses on exploring novel TL approaches for autonomous vehicles scene

semantic segmentation and detection accross 3 different environments (as depicted in the

Figure above: simulation, a robotic platform and a real-world autonomous shuttle test-bed).

This position is funded by the Agence Nationale de la Recherche (ANR) under grant reference

ANR-21-CE23-0032.