Véronique Cher­faoui is a tenured uni­ver­si­ty pro­fes­sor attached to UTC-Heudi­asyc, a CNRS/UTC UMR. She is also head of the “Inter­act­ing Robot­ic Sys­tems” team, one of the lab­o­ra­to­ry’s three teams and Direc­tor of SIVAL­ab, a joint lab­o­ra­to­ry between UTC, CNRS and Renault.

“The research project we’re devel­op­ing in this joint lab­o­ra­to­ry con­cerns the local­iza­tion and per­cep­tion integri­ty of intel­li­gent vehi­cles. Research scientists/engineers and man­u­fac­tur­ers pre­fer this char­ac­ter­i­za­tion, as total auton­o­my is not for tomor­row. How­ev­er, by aim­ing for auton­o­my, we can devel­op inter­me­di­ate sys­tems such as dri­ving aids, while ensur­ing vehi­cle safe­ty. Intel­li­gent vehi­cle research is part of the field of robot­ics, and robot­ics needs AI to increase its deci­sion-mak­ing auton­o­my,” she says.

The autonomous or intel­li­gent vehi­cle is equipped with high­ly com­plex sys­tems. “These are sys­tems that must enable it to nav­i­gate at high speed and evolve in extreme­ly diverse, com­plex envi­ron­ments where the dif­fer­ent dynam­ics of all road users must be tak­en into account, such as trucks, oth­er vehi­cles, pedes­tri­ans, cyclists, scoot­ers, etc.,” she stresses.

So, rather than full auton­o­my being the hori­zon for the time being, efforts are being made to devel­op dri­ving part­ner vehi­cles aimed at increas­ing peo­ple’s mobil­i­ty and safety.

The role of AI in the com­plex sys­tems imple­ment­ed in intel­li­gent vehi­cles? “AI is involved at var­i­ous lev­els in autonomous vehi­cles. In robot­ics, we talk about the “per­cep­tion-deci­sion-action” cycle. In the first stage, we try to per­ceive our envi­ron­ment and under­stand it. Based on this, the sec­ond step is to decide on the manoeu­vre to be car­ried out, then final­ly to imple­ment this deci­sion through actions on the motor, the wheels, the steer­ing wheel and so on. AI can be applied at all lev­els. In recent years, learn­ing tech­niques based on deep neur­al net­works (deep learn­ing) have made enor­mous progress in envi­ron­men­tal per­cep­tion. These sys­tems can detect vehi­cles, pedes­tri­ans, road mark­ings, signs and many oth­er ele­ments of a road scene with very high suc­cess rates.

How­ev­er, they are “black box­es” and it is dif­fi­cult to pre­dict and explain the cas­es that don’t work. There is AI at the manoeu­vre deci­sion lev­el too, with tech­niques that can be based on tra­jec­to­ry plan­ning and pre­dic­tion and risk cal­cu­la­tion. On the oth­er hand, there is cur­rent­ly lit­tle AI in the action phase, since com­mand and con­trol algo­rithms based on vehi­cle dynam­ic mod­els are effi­cient”, explains Véronique Cher­faoui. How­ev­er, oth­er avenues are being explored. “Research is being car­ried out in what is known as “End to End”. This involves devel­op­ing deep neur­al net­works that take sen­sor data (cam­eras, lidars, etc.) as input, where out­put is a direct action on the steer­ing wheel, gas ped­al or brake. How­ev­er, this approach has its lim­its, because we can’t imag­ine every pos­si­ble sit­u­a­tion and we can’t explain what led to a par­tic­u­lar deci­sion. Car man­u­fac­tur­ers are not yet ready to embark on this type of sys­tem, because their lia­bil­i­ty is at stake, and we can’t guar­an­tee oper­a­tional reli­a­bil­i­ty”, she believes.

In any case, this is not the path cho­sen by Véronique Cher­faoui, who insists that she is first and fore­most a roboti­cist. “I don’t devel­op gen­er­a­tive AI tools or neur­al net­works, for exam­ple. My role is to adapt AI tools to my ques­tions as a roboti­cist. In the deci­sion­mak­ing phase, for exam­ple, I can use deep neur­al net­works, but also rea­son­ing, tak­ing uncer­tain­ties into account”, she explains. From per­cep­tion to action, uncer­tain­ties are many and var­ied. Yet neur­al net­works have a hard time mod­el­ling uncer­tain­ties. “The project we’re run­ning with Renault is based on tak­ing uncer­tain­ties into account, from the sen­sor right through to deci­sion­mak­ing. If our infor­ma­tion is too uncer­tain, we com­mu­ni­cate this to the sys­tem, which could give the hand back to the dri­ver because it is unable to decide”, she explains.

These issues are at the heart of the CAP TWINNING col­lab­o­ra­tive project, part of the PostGenAI@Paris AI clus­ter sup­port­ed by Sor­bonne Uni­ver­si­ty. “The idea is for the vehi­cle to be a dri­ving part­ner, with dri­ving shared between the human and the vehi­cle. We assume that the vehi­cle knows how to man­age a giv­en sit­u­a­tion, but when it feels it is in dif­fi­cul­ty, it must tell the dri­ver, who can then take over again at any time. In this way, the dri­ver is always in the loop. In this project, the idea is to imple­ment inter­faces and AI that enable the dri­ver to under­stand what the car is doing, and con­verse­ly, for the dri­ver to under­stand what the car is doing. When the sys­tem makes a deci­sion, it has to be able to explain it to the part­ner dri­ver. The aim is to give the vehi­cle max­i­mum auton­o­my, an auton­o­my designed to facil­i­tate the mobil­i­ty of cer­tain peo­ple and improve road safe­ty,” con­cludes Véronique Cherfaoui.

MSD