On November 21 and 22, Île-de-France Mobilités organized the AI and Mobilities Hackathon, with the aim of exploring the uses of AI in the service of mobility.

Over two days, the 10 teams formed " triturated " mobility data, manipulated it, coupled it with weather data, with event agendas, used it in machine learning algorithms, exploited the possibilities offered by language models (LLM)... to end up with mostly functional prototypes presented to the jury.

Some of this Île-de-France Mobilités data, supplemented by specific datasets provided by RATP, SNCF and the Grand Paris Seine et Oise urban community, was never before seen. They enabled the teams to imagine new services for passengers, new tools for agents, and to take up one or more of the challenges we set them:

●      Improving the accessibility of mobility services ;

●      Building an AI toolbox to accelerate the development of AI in the service of users ;

●      Improve forecasts at the service of mobility;

●       Personalize the user experience of digital services to the traveler.

To these challenges was added a final cross-cutting challenge to explore the issue of AI system frugality, with the aim of approaching it from the angle of project improvement, notably through the design of solutions that are potentially deployable locally.

Accessibility in the spotlight

Among the most remarkable features of this hackathon were the large number of teams taking up the challenge of accessibility and inclusion, as well as that of personalization. Also noteworthy is the diversity and quality of the participants' profiles, as well as the level of technical work carried out by the teams, whether on data, algorithms, or the use of LLMs.

Une mise en-place-de-moyens-techniques-en-amont...

A special effort was made to provide teams with data and resources to help them develop prototypes.

A datalab was deployed by Île-de-France mobilités on the occasion of the hackathon, this being an instance of Onyxia, an INSEE product distributed under an open license that relies on " cloud " technologies while limiting adhesions with hosting solution providers.

Through Onyxia, the teams had at their disposal development environments python, SQL, an Elasticsearch vector database and the ability to instantiate and prototype chatbots (with Ollama and OpenWebUI). The datasets themselves could also be uploaded to the Datalab (via an S3 storage service - minio). Lastly, IDFM provided candidates with access to generative AI services via team access keys.

To encourage the emergence of ideas and proposals, upstream work was carried out to identify data of various typologies: long, detailed time series (train passage time history communicated by SNCF Transilien, validation data, aggregated Itinerary search data, etc.), textual passenger information data, real-time API from the PRIM platform, location information such as the position of elevators, stairs and call points (RATP data), etc.

By targeting relevant datasets for each challenge, we aimed to make it easier for teams to get started and project their choice of solution.

... and operational results...

One of the hackathon's objectives was to pragmatically identify the keys to rendering operational services to users thanks to AI.

The possibilities for classification, text transformation, and prediction, which had been identified ahead of the hackathon by IDFM were well highlighted, particularly by projects aimed at simplifying and contextualizing traveler information.

The teams were also interested in the multimodal nature of generative AI (ability to process various sources : images, text, sound, video...) to imagine dynamic and fluid uses for the benefit of effective passenger information, particularly when they are cognitively impaired.

Thus, in addition to LLM call-for-service techniques and PRIM APIs, several teams have been working on implementing ergonomic traveler information solutions, often making use of audio (whisper, local text2speech, and OpenAI APIs) with a focus on responsiveness (using websocket technologies). Solutions using image analysis were also demonstrated, notably for the simplified and contextualized interpretation of traffic information display screens.

... which allow us to imagine tomorrow's challenges for IDFM.

This work highlights the wealth of perspectives offered by multimodal language models, which provide direct operational means to enable traveler information to be adapted to the multiplicity of types of users and uses of public transport.

To activate these opportunities, one of the challenges for Île-de-France mobilités will be to assume its role as a regional platform for the centralization of quality data.

These multimodal AIs would indeed benefit from being fed by specifically annotated datasets as yet non-existent on a regional scale and which will have to be built up such as images, sounds in stations, on trains or station plans.

It should be noted, moreover, that other data needs were expressed by the candidates and would motivate a centralization work:

  • detailed data relating to the maintenance of the elevators, their environment (hygrometry, ambient temperature etc..), and their use, to build predictive maintenance models;
  • data relating to operating incidents and their causes, to develop forecasting models;
  • or data relating to the presence of agents in stations and stations, and the location of call points to propose use cases around the feeling of safety.

In a context of rapidly accelerating information processing technologies (around AI and data), the Hackathon format as a collective effort to federate ideas and resources, is an interesting contribution to the maturation of a collective vision. It brings public policy issues to the fore and enables them to be taken into account at a key moment in the adoption of technologies and their implementation on an industrial scale.

This format is also an opportunity to animate a community of experts who, in their daily work, are ambassadors for the collective vision and the necessary improvements noted collectively.

Shared software bricks

With a view to capitalizing on the future AI projects of IDFM and the mobility ecosystem, all the developments made have their open code and functional code extracts are highlighted to encourage inspiration and reuse.

While we can't mention all the potential Reuses, we can nevertheless highlight the bricks proposed by the Accit/FALC team that enable content to be " translated " into Easy to Read and Understand. A commonality to be developed for as many people as possible, beyond mobility services !

You too can share your data Reuses with us by clicking on this link.

We naturally remain available to welcome your comments and answer your questions at [email protected] and on Slack. 

Voluntary participants and a mobilized organization team !

We'd like to say another huge thank you to the 55 participants and attendees divided into 10 teams made up of both Île-de-France Mobilités partners (transport operators, service companies...) and individual participants (AI students, mobility experts). They were willing, very dynamic and fertile in their productions and exchanges. All the teams contributed their energy, their approach, their desire and their involvement in the challenges to be met. We would also like to thank all the experts who supported the teams over these two days.

Thank you also to the Hackathon jury who accepted the difficult responsibility of evaluating the projects : David Assouline (SNCF Transilien), Mounia Latrech (RATP), Jean-Daniel Alquier (GPSEO), Jules Pondard (DINUM) and Hélène Brisset, IDFM's Digital Director.

Finally, we would also like to thank Delphine Bürkli, Mayor of the 9th arrondissement of Paris and Director of Île-de-France Mobilités, for her presence at the event and the many exchanges she was able to have with the teams.

Tour d'horizon des 10 projets proposés

Mobil'IA project 🥇 (First prize), led by Alexandre Le Ray, Jérémy André, Alexandre Meyer and Pierre Farret.

Mobil'IA is an inclusive application that uses AI to make public transport travel accessible to people with disabilities.

MAIdfm project 🥈 (Second prize), led by Simon Aharonian, Nawel Boumerdassi, Corentin Barand, Achille Popelier and Zakaria Hammal.

mAIdfm is a project focused on personalizing the use of transport applications through artificial intelligence. It aims to transform the public transport experience through:

  • Personalized recommendations based on user clustering.
  • Gamification with playful badges, rankings and challenges.
  • Intelligent notifications on inspired traffic, weather or local events, tailored to user preferences.

" Our project focused on using artificial intelligence to offer each Île-de-France Mobilités traveler unique, personalized and fun content, based on their profile, favorite routes, and transportation habits," explains the team.

" It was an intense experience, but very rewarding "confirms Achille, team member, " The time pressure was there, but it motivated us to give our best. ".

Accit/FALC project 🥉 (3rd prize), led by Gaëtan Bloch, Marc Bresson, Etienne Vautherin, Cécile Balsier, Katia Awona and Théo Lartigau.

Accit/FALC is a personalized application, accessible online and offline, that provides the necessary information at the right time, on the right channel and according to the identified traveler profile and disability. In particular, this project attempts to implement the FALC format, Facile A Lire et à Comprendre. This application is coupled with an API enabling the centralization of a text simplification and scoring service, making it a potentially deployable and reusable project beyond the scope of Ile-de-France mobility.

PRIM news in your mailbox every month

PRIM news and upcoming data available in your mailbox every month

Subscribe to the newsletter