PostGenAI@Paris
Selected in 2024 under the Cluster AI call for projects with €35M in funding, PostGenAI@Paris, led by Sorbonne University and operated through its Sorbonne Cluster for Artificial Intelligence (SCAI), aims to contribute to France’s national AI strategy by establishing an international hub of excellence dedicated to post-generative AI. This new era of AI moves beyond mere content generation to foster deeper understanding and more autonomous, context-aware interactions with its environment. PostGenAI@Paris focuses on three main areas: breakthrough technologies, health, and resilient societies.
The scientific program of PostGenAI@Paris is built on two main pillars. The first consists of 21 collaborative acceleration programs (PACs), which integrate research, training, and innovation. These PACs are essential for developing research-based education and promoting industrial collaborations on cutting-edge topics. The second pillar is a program designed to foster interconnection among consortium members while ensuring the cluster’s agility and long-term evolution.
IRCAM contributes to PostGenAI@Paris both as leader of the PAC AI-MADE (Analysis/Synthesis team) and as participant in PAC TWINNING (Sound Perception and Design team).
AI-MADE
AI for Music And voice meDia gEneration
AI-MADE aims to develop innovative generative models for professional audio content generation, covering applications for music production and live performance (real-time use), cinema, and video games. The central scientific problem to be investigated is the intuitive and precise control of sound properties, notably related to translating from idea/concept into sound. Particularly interesting are questions related to contextualization, style, and expressivity, but also hybridization of sounds. AI-MADE will develop new disentanglement strategies that allow representing sound in latent spaces that directly expose perceptually relevant high-level parameters, and avoid the cumbersome task of coherently tuning a large number of low-level parameters to achieve a desired sound.
AI-MADE will search for properly adapted inductive bias to ensure computational efficiency as well as data efficiency. AI-MADE will favour grey box models that support analysis/synthesis procedures such that the resulting generative models may serve for sound and style analysis.
AI-MADE will develop prototypes for the evaluation in the context of audio production, music performances, as well for analysis in the emerging field of computational humanities (musicology, phonetics).
TWINNING
TrustWorthy Interactive augmeNted autoNomous drivING
TWINNING addresses the broad challenge of autonomous driving, an emerging and rapidly expanding field within the transport industry, where AI technologies play a central role. Its overall goal is twofold: on the one hand, to study shared driving between a human driver and an autonomous system in a dynamic, uncertain road environment with interactions from other road users; on the other hand, to enhance the driver’s collaborative driving capabilities through reliable, high-performance human-AI interactions that support perception, control, and decision-making.
Specifically, PAC TWINNING focuses on co-driving between a driver and a vehicle whose autonomous navigation capabilities, powered by machine learning, are monitored by an integrity controller. Autonomous navigation on open roads remains a challenge due to the uncontrolled, uncertain nature of the environment, the multitude of driving interactions with other users, and the limits of situational understanding that hinder fully autonomous decision-making. The project therefore seeks to enhance the driver’s collaborative driving skills by providing effective interaction mechanisms that improve situational awareness, decision-making, and system control. These interactions will be based on outputs from AI models that are carefully monitored to ensure reliability and trustworthiness.
The TWINNING structure is organized around three main research areas:
1. Scene understanding and robust perception under uncertainty. This involves managing and propagating uncertainties in perception functionalities, with a focus on environment monitoring to ensure compliance with predefined operational design domains.
2. Decision-making and trajectory planning in dynamic, uncertain environments. The goal is to develop AI-based methods for reactive planning that comply with driving rules, traffic codes, and more; these methods will also provide predictive and anticipatory capabilities in highly dynamic contexts to handle interactions with other road users.
3. Shared human/vehicle control, including mechanisms to build driver trust, and the development of augmented, adaptive human-machine interfaces. The driver acts as an operator interacting with the intelligent vehicle via interfaces tailored to its capabilities. Research on haptic, auditory, and augmented reality interfaces will enable the design and study of new modes of driver-vehicle interaction.