Invited Speakers
Below you will find the confirmed speakers for SoRAIM’27. Course titles and abstracts will be added once the programme is finalised.
Prof. Dr. Elisabeth André
University of Augsburg, Germany
Prof. Dr. Elisabeth André is Full Professor of Computer Science and Founding Chair of Human-Centered Artificial Intelligence at the University of Augsburg, a position she has held since 2001. Prior to joining Augsburg, she spent 13 years as a researcher at the German Research Center for Artificial Intelligence (DFKI) in Saarbrücken, where she rose to Principal Researcher. She holds a diploma and a doctorate in Computer Science from Saarland University. Her research spans multimodal human-machine interaction, embodied conversational agents, affective computing, social signal processing, and social robotics. A flagship contribution is the open-source SSI (Social Signal Interpretation) framework for recording and analysing multimodal signals — including gaze, speech, and gesture — which is now used worldwide to endow robots and virtual agents with the ability to perceive and respond to human emotions.
Course: To be announced.
Dr. Sylvain Calinon
Idiap Research Institute & EPFL, Switzerland
Dr. Sylvain Calinon is a Senior Research Scientist at the Idiap Research Institute and a Lecturer at the École Polytechnique Fédérale de Lausanne (EPFL). He heads the Robot Learning & Interaction group at Idiap, with expertise in human-robot collaboration, robot learning from demonstration, geometric representations, and optimal control.
The approaches developed in his group can be applied to a wide range of applications requiring prehensile and non-prehensile manipulation skills, with robots that are either close to us (assistive and industrial robots), parts of us (prosthetics and exoskeletons), or far away from us (shared control and teleoperation).
Course: To be announced.
Dr. Oya Celiktutan
King’s College London, United Kingdom
Dr. Oya Celiktutan is a Reader in AI and Robotics in the Department of Engineering at King’s College London, where she leads the Social AI & Robotics Laboratory. She is also the Honorary Robotics Lead at Guy’s and St Thomas’ NHS Foundation Trust, working closely with two hospitals to translate socially assistive robotic technologies into clinical settings. Her research focuses on multimodal machine learning for autonomous robots and virtual agents that interact naturally with humans, including multimodal perception, human behaviour understanding and generation, and socially aware navigation and manipulation. Her work has been supported by EPSRC, The Royal Society, and the EU Horizon programme, as well as industrial partners such as Toyota Motor Europe and NVIDIA.
Course: To be announced.
Dr. Alexandre Défossez
Kyutai, Gradium, France
Dr. Alexandre Défossez is a co-founder of Kyutai, a non profit lab for research in artificial intelligence based in Paris committed to open science. His work covers generative speech and multimodal AI (Moshi, Hibiki, DSM) with a strong focus on handling multiple streams jointly across modalities in a streaming and low latency fashion. He is also a co-founder and Chief Science Officer at Gradium, a startup launched in 2025 whose mission is to commercialize the best possible voice AI experience. Before that, Alexandre was a scientist for 3 years at Facebook AI Research in Paris, where he led the development of models for audio compression and modeling (AudioCraft, MusicGen, EnCodec). He graduated in mathematics from École Normale Supérieure, and did his PhD between INRIA and FAIR Paris on music source separation under the supervision of Francis Bach, Nicolas Usinier et Léon Bottou.
Course: To be announced.
Fundamental Courses
Speech Processing: From Modular Pipelines to SpeechLMs
Instructors: Thomas Hueber & Olivier Perrotin
Topic area: Speech Processing
Keywords: Speech signal processing, source-filter model, neural vocoders, self-supervised learning, neural audio codecs, text-to-speech synthesis, ASR, speech tokenization, SpeechLMs, conversational AI
Summary: Builds from speech fundamentals (production/perception, source-filter model, MFCCs) through the four-step evolution of speech representations (signal-based vocoders → neural vocoders → self-supervised encoders → discrete tokenization), then applies these to interactive systems: modular ASR-LLM-TTS pipelines vs. unified multimodal SpeechLMs (SpiRitLM, Moshi, LLaMA-Omni). Historical/technical narrative where each concept motivates the next, richly illustrated with embedded audio demos.
Highlights:
- Historical overview of speech synthesis, putting the most recent techniques into context
- Unifying “4-step” voice-coding framework tying together WaveNet, HiFi-GAN, HuBERT, VQ-VAE/EnCodec
- Live audio demos: WaveNet, VALL-E/IndexTTS2 zero-shot voice cloning, Moshi full-duplex conversation
- Deep dive into a recent full-duplex 3-stream architecture
- Coverage of multimodal frontiers: speech-gesture generation, omni-style vision+audio+text models
Vision and Multimodal Learning
Instructor: Stéphane Lathuiliere
Topic area: Vision and Multimodal Learning
Keywords: self-supervised learning, contrastive learning, vision transformers, CLIP, multimodal LLMs, visual instruction tuning, human-robot interaction, foundation models, image-text alignment
Summary: Opens via human-robot interaction (detection, pose, segmentation), then surveys the shift from supervised ImageNet pretraining to self-supervised representation learning (SimCLR, DINO). Moves to multimodal representation learning centered on CLIP and its extensions (CLAP, ImageBind), then traces multimodal LLMs from Flamingo/LLaVA to Qwen2.5-Omni, closing with open challenges for MLLMs in social robotics. Example- and paper-driven with architecture diagrams and loss derivations.
Highlights:
- Step-by-step SimCLR/CLIP loss derivations plus DINO’s emergent unsupervised segmentation
- CLIP as “foundation model” with zero-shot results across 27+ datasets, extended to CLAP and ImageBind
- Architecture deep dives: Frozen → Flamingo → LLaVA (few-shot VQA to instruction-tuned agents)
- Qwen2.5-Omni streaming audio+video+text demo (live equation-sketching)
- Closing discussion of open MLLM-for-robotics challenges
Robot Navigation: From Physics to Preferences
Instructor: Xavier Alameda-Pineda
Topic area: Control and Reinforcement Learning
Keywords: robot navigation, Model Predictive Control, Q-learning, Deep Q-Networks, policy gradient, PPO, Soft Actor-Critic, RLHF, Direct Preference Optimization, GRPO.
Summary: Uses 2D robot navigation as a running example to trace the arc from classical control to modern preference alignment: (1) classical model-based control (Dynamic Window Approach, MPC), (2) value-based RL (Q-learning, DQN), (3) policy-gradient/actor-critic RL (PPO, SAC), (4) preference-based alignment (RLHF, DPO). Derivation-driven and Socratic, building each method from first principles with toy examples and a closing comparison table across all methods.
Highlights:
- Live MPC demo derived directly from the robot’s kinematics and cost function
- Gridworld walkthrough of Q-value convergence via value iteration
- Clean unifying comparison of DQN vs. SAC critics
- Full derivation of DPO showing the reward model can be eliminated entirely
- Final summary table mapping all four methods by domain, signal, model type, action space