Olivier Sigaud - Foundational models for manipulation robotics: A modular approach
- Date: 09 mars 2026 à 13h
- Salle: 65-66 304
Vision–Language–Action (VLA) models promise “end-to-end” learning of generic manipulation policies for robots operating in arbitrary environments. However, these approaches suffer from the need for large, costly-to-collect datasets of expert trajectories, and their success rates remain insufficient for safety-critical applications. In this talk, I will argue for a more modular approach that leverages reinforcement learning mechanisms, and I will present our current efforts to concretely flesh out such an approach.