PREVLA: Unified Vision-Language-Action Model via Integrated Perception-Reasoning-Execution for Generalized Embodied Robotic Intelligence

Current embodied AI systems are fundamentally limited by core generalization bottlenecks, preventing robots from adapting to the complexities of real-world tasks. This deficit creates a critical disconnect between perception, reasoning, and execution, hindering the development of truly autonomous systems.This fellowship introduces PREVLA, a unified cognitive framework designed to systematically solve these challenges. The project will achieve generalized Vision-Language-Action capabilities through three synergistic objectives: (1) to achieve Perception Generalization across complex and diverse scenarios by developing novel multimodal alignment techniques; (2) to achieve Reasoning Generalization across open-vocabulary instructions by creating semantically-stabilized architectures capable of robust visual imagination; and (3) to achieve Execution Generalization across diverse robotic platforms by pioneering real-time, parallel action frameworks that eliminate discretization degradation.This ambitious vision will be realized through a work plan engineered to deliver a pathway from theoretical breakthrough to industrial impact. The project will translate cutting-edge machine learning methodologies—masked self-supervised learning for perception, mixture-of-experts (MoE) architectures for reasoning, and flow matching for execution—into a unified framework. The project's breakthrough will empower robots to perform complex, multi-step manipulation from natural language, significantly advancing the state-of-the-art. The project's scientific impact will be driven by a strategy of targeting high-impact publications and the full open-source release of the PREVLA framework. By addressing key market deployment barriers, this research holds significant potential to enhance European competitiveness in manufacturing and healthcare. The project's ultimate vision is to contribute to a future where human-robot collaboration is safe, intuitive, and efficient.