Vesta: A Generalist Embodied Reasoning Model

Vesta is introduced as a unified embodied reasoning model that integrates localization, spatial reasoning, navigation, and long-horizon planning into a single foundation model, addressing the inefficiencies of multi-model stacks. It utilizes a large curated corpus for spatial grounding and a multimodal memory system, achieving over 20% improvement on average against individual state-of-the-art (SOTA) models and over 10% against ensembles of top-performing models across various benchmarks. This model significantly enhances task success rates by more than 35% in real-world robotic applications, suggesting that generalist models can effectively replace specialized systems, offering a scalable solution for practitioners in AI and robotics.