Vesta: A Generalist Embodied Reasoning Model

NVIDIA’s Vesta is a unified embodied reasoning model that collapses localization, navigation, embodied question answering, and long-horizon action planning into a single Qwen3-VL-8B-based planner, instead of a brittle multi-model stack. The core technical idea is a curated supervised fine-tuning mixture biased toward spatial intelligence, navigation, grounding, embodied reasoning, and real-robot data, combined with a minimalist multimodal memory harness that interleaves retained image frames with a running text cache of prior subtasks. Across embodied cognition and localization benchmarks, Vesta reaches 68.7 average cognition and 69.9 average localization, outperforming RynnBrain and RoboBrain 2.5 in most categories; on offline real-robot planning it scores 75.4 average versus 38.5 for RoboBrain-2.5-8B. On navigation, it matches the specialist InternVLA-N1 with 55.5 SR and 61.4 OS on R2R-CE val_unseen. Most notably, on a bimanual YAM gripper robot with Gr00t-N1.6 as the low-level actor, Vesta improves success by 38.3% over actor-only control and 25% over a Qwen3-VL planner, showing that a single generalist planner can outperform specialized systems while remaining directly deployable in hierarchical robotic control.