NVIDIA AI Presents ThinkAct: Vision-Language-Action Reasoning via Reinforced Visual Latent Planning

NVIDIA and National Taiwan University researchers have developed ThinkAct, an embodied AI framework that advances vision-language-action (VLA) reasoning by integrating reinforced visual latent planning to connect high-level multimodal reasoning with low-level robotic control. Unlike traditional end-to-end VLA models, ThinkAct employs a dual-system architecture featuring a multimodal large language model (MLLM) that generates structured, step-by-step visual plan latents, enabling improved long-term planning, adaptability, and robustness in complex, dynamic environments.