Nvidia has introduced Cosmos 3, a groundbreaking open omni-model designed for physical AI, integrating reasoning, world generation, and action into one cohesive system. Unveiled on May 31, 2025, this innovative model aims to enable robots and autonomous vehicles to comprehend the unpredictable nature of the real world effectively.
What are the capabilities of Cosmos 3? Cosmos 3 can create predictive video sequences lasting up to 30 seconds, deriving insights from text, images, or videos. This ability allows robots to envision their next moves in their environments before initiating any physical actions, enhancing operational efficiency and safety.
How does Cosmos 3 operate? Utilizing a Mixture of Transformers architecture, Cosmos 3 simultaneously processes various input forms. This model supports sound and action modalities, meaning a robot empowered by Cosmos 3 can interpret visual, auditory, and physical data within a unified framework, leading to more intelligent interactions with its surroundings.
What is the model’s practical application? The core function of Cosmos 3 revolves around robot policy learning. It acts as the foundational support for what Nvidia describes as World Action Models (WAMs), enabling robots to navigate new environments seamlessly, even those they have never encountered, thereby expanding their operational envelope.
What developments led to Cosmos 3? Following a series of earlier iterations released in 2025 that emphasized prediction, transfer learning, and reasoning, Cosmos technology caught the attention of significant players in various industries. For instance, Figure AI and Agility Robotics, both working with humanoid robots, have adopted Cosmos technology for their advanced bipedal systems. Likewise, companies in the autonomous vehicle domain such as Uber, Waabi, and Wayve have integrated previous models of Cosmos into their self-driving technologies.
How could this impact the robotics market? The introduction of an open-source Cosmos 3 could significantly boost adoption among smaller companies that may not have the means to develop their world models independently. The model's robust synthetic data generation feature addresses a long-standing challenge in the robotics industry: acquiring sufficient real-world training data without risking valuable equipment.