At the recent SIGGRAPH 2026, a beacon event for computer graphics and interactive techniques, NVIDIA took center stage with an astonishing presentation by Ming-Yu Liu, the Vice President of Cosmos Lab. Liu’s talk delved into groundbreaking advancements in the field of Physical AI, focusing on the powerful capabilities of NVIDIA’s Cosmos platform. This transformative technology promises to unify world understanding, generation, and action within a single model.
Cosmos 3 is NVIDIA’s latest frontier, aiming to bridge the gap between AI and the intricate realities of the physical world. This platform stands as a world foundation model, designed to redefine the data paradigm in AI-driven robotics and autonomous systems. Cosmos 3 introduces advanced capabilities such as Cosmos 3 Edge, enabling real-time inference on devices, and Cosmos Dreams, facilitating closed-loop simulations that benefit autonomous vehicles.
Based on content from NVIDIA Developer
In his keynote, Liu walked the audience through a historical overview of NVIDIA’s journey, highlighting pivotal innovations like the invention of the programmable GPU in 2001 and the introduction of CUDA, which transformed GPUs into potent parallel processors. He demonstrated how GPUs, originally crafted for computer graphics, ignited a revolution in AI, ultimately culminating in breakthroughs like real-time ray tracing and neural rendering.
The foundational premise of Cosmos is its ability to generate scalable data from a robot’s perspective, overcoming the constraints of slow data collection and scenario hand-crafting present in existing systems. “Compute is data,” Liu declared, explaining how every cycle on a GPU can now produce new test scenarios and variations, offering a robust data engine for scaling robot capabilities. Cosmos offers developers improved data, enriched environments, and solid starting points, thus making Physics-accurate simulation generation and video analytics more accessible.
During the presentation, Liu outlined how Cosmos employs a unique mixture of transformer architecture, which includes an auto-regressive tower for reasoning and a diffusion tower for generating vision, audio, and action tokens. These dual towers share multimodal attention, aligning various informational inputs into a coherent output, crucial for generating physically accurate simulations.
One of the most striking features of Cosmos is the establishment of a common action language utilizing geometric primitives such as ego pose and gripper actions. This universal representation allows varied embodiments, from humanoid robots to self-driving cars, to benefit from shared learning insights, signifying a leap towards more nuanced cross-embodiment transfers and adaptability in robotics.
NVIDIA’s Cosmos platform is not a one-off innovation but rather a compounding suite. Since its introduction in early 2025, it has integrated world understanding, policy action, and simulation into a cohesive framework, now culminating in Cosmos 3. NVIDIA is committed to making Cosmos an open platform, offering models, datasets, and libraries that empower developers to build and evolve custom world models tailored to specific needs.
The session concluded with the unveiling of Cosmos Dreams, a neural closed-loop simulator catering initially to autonomous vehicles. This technology was deftly demonstrated live, showcasing how artificial environments are generated in real-time to optimize training and policy generation without the need for physical trials.
The advancements exhibited at SIGGRAPH 2026 by NVIDIA serve as a harbinger of the AI-driven future, spotlighting the role of advanced modeling platforms like Cosmos in crafting a world where technology seamlessly converges with physical reality.
Get started with NVIDIA Cosmos
For those who couldn’t attend the conference, more detailed insights about Cosmos 3 can be explored through the NVIDIA blog and Hugging Face article.
