Cosmos 3 Edge: Global Model for Embedded Physical AI

NVIDIA has launched Cosmos 3 Edge, a 4-billion-parameter global model designed for physical AI systems and robots operating at the edge. This open-source model enables machines to understand their environment, reason in real-time, and generate robotic actions directly on resource-constrained devices, such as NVIDIA RTX PRO and Jetson GPUs.

Cosmos 3 Edge stands out for its ability to model the world by learning how an environment changes over time, representing objects, movements, and the effects of actions. It combines two transformer towers (autoregressive and diffusion) sharing a common multimodal representation, enabling understanding, prediction, and action generation. The model can simulate possible futures and connect these simulations to concrete actions, thus providing a basis for robotic policy learning. It achieves state-of-the-art performance in real-time inference and accuracy on key benchmarks.

NVIDIA also provides post-trained checkpoints and scripts to enable developers to adapt and optimize Cosmos 3 Edge for their specific workloads, particularly for robotic manipulation tasks. The goal is to provide an open platform for building world models tailored to specific domains, with continuous improvements planned for interactive world generation, driving scenario simulation, and robotic policies.