NVIDIA Introduces Cosmos 3 Edge for On-Device Robot Control

Robots require policies that adapt to sensors, environments, and tasks while running on onboard computing hardware. [1] World models provide a foundation for learning physical interactions, but their large size often complicates on-device deployment. [1] NVIDIA Cosmos 3 Edge addresses this by offering a 4B omni-model that includes a 2B NVIDIA Nemotron-based reasoner. [1] This model is small enough to operate on-device using NVIDIA Jetson Thor. [1]

Cosmos 3 Edge was pretrained on the same physical-world data as the larger Cosmos 3 Nano and Cosmos 3 Super models. [1] This shared pretraining provides the model with a grounding in object movement and interactions. [1] Developers can post-train Cosmos 3 Edge to predict robot actions and serve the resulting policy on Jetson Thor. [1] The framework allows for running inference within a receding-horizon control loop and evaluating the policy's behavior in a closed-loop simulation. [1]

NVIDIA's release of Cosmos 3 Edge attempts to solve the computing constraints of deploying complex world models directly onto robots. By shrinking an omni-model to 4 billion parameters, NVIDIA enables onboard processing via hardware like Jetson Thor, reducing reliance on cloud infrastructure for real-time physical interaction tasks. This approach contrasts with larger, cloud-dependent models by prioritizing edge deployment for physical manipulation policies. What the sources don't address: The specific performance benchmarks or latency metrics when running the 4B model inference loops on Jetson Thor compared to its larger counterparts.