NVIDIA Cosmos 3 Edge Brings a 4B World Model to On-Device Robot Control

A smaller Cosmos model for robots at the edge

NVIDIA published a new developer workflow on August 19, 2026 for post-training Cosmos 3 Edge into an on-device robot manipulation policy. Cosmos 3 Edge is a 4-billion-parameter omni-model in the Cosmos 3 family and includes a 2-billion-parameter NVIDIA Nemotron-based reasoner. NVIDIA says it was pretrained on the same physical-world data foundation used for Cosmos 3 Nano and Cosmos 3 Super.

The key difference is deployment size. The model is small enough to fit on NVIDIA Jetson AGX Thor, allowing a robot policy to run locally instead of sending every control decision to a data-center GPU. That matters for robotics because control loops are sensitive to network delay, availability and device memory.

Turning a world model into an action policy

The tutorial demonstrates how developers can post-train Cosmos 3 Edge to predict robot actions, serve the resulting policy on Jetson Thor and evaluate it in a closed-loop simulation. The training workflow is reproducible from NVIDIA's open cosmos-framework repository, and NVIDIA provides a released checkpoint through Hugging Face.

The approach starts from a world model that already encodes patterns about object motion and physical interaction, then adapts it to action prediction for a specific manipulation setting. This can reduce how much task-specific data the policy needs to learn basic physical relationships from scratch.

On-device timing and control loop

NVIDIA reports that on a Jetson AGX Thor T5000, the DROID action policy generates an action chunk in about 1.53 seconds at 640x540 resolution and 15 Hz. Each generated chunk spans about 2.13 seconds of robot motion, which means the next chunk can be ready before the previous one finishes.

The model therefore operates as a receding-horizon controller: it generates a chunk of future actions, executes them, then replans on the next inference cycle rather than making a new model call for every single observation. NVIDIA says this lets the robot move continuously without keeping a data-center GPU in the loop.

Reported closed-loop results

In NVIDIA's RoboLab closed-loop tasks, the post-trained policy reached a 22.9% success rate. That number shows the workflow is still far from universal manipulation reliability, but it is a useful proof that a relatively compact world model can be adapted into a complete on-device policy backbone and run in real time on an embedded robotics platform.

The result should be interpreted as a vendor-reported task result rather than a general robotics benchmark. Performance will depend heavily on embodiment, camera setup, action space, task difficulty, training data and control frequency.

Why this matters

Robot foundation models are often too large or too slow for local deployment. Cosmos 3 Edge represents a different design point: retain enough physical-world representation to support post-training, but keep the model compact enough to serve on an edge robotics computer.

For developers, the release is especially relevant because NVIDIA provides a concrete end-to-end path from pretrained world model to action policy, deployment and closed-loop evaluation. That makes Cosmos 3 Edge more than a research checkpoint; it is a reproducible developer workflow for testing world-model-based robot control without mandatory cloud inference.