# New Era in Robot Control with NVIDIA Cosmos 3 Edge

Next-gen robot policies running on Jetson Thor

- **Source:** [Nvidia Developer Blog](https://developer.nvidia.com/blog)
- **Published:** 2026-08-19
- **Topic:** [Data & AI](https://brifing.si/c/data-ai)
- **Original article:** <https://developer.nvidia.com/blog/post-train-nvidia-cosmos-3-edge-for-on-device-robot-control/>
- **On Brifing:** https://brifing.si/s/ff2a68b9-977a-4798-a530-d3977334f8cd

## Next-Gen Robot Control

NVIDIA's Cosmos 3 Edge enables robots to develop policies that adapt to their sensors, environments, and tasks. This model compresses world models' physical interaction learning capability into a size that can run on-device. The 4B parameter omni-model is optimized for embedded systems like Jetson Thor.

## Technical Details of Cosmos 3 Edge

Cosmos 3 Edge features a 2B NVIDIA Nemotron-based reasoning engine. It was pretrained on the same physical-world data as NVIDIA Cosmos 3 Nano and Super. The model provides solid grounding in how objects move and interact. These features simplify retraining robot policies with task-specific data.

## Real-Time Performance

Cosmos 3 Edge runs at 15 fps on NVIDIA Jetson AGX Thor T5000. It generates each action chunk in about 1.53 seconds while covering roughly 2.13 seconds of robot motion. The model supports continuous streaming without needing a central GPU. This marks a significant milestone in industrial robotics.

## Training Process and Dataset

The model was trained on the nvidia/Cosmos3-DROID dataset containing 76k successful teleoperation trajectories. Collected using Franka Panda arm and Robotiq gripper, it includes approximately 350 hours across 86 tasks and 564 scenes. Data is provided in LeRobotDataset v3.0 format at 640×360 resolution.

## Deployment and Usage

The trained model is served by a WebSocket policy server using OpenPI protocol. The server runs locally on Jetson Thor, fitting 9GB BF16 weights into the device's onboard memory. This enables robots to operate independently without data center connectivity.

## Application Areas

Cosmos 3 Edge supports various platforms including Franka, UR, WidowX 250, and LeRobot SO101. All compatible platforms are listed in the model card. It offers ideal solutions for industrial automation, logistics, and robots in hazardous environments, achieving 22.9% success in closed-loop tests.

## Learning Resources and Community Support

NVIDIA provides comprehensive documentation for model usage. The cosmos-framework GitHub repo contains training scripts and configuration files. Checkpoints are shared via HuggingFace. The community continues using OpenPI protocol for the DROID policy ecosystem.

## Transformation in Robotics

Cosmos 3 Edge proves the practical use of world foundation models in robot control. A 4B parameter model running in real-time on embedded systems opens new horizons in autonomous robotics. This breakthrough promises revolutionary innovations in industrial automation and service robots.

## Sources covering this story

- https://developer.nvidia.com/blog/post-train-nvidia-cosmos-3-edge-for-on-device-robot-control/ — 

---

_Summary written by Brifing from the original reporting. Copyright in the original article belongs to its publisher; read the full piece at the link above._