---
title: "Post-Train NVIDIA Cosmos 3 Edge for On-Device Robot Control"
description: "Deploying large world models on robots is constrained by onboard memory and the control latency needed for continuous manipulation. The tutorial post-trains NVIDIA Cosmos 3 Edge, a 4B omni-model with a 2B NVIDIA Nemotron-based reasoner, into a DROID action policy using 76,000 successful teleoperated trajectories, then serves it through an OpenPI-compatible WebSocket server on Jetson Thor. The configuration predicts 32 future 8-dimensional joint-position actions at 15 Hz, while a receding-horizon loop executes a prefix and replans from fresh camera and robot-state observations. On an NVIDIA Jetson AGX Thor T5000, inference takes about 1.53 seconds per action chunk covering roughly 2.13 seconds of motion, and the policy reaches 22.9% success in closed-loop RoboLab evaluation; the source contrasts this with Cosmos 3 Nano’s 36.8%."
---

# Post-Train NVIDIA Cosmos 3 Edge for On-Device Robot Control

[NVIDIA](https://yomu.fyi/company/nvidia-developer-blog) · Michelle Horton · Aug 19, 2026

**Type:** Tutorial

## Summary

Deploying large world models on robots is constrained by onboard memory and the control latency needed for continuous manipulation. The tutorial post-trains NVIDIA Cosmos 3 Edge, a 4B omni-model with a 2B NVIDIA Nemotron-based reasoner, into a DROID action policy using 76,000 successful teleoperated trajectories, then serves it through an OpenPI-compatible WebSocket server on Jetson Thor. The configuration predicts 32 future 8-dimensional joint-position actions at 15 Hz, while a receding-horizon loop executes a prefix and replans from fresh camera and robot-state observations. On an NVIDIA Jetson AGX Thor T5000, inference takes about 1.53 seconds per action chunk covering roughly 2.13 seconds of motion, and the policy reaches 22.9% success in closed-loop RoboLab evaluation; the source contrasts this with Cosmos 3 Nano’s 36.8%.

## Context

On-device robot policies must fit the robot’s available memory and meet the control-latency requirements of the task. The post presents Cosmos 3 Edge as a smaller world foundation model intended to address those constraints on Jetson Thor without offloading inference to a data-center GPU.

## Approach / What changed

The tutorial post-trains Cosmos 3 Edge on the Cosmos3-DROID dataset, converts the base checkpoint to distributed-checkpoint format, and serves the resulting policy through an OpenPI WebSocket server on Jetson Thor. A receding-horizon controller executes a prefix of each action chunk and replans using new camera and robot-state observations, with RoboLab providing closed-loop simulation evaluation.

## Takeaways

- Cosmos 3 Edge is a 4B omni-model with a 2B NVIDIA Nemotron-based reasoner, and its weights occupy roughly 9 GB in BF16 for Jetson Thor deployment.
- The policy emits 32 future 8-dimensional joint-position actions at 15 Hz; the example loop executes the first 16 actions before requesting a new chunk.
- The post-trained Edge policy reaches 22.9% success across the closed-loop RoboLab task suite, compared with 36.8% for Cosmos 3 Nano, reflecting the stated trade-off for fully on-robot real-time operation.

**Tags:** [Deployment](https://yomu.fyi/topic/deployment), [Machine Learning](https://yomu.fyi/topic/machine-learning), [Performance](https://yomu.fyi/topic/performance), [Testing](https://yomu.fyi/topic/testing)

- Source: [NVIDIA](https://developer.nvidia.com/blog/post-train-nvidia-cosmos-3-edge-for-on-device-robot-control)
- Source URL: https://developer.nvidia.com/blog/post-train-nvidia-cosmos-3-edge-for-on-device-robot-control
- Ingested by Yomu: 2026-08-28T20:00:58.721Z

[Read original post](https://developer.nvidia.com/blog/post-train-nvidia-cosmos-3-edge-for-on-device-robot-control)