NVIDIA releases Isaac GR00T N1.6 with the Cosmos open models

Background

Humanoid robots are hard to program. A robot that can walk is not automatically a robot that can be told what to do and work out how to do it. Two problems sit behind this: the robot needs to understand an instruction that may be vague, and it needs to control its whole body to carry that instruction out. NVIDIA framed the Isaac GR00T N1.6 release around exactly these two problems, saying the model enhances reasoning, generalization and whole-body control for humanoid robots^[1].

What was released

NVIDIA published the open-model release on 28 October 2025 in a blog by Bryan Catanzaro^[1]. The centrepiece is Isaac GR00T N1.6, an open model for humanoid robots.

The company said GR00T N1.6 integrates Cosmos Reason, a reasoning vision-language model built for physical AI. The point of that component is to turn vague instructions into step-by-step action plans^[2]. In other words, the reasoning model handles the translation from something a person might say to something a robot can actually execute as a sequence of steps.

Alongside the model, NVIDIA released GR00T training data. The company said that this data had risen into the top 10 most-downloaded Hugging Face datasets of all time^[1]. GR00T N1.6 itself was made available on Hugging Face, and Isaac GR00T models are offered as NVIDIA NIM microservices^[1].

The same release moved Isaac Lab, NVIDIA's robot-learning framework, to an early developer release, version 2.3, with whole-body control and expanded teleoperation for data collection^[2]. Teleoperation matters here because it is one of the ways robot demonstrations are gathered; making it easier to collect data feeds directly into training.

Who was adopting it

NVIDIA said Agility Robotics, Amazon Robotics, Figure AI, Skild AI, Milestone Systems and Uber were adopting Cosmos or Isaac GR00T N models^[1].

Why it matters

Two things stand out about this release. The first is openness. A model, its training data and a simulation and learning stack being made available together lowers the barrier for robotics teams that do not have the resources to build every layer themselves. Training data that becomes one of the most-downloaded datasets on Hugging Face is one sign of how much demand there is for that layer^[1].

The second is the split of work between reasoning and control. Rather than asking one model to do everything, the release pairs a reasoning vision-language model that plans with a model that drives a whole body^[1][2]. That structure reflects a wider pattern in robotics: separating the part that decides what to do from the part that physically does it, so each can be improved and evaluated on its own.

Where it sits

The release is aimed at physical AI — AI systems that act in the world rather than only producing text or images. By packaging a humanoid control model, a reasoning model, training data, a simulation framework and deployment as microservices in a single announcement, NVIDIA presented the components as one route from an instruction to a working robot^[1][2]. The list of companies named as adopters shows the effort was aimed at commercial robotics from the start^[1].