NVIDIA Isaac GR00T N1.7

What it is

On 16 March 2026, at its GTC event, NVIDIA announced Isaac GR00T N1.7, describing it as an open reasoning vision-language-action model purpose-built for humanoid robots and now commercially viable for real-world deployment^[1]. A vision-language-action model is one that takes camera images and a language instruction as input and outputs motor commands directly, rather than only describing a scene. The model card uses the same framing: an open foundation model for generalised humanoid robot reasoning and skills that accepts multimodal input, including language and images^[2]. Where a conventional robot controller is written for one task in one workspace, a foundation model of this kind is meant to be adapted to many.

How it is built

GR00T N1.7 is the medium-sized member of the family^[2]. Its architecture combines pre-trained vision and language encoders with a flow-matching action transformer that models a chunk of actions conditioned on vision, language and proprioception^[2]. In plain terms: the encoders turn images and instructions into internal representations, proprioception supplies the robot's own sense of joint position and motion, and the action transformer produces a short sequence of movements at once instead of a single step. Flow matching is a generative technique that learns to carry samples from a simple noise distribution towards a target distribution; here the target is a trajectory of actions.

The vision-language backbone is Cosmos-Reason2-2B, and RGB camera frames are passed through a pre-trained SigLIP2 vision transformer^[2].

Post-trained variants and their data

The family ships four post-trained variants, tuned for SimplerEnv Bridge, SimplerEnv Fractal, DROID and LIBERO^[2]. Some of the data behind them is documented in the model card. The Bridge post-training data is a LeRobot-format conversion of BridgeData V2, containing 60,096 robot manipulation trajectories across 24 environments^[2]. The DROID post-training data has about 76,000 demonstration trajectories, roughly 350 hours, collected across 564 scenes in 52 buildings and covering 86 manipulation tasks^[2]. LIBERO supplies 130 language-conditioned manipulation tasks for lifelong robot learning^[2]. The mix matters because each dataset stresses something different: the breadth of scenes and buildings, the number of distinct tasks, or the requirement to keep learning new tasks over time.

License and training footprint

The model is released under the NVIDIA Open Model License Agreement and is ready for commercial and non-commercial use^[2]. The model card also estimates training energy use at 41,288 kWh and emissions at 16.949 tonnes of CO2 equivalent^[2], figures that make the cost of building a model of this size visible alongside its weights.

Where it sits

Humanoid, LG Electronics, NEURA and Noble Machines are adopting GR00T N1.7 to scale humanoid robot deployment^[1]. For teams that want a humanoid to follow instructions and act on what its cameras see, the value of an openly licensed model is that the perception-and-action stack comes as a starting point that can be post-trained on a team's own data, rather than something each lab has to assemble from scratch.

^[1]: NVIDIA, "NVIDIA Expands Open Model Families to Power the Next Wave of Agentic, Physical and Healthcare AI" ^[2]: NVIDIA / Hugging Face, "GR00T-N1.7-3B model card"