NVIDIA Isaac GR00T 1.7: End-to-end humanoid robot workflow

NVIDIA today details the Isaac GR00T development platform and its GR00T 1.7 vision-language-action model for humanoid robots, presenting a complete, open workflow for simulation, data collection, post-training and on-robot deployment. The platform bundles Isaac Lab-Arena for environments, Isaac Teleop for demonstrations, GR00T 1.7 for policy training, and Isaac ROS plus Jetson Thor for deployment.

What is the NVIDIA Isaac GR00T platform?

The Isaac GR00T platform is an open, modular end-to-end workflow that spans simulation environment setup, teleoperation data collection, model training, large-scale evaluation and deployment. Developers can use individual components or the full validated NVIDIA software stack: Isaac Lab-Arena for environment composition, Isaac Teleop for collecting demonstrations, Isaac GR00T for policy training and Isaac ROS with Jetson Thor for exporting and deploying models as LEAPP bundles.

The platform is intended to reduce integration work by unifying formats and tooling across stages, enabling teams to move from humanoid bring-up to task-specific skill development more quickly.

What’s new in GR00T 1.7 and what are its concrete specs?

GR00T 1.7 is a commercially usable, open VLA model released under the Apache 2.0 license with a base checkpoint of 3 billion parameters. It was pretrained on approximately 32,000 hours of real human demonstration and ego-centric video plus about 8,000 hours of simulated rollouts and demonstrations from BEHAVIOR, RoboCasa and Simulated GR-1.

The model swaps its previous Eagle backbone for a Cosmos-Reason2-2B backbone (the Qwen3-VL architecture), which NVIDIA says supports flexible resolution and encodes images in native aspect ratio without padding. GR00T 1.7 adds full pipeline export to ONNX and TensorRT, improves export reliability, and ships with a stronger task decomposition that the company links to better long-horizon reasoning, motion quality and cross-embodiment generalization.

NVIDIA published benchmark improvements versus GR00T N1.6: DROID-F0 up by 10% and DROID-F6 up by 61%, plus SimplerEnv Bridge +5% and Fractal +2%. The model and weights are accessible via GitHub and Hugging Face.

How does the end-to-end workflow work in practice?

The GR00T workflow moves from simulated scene composition to demonstration capture, conversion into GR00T’s LeRobot format, post-training and evaluation, then on-device deployment. In the platform walkthrough NVIDIA shows a pick-and-place example where a humanoid uses AGILE whole-body control and PinkIK to move an apple onto a plate.

Demonstrations are collected with Isaac Teleop using a supported VR headset streamed over CloudXR. NVIDIA’s example records 400 trajectories for the apple pick-and-place tutorial; the provided record command shows flags such as --num_demos 400, --enable_cameras, and output to an HDF5 dataset file. Successful HDF5 recordings are converted to LeRobot format using a YAML mapping (the example uses fps: 50 and chunks_size: 1000), producing parquet state/action files, MP4 camera recordings and dataset metadata for post-training.

Post-training fine-tuning runs in a standalone checkout of the Isaac-GR00T repo; the example tunes the visual backbone, projector and diffusion model while keeping the language model frozen. Once post-trained, models can be exported to ONNX/TensorRT and packaged for on-device inference and control via Isaac ROS and Jetson Thor.

Why it matters

GR00T 1.7 and the Isaac GR00T platform package the previously fragmented pieces of humanoid development into a single, open pipeline. The specific pretrained scale (about 32,000 hours of real demonstrations plus 8,000 hours of simulation), the 3 billion-parameter base checkpoint, and explicit ONNX/TensorRT export targets lower the friction between research prototypes and robot deployment. The benchmark gains NVIDIA reports (DROID-F0 +10%, DROID-F6 +61%) suggest material improvements in robustness and long-horizon tasks, which are central pain points for multi-step humanoid behaviors.

What to watch

Watch adoption of the GR00T components in third-party robotics stacks and the public checkpoints on GitHub and Hugging Face. The next confirmatory signals will be independent evaluations using the cited benchmarks and published examples of cross-embodiment transfers or on-robot deployments using the ONNX/TensorRT export path.