NVIDIA has introduced the Isaac GR00T Development Platform together with the open-source GR00T 1.7 vision-language-action (VLA) model to unify humanoid robot development workflows. The platform covers the full pipeline from simulation and data collection through post-training, large-scale evaluation and deployment readiness.
Why this matters
Humanoid robotics development remains complex and fragmented: separate tools, incompatible data formats and manual integrations slow teams down. Isaac GR00T aims to reduce integration overhead by providing a modular, fully open stack that produces clean artifacts at every stage so developers can move faster from robot bring-up to task-specific skills.
Platform components and roles
The GR00T development pipeline maps to specific NVIDIA technologies:
- Simulation Environment Setup: NVIDIA Isaac Lab-Arena — build simulation scenes for data generation, training and testing.
- Data Creation: NVIDIA Isaac Teleop — capture robot demonstration data (e.g., via VR teleoperation).
- Policy Training: NVIDIA Isaac GR00T 1.7 + training scripts — train humanoid policies on simulated and real demonstrations.
- Policy Evaluation: NVIDIA Isaac Lab-Arena — test policies in simulation prior to real-world deployment.
- Policy Deployment: NVIDIA Isaac ROS + Jetson Thor — export models into a deployable LEAPP bundle and run on-device inference and control.
The stack is modular: teams can adopt individual components or the full pipeline and integrate their own tooling with NVIDIA’s validated software.
GR00T 1.7: features and updates
- License and availability: GR00T 1.7 is released under the permissive Apache 2.0 license. Model weights and code are available on GitHub and Hugging Face. A base checkpoint of approximately 3 billion parameters is provided.
- Multimodal VLA model: GR00T 1.7 accepts image, language and robot-state inputs and outputs the actions needed for tasks across diverse embodiments.
- Pretraining scale: the model was pretrained on roughly 32,000 hours of real demonstration and human egocentric data plus about 8,000 hours of simulated rollouts and demonstrations (including BEHAVIOR, RoboCasa and Simulated GR-1) to produce more human-like actions.
- New VLM backbone: Cosmos-Reason2-2B (Qwen3-VL architecture) replaces the Eagle backbone used in N1.6, enabling flexible resolutions and encoding images at native aspect ratios without padding.
- Expanded deployment support: full pipeline export to ONNX and TensorRT with improved export reliability and more frequent updates.
- Enhanced performance: better long-horizon task reasoning via task- and subtask-level decomposition, improving reliability, motion quality and cross-embodiment generalization.
- Benchmark improvements: consistent gains over N1.6 across DROID and SimplerEnv (examples: DROID-F0 +10%, DROID-F6 +61%, SimplerEnv Bridge +5%, Fractal +2%).
How to get started (pick-and-place example)
The announcement includes a concrete simulation walkthrough that demonstrates building a dexterous pick-and-place policy using the GR00T platform and GR00T 1.7. The steps are:
- Set up an environment in Isaac Lab-Arena: compose the scene with robot and object assets, define the task and select a teleoperation device. The example places a robot in front of a shelf that moves an apple onto a plate. For whole-body control in this static task, AGILE WBC is used.
- Collect demonstrations with Isaac Teleop: use a supported VR headset and CloudXR streaming to record demonstrations. In the example, 400 trajectories were collected across multiple sessions; demonstrations are saved to an HDF5 file.
- Convert data to LeRobot format: GR00T 1.7 consumes LeRobot datasets, so the recorded HDF5 is converted inside the Arena container into parquet files, MP4 camera recordings and metadata via a YAML config.
- Post-train (fine-tune) GR00T 1.7: run post-training outside the Arena container to tune the visual backbone, projector and diffusion model while keeping the language model frozen. Example training flags include global batch size 12 and max steps 20,000.
- Evaluate the policy: load the checkpoint into a GR00T server and run closed-loop evaluation from the Arena client. Metrics such as success_rate and object_moved_rate print during evaluation; a smoke test uses --num_steps 600, while representative evaluation should run many complete episodes (e.g., --num_episodes 100).
Every stage generates a reusable artifact that feeds the next step, enabling a reproducible end-to-end workflow that can be adapted to other embodiments and tasks, including physical robots.
Ecosystem and partners
The GR00T platform and workflow are already being adopted by a growing set of humanoid robotics partners and institutions:
- Humanoid makers and AI providers: 1X, Agility, ANYBotics, Bellboy Robotics, FieldAI, Lightwheel AI, NEURA Robotics, Nexuni, Noble Machines, Schaeffler, Skild AI, Techman Robot.
- Research institutions experimenting with the unified workflow: Stanford, Carnegie Mellon University (CMU), University of California San Diego (UCSD), ETH Zurich, Allen Institute for AI (AI2).
- Wearable and XR device makers with native Isaac Teleop support: Haptikos, Manus, Noitom, OpenGraph Labs, PICO, Sensing, HTC Vive, Wuji, Xsens.
These partners integrate Isaac Teleop, Sim, Lab and ROS components to accelerate their development pipelines and move AI-enabled humanoid systems toward industrial use.
Availability and resources
The GR00T platform and the GR00T 1.7 model are available now. NVIDIA provides an end-to-end reference workflow and learning content (including “How to Develop and Deploy Humanoid Robots End-to-End with NVIDIA Isaac GR00T” and a GR00T Platform Introduction Video), as well as blog posts and launchable packages for LeRobot integration. Code and model checkpoints are hosted on GitHub and Hugging Face.
By combining an open, modular platform with a pretrained multimodal foundation model, NVIDIA aims to make humanoid policy development more reproducible and to shorten the path from research prototypes to deployed robot behaviors.
