NVIDIA Isaac GR00T
NVIDIA's open vision-language-action foundation models for humanoid and other robots, with Apache-2.0 code and open weights for fine-tuning
Overview
Isaac GR00T is NVIDIA's family of open robot foundation models: vision-language-action (VLA) models that take camera images and a language instruction and output robot actions. NVIDIA announced Project GR00T for humanoid robots at GTC in March 2024; the current release, GR00T N1.7 (3B parameters, early access), is published on Hugging Face under the NVIDIA Open Model License, which allows commercial use, while the training and fine-tuning code is Apache-2.0. It ships fine-tuned variants for the LIBERO, DROID and SimplerEnv benchmarks, runs on a single GPU with 16 GB or more (including Jetson AGX Thor on the robot), and is also distributed through Hugging Face LeRobot. NVIDIA uses Agile Robots' humanoid as a reference robot. Also from NVIDIA: NVIDIA Isaac Sim and Isaac Lab, the simulator and robot-learning framework used to train and test GR00T policies, and Nemotron.
The Verdict
Who Should Use NVIDIA Isaac GR00T?
Best For
- Robotics teams that want an open, commercially usable VLA to fine-tune
- Humanoid and manipulator builders on NVIDIA Jetson hardware
- Researchers benchmarking on LIBERO, DROID or SimplerEnv
Not Ideal For
- Teams without NVIDIA GPUs (fine-tuning wants 40 GB or more of VRAM)
- Anyone wanting a hosted robot API instead of running models
- Production users who need a stable, non-early-access release
What's Great
- Open weights with a license that permits commercial use
- Runs on one GPU and on Jetson AGX Thor; TensorRT reaches about 36 Hz on H100
- Ready fine-tunes for common robots and benchmarks
- Fits the NVIDIA stack: Isaac Sim, Isaac Lab, Jetson, plus LeRobot
Watch Out For
- Fine-tuning needs 40 GB+ VRAM; NVIDIA's sim workflow recommends 48 GB+ and 128 GB RAM
- The vision-language backbone (Cosmos-Reason2-2B) is gated on Hugging Face
- Strict toolchain: Linux, CUDA 12.8+, Python 3.12, FFmpeg 4 to 7 only
- N1.7 is labelled early access
Pricing
View all features & details
Releases
- Project GR00T announced at GTC, 18 March 2024
- GR00T N1, N1.5 and N1.6 open releases
- GR00T N1.7 (3B, early access): current release
Hardware
- Inference: 1 GPU with 16 GB+ (RTX 4090, L40, H100, Jetson AGX Thor/Orin, DGX Spark)
- Fine-tuning: 40 GB+ VRAM recommended (H100 or L40)
- About 10.7 Hz on Jetson AGX Thor with TensorRT
Released Fine-Tunes
- LIBERO (Franka arm)
- DROID
- SimplerEnv Bridge (WidowX) and Fractal (Google Robot)
How It Compares
| Feature | NVIDIA GR00T | openpi | Gemini Robotics |
|---|---|---|---|
| Open weights | Yes (NVIDIA Open Model License) | Yes (Apache-2.0) | No |
| Hosted API | No | No (Physical Intelligence runs a separate Pi API beta) | Yes, Robotics-ER 2 in the Gemini API |
| Inference hardware | 16 GB+ GPU | 8 GB+ GPU | Cloud (On-Device model for partners) |
| Fine-tuning hardware | 40 GB+ VRAM | 22.5 GB+ (LoRA), 70 GB+ (full) | See vendor |
| Price | Free | Free | ER 2: free tier, paid from $1 per 1M input tokens |
Is this your tool? Add a badge