> ## Documentation Index
> Fetch the complete documentation index at: https://mintlify.com/NVIDIA/Isaac-GR00T/llms.txt
> Use this file to discover all available pages before exploring further.
# Quick start
> Run inference with GR00T in minutes
Get started quickly by downloading a pre-trained checkpoint and running inference on the GR1 embodiment.
## Prerequisites
* GR00T installed (see [Installation](/getting-started/installation))
* CUDA-enabled GPU
* Internet connection to download model checkpoint
## Start the policy server
GR00T uses a server-client architecture for inference. Start the policy server with a pre-trained checkpoint:
```bash theme={null}
uv run python gr00t/eval/run_gr00t_server.py \
--embodiment-tag GR1 \
--model-path nvidia/GR00T-N1.6-3B
```
The server will:
* Download the `nvidia/GR00T-N1.6-3B` checkpoint from Hugging Face (if not already cached)
* Load the model on GPU
* Start listening on `localhost:5555` for inference requests
On first run, the model checkpoint (approximately 6GB) will be downloaded from Hugging Face. Subsequent runs will use the cached version.
## Run standalone inference
Run inference on a sample dataset to verify everything works:
```bash theme={null}
uv run python scripts/deployment/standalone_inference_script.py \
--model-path nvidia/GR00T-N1.6-3B \
--dataset-path demo_data/gr1.PickNPlace \
--embodiment-tag GR1 \
--traj-ids 0 1 2 \
--inference-mode pytorch \
--action-horizon 8
```
This script:
* Loads the GR00T N1.6 3B parameter model
* Runs inference on trajectories 0, 1, and 2 from the demo dataset
* Uses PyTorch mode (no TensorRT acceleration)
* Predicts 8 future action steps per inference call
### Expected output
You should see timing information similar to:
```
Data Processing: 2 ms
Backbone: 18-25 ms
Action Head: 16-38 ms
End-to-End: 37-58 ms
Frequency: 17-27 Hz
```
Actual timing depends on your GPU. See the [performance table](#inference-performance) below.
## Inference performance
GR00T N1.6 3B inference timing (4 denoising steps, single view):
| Device | Mode | Data Processing | Backbone | Action Head | E2E | Frequency |
| -------- | ------------- | --------------- | -------- | ----------- | ------ | --------- |
| RTX 5090 | torch.compile | 2 ms | 18 ms | 16 ms | 37 ms | 27.3 Hz |
| H100 | torch.compile | 4 ms | 23 ms | 11 ms | 38 ms | 26.3 Hz |
| RTX 4090 | torch.compile | 2 ms | 25 ms | 17 ms | 44 ms | 22.8 Hz |
| Thor | torch.compile | 5 ms | 39 ms | 61 ms | 105 ms | 9.5 Hz |
For 2x faster inference, see [TensorRT optimization](/deployment/tensorrt-optimization).
## Run zero-shot evaluation
For a more complete example with simulation environments, try the RoboCasa GR1 tabletop tasks:
```bash theme={null}
uv run python gr00t/eval/run_gr00t_server.py \
--embodiment-tag GR1 \
--model-path nvidia/GR00T-N1.6-3B
```
In a separate terminal, navigate to the RoboCasa example:
```bash theme={null}
cd examples/robocasa-gr1-tabletop-tasks
```
Follow the instructions in `examples/robocasa-gr1-tabletop-tasks/README.md` for environment setup and evaluation.
## Using the policy API
To integrate GR00T into your own environment, use the Policy API:
```python theme={null}
from gr00t.policy.server_client import PolicyClient
# Connect to the policy server
policy = PolicyClient(host="localhost", port=5555)
# Verify connection
if not policy.ping():
raise RuntimeError("Cannot connect to policy server!")
# Create observation in the expected format
observation = {
"video": {
"camera_name": video_array, # Shape: (B, T, H, W, 3), dtype: uint8
},
"state": {
"state_name": state_array, # Shape: (B, T, D), dtype: float32
},
"language": {
"task": [["pick up the apple"]], # Shape: (B, 1)
}
}
# Get action from policy
action, info = policy.get_action(observation)
# Execute action in your environment
env.step(action)
```
For detailed information on observation/action formats and policy integration, see the [Policy API guide](/core-concepts/policy-api).
## Available pre-trained models
### Base models
| Model | Parameters | Use Case | Checkpoint |
| ---------- | ---------- | ---------- | ------------------------------------------------------------------- |
| GR00T N1.6 | 3B | Finetuning | [nvidia/GR00T-N1.6-3B](https://huggingface.co/nvidia/GR00T-N1.6-3B) |
| GR00T N1.5 | 3B | Finetuning | [nvidia/GR00T-N1.5-3B](https://huggingface.co/nvidia/GR00T-N1.5-3B) |
### Finetuned models
| Model | Embodiment | Description | Checkpoint |
| ----------------------------- | -------------- | ---------------------------------- | --------------------------------------------------------------------------------------------------- |
| GR00T-N1.6-bridge | WidowX | Bridge dataset manipulation tasks | [nvidia/GR00T-N1.6-bridge](https://huggingface.co/nvidia/GR00T-N1.6-bridge) |
| GR00T-N1.6-fractal | Google Robot | Fractal dataset manipulation tasks | [nvidia/GR00T-N1.6-fractal](https://huggingface.co/nvidia/GR00T-N1.6-fractal) |
| GR00T-N1.6-BEHAVIOR1k | Galaxea R1 Pro | BEHAVIOR-1K loco-manipulation | [nvidia/GR00T-N1.6-BEHAVIOR1k](https://huggingface.co/nvidia/GR00T-N1.6-BEHAVIOR1k) |
| GR00T-N1.6-G1-PnPAppleToPlate | Unitree G1 | Loco-manipulation pick-and-place | [nvidia/GR00T-N1.6-G1-PnPAppleToPlate](https://huggingface.co/nvidia/GR00T-N1.6-G1-PnPAppleToPlate) |
| GR00T-N1.6-DROID | DROID | DROID manipulation tasks | [nvidia/GR00T-N1.6-DROID](https://huggingface.co/nvidia/GR00T-N1.6-DROID) |
## Next steps
Prepare your robot data for training
Finetune GR00T on your custom data
Learn the Policy API for integration
Evaluate your trained models