DROID benchmark

Real-world manipulation with the DROID dataset

GR00T N1.6-DROID is fine-tuned on the DROID dataset, a large-scale dataset for real-world robotic manipulation. This checkpoint can be deployed on DROID-compatible robot hardware.

Checkpoint: nvidia/GR00T-N1.6-DROID

Deployment

Setup inference server

1. Start policy server

   On a machine with a sufficiently powerful GPU, start the policy server from the root folder of the repository:

   uv run python gr00t/eval/run_gr00t_server.py \
       --embodiment-tag OXE_DROID \
       --use_sim_policy_wrapper \
       --model-path=nvidia/GR00T-N1.6-DROID
   

Setup robot control

1. Install DROID package

   Install the DROID package on the robot control laptop/workstation following the official instructions.

2. Install GR00T dependencies

   Install additional dependencies for the GR00T control script in the DROID environment:

   pip install tyro moviepy==1.0.3 pydantic numpy==1.26.4
   
3. Configure cameras

   Enter the camera IDs for your ZED cameras in examples/DROID/main_gr00t.py.

4. Start control script

   Run the control script with your chosen external camera:

   python examples/DROID/main_gr00t.py --external_camera="left"  # or "right"
   

Hardware requirements

For the DROID setup, you’ll need:

• GPU machine for running the policy server (RTX 4090, H100, or similar)
• DROID-compatible robot hardware
• ZED cameras configured as per DROID specifications
• Network connection between control laptop and GPU server

Dataset information

The DROID dataset provides:

• Large-scale real-world manipulation demonstrations
• Multi-camera observations
• Diverse manipulation tasks
• Standardized data format compatible with GR00T’s LeRobot schema