GR00T Lerobot on DGX Spark playbook

The Dockerfile in this repo will build an image ready to train/inference gr00t via lerobot, on NVIDIA DGX Spark 🚀

lerobot[gr00t]

This dockerfile will allow you to use the lerobot[gr00t] model. Read more about it on Huggingface 🤗

• Huggingface documenation
• NVIDIA blog post

Configuration

• Arch: ARM
• GPU: GB10
• CUDA: 13.1
• torch: 2.9
• flash-attn: 2.8.4

Compiled binares for torch via https://pypi.jetson-ai-lab.io/sbsa/cu130

Instructions

On your DGX Spark, run:

git clone https://github.com/liorbenhorin/gr00t-lerobot-dgx-spark-docker
cd gr00t-lerobot-dgx-spark-docker
./start.sh

Note: First time will take ~45m to build due to flash-attn, so grab a huge coffee ☕