NVIDIA DGX Spark Inference

Deployment uniformly uses lerobot-rollout; for usage and parameters, see Command line reference.

Install the environment

• Pytorch

Install the cuda13.0 version of pytorch separately from the official website

pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu130

• Then comment out torch separately in the pyproject.toml file

Then pip install -e .

• Note the path of policy.path in the inference command line; replace it with the actual model path inside spark

Delete the existing dataset starting with rollout (if any)

sudo chmod 666 /dev/ttyACM*

sudo rm -rf /home/apx103/.cache/huggingface/lerobot/Tommymy/rollout_lerobot_my_dataset_shake_hands

ACT

lerobot-rollout  \
  --strategy.type=episodic \
  --robot.type=so101_follower \
  --robot.port=/dev/ttyACM0 \
  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 60, fourcc: "MJPG"}}" \
  --robot.id=my_follower_arm \
  --display_data=false \
  --dataset.repo_id=Tommymy/rollout_lerobot_my_dataset_shake_hands \
  --dataset.single_task="Shake Hands" \
  --dataset.episode_time_s=1000 \
  --policy.path=/home/apx103/Downloads/lerobot_output/shake/ACT/5K/pretrained_model

SmolVLA

• Install the environment

conda activate base
conda create -y -n lerobot-smolvla python=3.10 -y
conda activate lerobot-smolvla
conda install ffmpeg=7.1.1 -c conda-forge -y

pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu130

cd lerobot
pip install -e ".[feetech,smolvla]"

• Inference

lerobot-rollout  \
  --strategy.type=episodic \
  --robot.type=so101_follower \
  --robot.port=/dev/ttyACM0 \
  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 60, fourcc: "MJPG"}}" \
  --robot.id=my_follower_arm \
  --display_data=false \
  --dataset.repo_id=Tommymy/rollout_lerobot_my_dataset_shake_hands \
  --dataset.single_task="Shake Hands" \
  --dataset.episode_time_s=1000 \
  --policy.path=/home/apx103/Downloads/lerobot_output/shake/smolvla/40K/pretrained_model

WALL-OSS

• Install the environment

cd lerobot
pip install -e ".[feetech,wallx]"

• Add the code

lerobot-rollout  \
  --strategy.type=episodic \
  --dataset.repo_id=Tommymy/rollout_lerobot_my_dataset_shake_hands \
  --dataset.single_task="Shake Hands" \
  --policy.path=/home/apx103/Downloads/lerobot_output/shake/wallx/30K/pretrained_model \
  --dataset.push_to_hub=false \
  --robot.type=so101_follower \
  --robot.id=my_follower_arm \
  --robot.port=/dev/ttyACM0 \
  --robot.cameras="{front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 60, fourcc: "MJPG"}}" \
  --display_data=false \
  --dataset.episode_time_s=1000

pi0

• Install the environment

conda create -y -n lerobot-pi python=3.10 -y
conda activate lerobot-pi
conda install *ffmpeg*=7.1.1 *-c* conda-forge -y

cd lerobot
pip3 install torch torchvision --index-url https://download.pytorch.org/whl/cu130
pip install -e ".[feetech,pi]"

• Inference

lerobot-rollout  \
  --strategy.type=episodic \
  --robot.type=so101_follower \
  --robot.port=/dev/ttyACM0 \
  --robot.cameras="{ front: {type: opencv, index_or_path: 0, width: 640, height: 480, fps: 60, fourcc: "MJPG"}}" \
  --robot.id=my_follower_arm \
  --display_data=false \
  --dataset.repo_id=Tommymy/rollout_lerobot_my_dataset_shake_hands \
  --dataset.single_task="Shake Hands" \
  --dataset.episode_time_s=1000 \
  --dataset.push_to_hub=false \
  --policy.path=/home/apx103/Downloads/lerobot_output/shake/pi0/50K/pretrained_model