yunkao/SonoGym_lerobot_dataset

SonoGym Lerobot Dataset

This repository contains the expert datasets collected for the paper SonoGym: High Performance Simulation for Challenging Surgical Tasks with Robotic Ultrasound. These datasets are specifically designed to facilitate the training of imitation learning policies for complex robotic surgical tasks, particularly those involving robotic ultrasound.

The dataset is a part of the broader SonoGym project, a scalable simulation platform built on NVIDIA IsaacLab that enables parallel simulation for challenging robotic ultrasound tasks.

Project Page: https://sonogym.github.io/ Code/GitHub Repository: https://github.com/SonoGym/SonoGym

Dataset Details

The SonoGym_lerobot_dataset is an expert dataset designed for training imitation learning policies, particularly with the lerobot library. It contains trajectories collected under various settings, covering different ultrasound guidance and surgical tasks with single or multiple patient models, and both physics-based (model-based) and generative modeling (learning-based) ultrasound simulations.

The dataset is structured into several components, each representing a specific task and simulation configuration:

• Isaac-robot-US-guidance-v0-single: Ultrasound guidance task, single patient, using model-based US simulation.
• Isaac-robot-US-guidance-v0-single-net: Ultrasound guidance task, single patient, using learning-based US simulation.
• Isaac-robot-US-guidance-5-models-v0: Ultrasound guidance task, single patient, using 4 learning-based US simulation networks.
• Isaac-robot-US-guided-surgery-v0-single-new: Ultrasound-guided surgery task, single patient, using model-based US simulation.
• Isaac-robot-US-guided-surgery-v0-single-net-new: Ultrasound-guided surgery task, single patient, using learning-based US simulation.
• Isaac-robot-US-guided-surgery-v0-5-net: Ultrasound-guided surgery task, single patient, using 4 learning-based US simulation networks.
• Isaac-robot-US-guided-surgery-v0-5: Ultrasound-guided surgery task, 5 patients, using model-based US simulation.

Sample Usage

This dataset is primarily used for training imitation learning (IL) policies. After downloading the dataset, you can train ACT or Diffusion Policy models using the lerobot library.

Example training command for ACT or Diffusion Policy with lerobot:

First, ensure you have the lerobot repository cloned and its dependencies installed.

# Example for training Diffusion Policy or ACT
python /path-to-lerobot/lerobot/scripts/train.py --config_path=workflows/lerobot/train_surgery_{method}_cfg.json

Replace {method} with either diffusion or act. Make sure to update the "dataset" root path in your configuration file (train_surgery_{method}_cfg.json) to point to the local path of your downloaded dataset. For example:

{
  "dataset": "/path-to-your-local-repo/SonoGym/lerobot-dataset/Isaac-robot-US-guidance-v0-single-net",
  // ... other config parameters
}

For more detailed usage instructions, including teleoperation, reinforcement learning training, and environment settings, please refer to the main SonoGym GitHub repository.