dVRK Internal Model

Lightweight research code for experimenting with human/internal dynamics models used with the dVRK teleoperation data. This repository contains dataset loaders, modeling code, and example notebooks that explore recovering internal-dynamics parameters and training sequence models.

Key Contents

• data/ — place to keep CSV trajectories from either synthetic LQR data or recorded human teleoperation data.
• dataset/ — dataset utilities and loaders.
• model/ — model implementations.
• notebooks/ — example notebooks demonstrating optimization and training workflows.

Quickstart

1. Create and activate a Python environment.

2. Install the main dependencies (adjust versions to match your system / CUDA):

   pip install torch torchvision torchcodec lightning

Open:

• notebooks/test_optimization.ipynb — experiments to recover LQR/internal-model parameters from synthetic or human trajectories.
• notebooks/test_train.ipynb — training experiments for sequence-prediction models.

Notes & Known Issues

• Some optimization runs can be unstable: for example, optimization becomes unstable when the LQR policy B parameter is below ~5e-3. Further investigation is required.
• Human data are non-stationary (dynamics vary over time), which makes optimizing a single, fixed-parameter internal model useless.
• Model architectures and training hyperparameters are experimental and may need tuning for successful training.