FINGR: Learning Dexterous Hand Control
for Real-World Rubik's Cube Solving

Tested on Ubuntu 22.04 with Python 3.12, PyTorch 2.7.1 (CUDA 11.8), an Intel i9-13900K, 32 GB RAM, and one NVIDIA RTX 4090 (24 GB). Estimated training time is 45-50 minutes for 1,000 epochs on the full dataset.

Clone the repository and enter its directory.

git clone https://github.com/EmptyBlueBox/FINGR.git
cd FINGR

Visualize a training trajectory, downloading the dataset automatically on first use.

uv run -m fingr.visualize_training_traj

Train the FINGR policy on the demonstration dataset.

uv run -m fingr.train_policy