Pose Estimation Lunar Robot (Dataset for camera pose estimation research using computer simulated images from rovers on the lunar surface)

Overview

The goal: using simulation data to train neural networks to estimate the pose of a rover's camera with respect to a known target object

The mission context: A simulated lunar surface, with lunar landers and lunar rovers. To accomplish their ressource extraction mission, the rovers must dig, transport and deliver regolith to a processing plant. For each of these tasks, a central need is for rovers to accurately estimate the relative pose both between themselves and with the landers.

The dataset:

• tf.data dataset ready for training (RGBD images and ground truth pose labels)
• five different scenarios for relative pose estimation, some easier, some harder!

The code:

Utilities for manipulating the dataset and calculating training metrics + example jupyter notebooks for data exploration and model training + more details on the dataset are available on github