Learning Temporally Consistent 3D Robot Mapping

Weining Ren$^*$, Haoran Chen$^*$. Supervised by Dr. Lukas Schmid and Prof. Jen Jen Chung.

Report Video

This project was undertaken at Autonomous Systems Lab, ETH Zürich.

In this project, we developed a 3D GRU fusion layer based on NeuralBlox and proposed an observation memory function with free space encoding to capture temporal changes when mapping in a dynamic environment.

Main Contributions

  • Proposed a 3D GRU fusion layer based on NeuralBlox to directly capture temporal changes in latent space when mapping in a dynamic environment;
  • Proposed a temporal attention mechanism using an observation space memory function;
  • Designed a novel training strategy to use free space encoding.

Overview of our pipeline

Overview of our pipeline

Training strategy

Training strategy

Qualitative results

Qualitative results

Please find more details in our report.