Learning to Map for Active Semantic Goal Navigation

Authors: Georgios Georgakis, Bernadette Bucher, Karl Schmeckpeper, Siddharth Singh, Kostas Daniilidis

ICLR 2022 | Conference PDF | Archive PDF | Plain Text | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We perform experiments on the Matterport3D (MP3D) (Chang et al., 2017) dataset using the Habitat (Savva et al., 2019) simulator.
Researcher Affiliation Collaboration Georgios Georgakis*1, Bernadette Bucher*1, Karl Schmeckpeper1, Siddharth Singh2, Kostas Daniilidis1 1University of Pennsylvania, 2Amazon
Pseudocode Yes Algorithm 1: L2M for Object Nav
Open Source Code Yes Trained models and code can be found here: https://github.com/ggeorgak11/L2M.
Open Datasets Yes We perform experiments on the Matterport3D (MP3D) (Chang et al., 2017) dataset using the Habitat (Savva et al., 2019) simulator.
Dataset Splits Yes We use the standard train/val split as the test set is held-out for the online Habitat challenge, which contains 56 scenes for training and 11 for validation.
Hardware Specification Yes We executed training and testing on our internal cluster on RTX 2080 Ti GPUs.
Software Dependencies No The paper mentions 'Py Torch Paszke et al. (2017)' and 'Habitat (Savva et al., 2019) simulator' with citations, but does not provide explicit version numbers for these or any other software dependencies.
Experiment Setup Yes The models are trained in the Py Torch Paszke et al. (2017) framework with Adam optimizer and a learning rate of 0.0002. All experiments are conducted with an ensemble size N = 4. For the semantic map prediction we receive RGB and depth observations of size 256 256 and define crop and global map dimensions as h = w = 64, H = W = 384 respectively.