Habitat Supervised Learning Baselines

Detailed report regarding this work can be accessed here: Supervised Learning Benchmarks for Point Goal Navigation in Photo-realistic indoor cluttered

Behavior Cloning

Installation

Install Habitat-api and Habitat-sim

Simlink to create the following structure

.
+- vln-instruction-generation
   +- config
   +- data
   |  +- datasets
   |  |  +- pointnav
   |  |     +- mp3d
   |  |        +- v1
   |  |           +- test
   |  |           +- train
   |  |           +- val
   |  +- scene_datasets
   |     +- mp3d
   |        +- 1LXtFkjw3qL
   |        +- 1pXnuDYAj8r
   |        +- ...
   |        +- zsNo4HB9uLZ
   +- semantic-path.py

Data Pre-processing

Run the following script to make data batches. Change source path and destination path in make_batches.py

python make_batches.py

DataLoader for training

expert.py #provides various functions to get optimal shortest path for any datapoint

Train

python train.py

Hyperparameters

Preliminary Results

Future Work

• Implement imitation learning pipeline
• Evaluating approach in real-time simulation in habitat-api and tracking various metrics including SPL