This repository is created to analyze the effect of human intentions in decision making during robot navigation task.
Overview : In this project, our aim is to leverage the power of human intentions in helping a robot find objects in an unknown environment. First, we want to investigate if these intentions will aid our robot in generalizing successfully to new environments, and second, we want to know if this language component will enable us to train with fewer samples than we could with only trajectory data.
The final goal is to compare the impact of using "language" and "not using language” on such a model.
Link to the data we generated is: (https://drive.google.com/drive/folders/1lpgI1vjZrilxdvaHta71ZR-B8J-OuZax?usp=sharing)
Training:
- Make a copy for the data in your drive from the link above.
- Use the script titled 'T5-TextGeneration_ActionPrediction.ipynb' which automatically loads the data from your Drive and creates a dataloader of the tokenized text from the data.
- Train the T-5 base model and you can tweak parameters like batch size, epochs and evaluation metric (currently we use rogue score).
- Test on your evaluation and test set.
- See plots for model performance graphs in the tensorboard.
Give us a like if you find our work useful.
Cheers!