Auxiliary Tasks for Navigation in Unknown Environments (class project)
We used a simulator
to train an agent that navigates itself in a photo-realistic 3-D indoor environment.
The agent only observes first-person view image.
We trained the agent with several different auxiliary tasks such as depth prediction,
and compare them to see what auxiliary tasks can help navigation.
The gif on the left shows a PPO agent trained with reward (distance to the goal) and auxiliary depth-prediction.