Applying a learning agent about ai2thor HOT 1 CLOSED

Hi Ross,

One of the changes that has occurred since that paper was released was the ability to randomize the arrangement of objects and the state of openable items (cabinets, fridges, etc.). When training this way there is an explosion of combinations that makes it infeasible to dump frames.

We still have the ability to detect collisions within our Unity code, we just aren't sending it down in the metadata. When a collision occurs during navigation, we will reset the agent to the spot it previously was and report a failed action back to the Python client. If collisions are still of interest to you, please open a separate issue for that.

We are planning on adding an example to the project showing how to train an agent to perform semantic navigation (follow #13 to be notified when this is added).

from ai2thor.