search

minosworld / minos

MINOS: Multimodal Indoor Simulator
MIT License
201 stars 33 forks source link

How to prevent the agent from moving to empty area in SUNCG env? #136

Open gliese581gg opened 5 years ago

gliese581gg commented 5 years ago

Hi!

I'm trying to run RL on SUNCG env but many SUNCG scenes have large empty area.

When the agent random-walks in the early stage of learning,

the agent stuck in these areas very frequently.

This causes the observations consist of only background (green color)

Is there any way to keep the agent stay in meaningful area? (area containing objects, walls, ETC)

angelxuanchang commented 5 years ago

How to have the agent learn to move about intelligently is part of current visual navigation research. In some ways, empty rooms should be "easy" since there is very few things to bump into.

You can consider having your agent build an mental map of the environment or giving it richer signals about the environment. You can use the measurements to get offset and distance to goal (please see https://github.com/minosworld/minos/blob/master/API.md#return-object). In addition to the color sensor you can also add additional sensors (such as depth) that will help your agent.

If you only want to train you agent in environments that are full of objects, then you should filter out scenes that are too empty. Also make sure that you are not using the empty_room and arch_only settings.