Open access
Author
Date
2023Type
- Master Thesis
ETH Bibliography
yes
Altmetrics
Abstract
Indoor navigation is an important skill for embodied agents to possess. This skill enhances the generality of intelligent agents by allowing more opportunities for task chaining. For example, many day-to-day tasks in a home require indoor navigation. The task “do the laundry” can be decomposed into “go to the bathroom, pick up the laundry basket, go to the laundry room, etc.”. Navigation is a necessary glue between other tasks. Our project, BeSAFE, is a framework that leverages an existing simulation platform for embodied AI research, Habitat, and extends it with a real-time photorealistic simulator using Unreal Engine. In addition to that, BeSAFE also provides an apartment environment that can be used for training and benchmarking. We currently only support point goal tasks. We used our framework to benchmark the safety of agents in indoor navigation tasks by defining a metric that computes the distance to the closest obstacle at every location. We employ this metric to assess the relative safety of each step the agent takes in our environment. Finally, we trained a PPO baseline agent in our environment and evaluated our metric on it. Show more
Permanent link
https://doi.org/10.3929/ethz-b-000651649Publication status
publishedPublisher
ETH ZurichOrganisational unit
03766 - Pollefeys, Marc / Pollefeys, Marc
More
Show all metadata
ETH Bibliography
yes
Altmetrics