Notes on Deep rl at scale: sorting waste in office building with a fleet of mobile manipulators

This paper describe a system to solve large-scale real-world task: sorting wastes in office buildings with a total training set of 9527 hours of robotic experience.

1. Hybrid data collecting system which contains simulated data and real-world data that is collected through a varity of policy bootstraping approaches.
2. Learning complex tasks by first bootstraping from simulation, and then use of multi-task training to learn simple tasks as a stepping stone.
3. RetinaGAN, a transformer which can transform simulated images to look more realistic.

Not found yet.