Domain Transfer for Reinforcement Learning Agents

PreciBake, is a company based in Munich, New York and Mumbai, developing AI solutions for food-tech and baking industry. Our AI team is continuously working on developing and improving our ML algorithms for tasks such as image classification, object detection and tracking.

Deep reinforcement learning agents have recently shown impressive performance complex tasks such as controlling robots and playing video games. Training reinforcement learning agents for real systems is in practice complicated since in order to learn the agent needs to act with the environment. Most RL agents are also very data hungry and might require millions of training episodes before performing on par with humans on the same task. Training an agent in the real physical environment is therefore most often neither safe or feasible and most RL agents are trained and tested in simulated environments before deployed into the real world.

Additionally there is no guarantee that an agent trained in the simulation will also perform well in the real physical domain. Since the dynamics of the simulation deviates from the dynamics of the physical world agents will most likely underperform or completely fail once deployed in the new domain. In this project we will explore recently developed methods to generalize RL agents trained in one environment, to also perform well when deployed into a new environment. The goal of the project is to develop a method to train RL agents that can perform in a real physical environment even if they were only trained on recorded and simulated data.