Skill Learning in Robots using Deep RL with Latent Variables
Recent contributions to the field of Reinforcement Learning (RL) have been concerned with the learning of skills, diverse behaviors of the agent that can be chosen by setting the values of latent variables which the policy is conditioned on. While some approaches omit the task reward during training, making it simply learn diverse behaviors, other approaches seek to train the agent to maximize the task reward in diverse manners.
In my university seminar Skills in Robot Reinforcement Learning I read about three recent approaches to diverse skill learning, in papers by Eysenbach et al. (2018), Kumar et al. (2020) and Osa et al. (2021). The seminar report goes into detail on the differences and similarities in the foundations and objectives. Finally, the experimental setups are discussed as well as which skills are manifested in the agents’ behaviors.
Aside from writing the report, which can be read and downloaded at the bottom of this page, I held a presentation that gives an overview over skill learning in robots using reinforcement learning.