Weekly review of Reinforcement Learning papers #10

Every Monday, I present 4 publications from my research area. Let’s discuss them!

Image by the author

[← Previous review][Next review →]

Paper 1: Distribution-conditioned reinforcement learning for general-purpose policies

Nasiriany, S., Pong, V. H., Nair, A., Khazatsky, A., Berseth, G., & Levine, S. (2021). Disco rl: Distribution-conditioned reinforcement learning for general-purpose policies. arXiv preprint arXiv:2104.11707.

Goal-conditioned reinforcement learning consists of incorporating a goal into the policy arguments. For example, consider a robot arm and a cube laying on a table; the task is to control the robot arm to move the cube to a desired position. The RL approaches you know solve this task if the target position is always the same. But what if I now want the cube to be moved to a new position? Do I have to learn a policy for each target position for the cube? By using a goal-conditioned policy, it is possible to learn the move task for any target position. The policy will take the state, and the target state as arguments.

Goal-conditioned policies can allow some generalization, but cannot capture all tasks that might be desired. For example, using the…