Researchers led by Madeleine Nouri have compared hierarchical reinforcement learning, which chains separately trained skills, with single flat policies for navigation and robot manipulation, without human demonstrations, according to a paper on arXiv. The paper has been accepted at the Australasian Conference on Robotics and Automation.
In navigation, two skills were reused in grid layouts the agent had not seen, and the chained approach clearly beat flat baselines with fewer training interactions. In manipulation, reaching, aligning a grasp and lifting were trained one after another. The chain needed fewer training steps than a flat policy and varied less between runs.



