Domains as Objectives: Multi-Domain Reinforcement Learning with Convex-Coverage Set Learning for Domain Uncertainty Awareness

Domain randomization (DR) is a powerful framework that has allowed the transfer of policies from randomized domain (a.k.a. simulation) to real robots with little to no retraining requirement. However, because the policy has to perform well for many different domain conditions, DR tends to produce sub-optimal policies that can be too conservative on the target real system. This problem is further exacerbated the larger the randomized domain is. To tackle this issue, recent works have proposed to learn universal policies (UP) with domain knowledge such that they can adapt their behavior to each domain when paired with an online system identifier (OSI). However, in most applications, perfect identifications of the target domain can be impossible. In this paper, by drawing similarities between DR as a multi-domain reinforcement learning and multi-objective reinforcement learning (MORL), we propose to learn a UP over the convex coverage set borrowed from the MORL theory. Thanks to this, our method learns a UP that effectively captures different sub-domains of the uncertainty set and can therefore adapt its behavior based on an OSI uncertainty, unlocking the power of stochastic system identification with no retraining requirement. This pseudo-MORL framework also contains previous works in DR and robust reinforcement learning. We conduct simulations on Mujoco tasks and experiments on a real D'Claw robot, revealing the effectiveness of our domain-uncertainty-aware UP for sim-to-real transfer.