BROWSE
Detail
Conference Paper
An Algorithm Design of Domain Adaptation for Sim-To-Real Robot Learning Via Meta-Reinforcement Learning
Cited - time in 
Share 
Share
- Authors
- Samyeul Noh, Junhee Park
- Issue Date
- 2021-07
- Citation
- International Conference on Ubiquitous Robots (UR) 2021, pp.451-452
- Publisher
- IEEE
- Language
- English
- Type
- Conference Paper
- Abstract
- In this paper, we present an algorithm design of domain adaptation for sim-to-real robot learning by means of meta-reinforcement learning (meta-RL). The goal of domain adaptation for sim-to-real robot learning is to adapt a model learned in simulation to real robots quickly and efficiently. To this end, the presented algorithm is designed by means of meta-RL and composed of two main procedures: meta-initialization in simulation and fast adaptation to real robots. In the first procedure, a model is meta-trained using a variety of learning tasks in simulated robotic environments with the goal of finding a good initialization for the model parameters so that new tasks can be learned using only a few gradient updates. In the second procedure, the meta-trained model is adapted to tasks for real robotic environments using only a small number of trials and gradient steps. With the assumption that learning tasks should be distributed from the same task distribution, the same robot configuration is used both in simulated and real robotic environments.
- KSP Keywords
- Algorithm design, Fast Adaptation, Gradient steps, Model parameter, Reinforcement learning(RL), Robot Learning, Task distribution, domain adaptation