An Analysis of the Impact of Dataset Characteristics on Data-Driven Reinforcement Learning for a Robotic Long-Horizon Task
Ingook Jang, Samyeul Noh, Seonghyun Kim, Donghun Lee · 2023
Data-driven reinforcement learning (RL) is a cost-effective method for training agents without online interaction with the real-world environment. This approach involves collecting and storing data from various sources such as expert demonstrations or random policies, and learning from these datasets without further online interaction with the environment. However, learning an optimal behavioral model from offline data is challenging as it may not cover the entire state-action space. The paper discusses an experimental study analyzing how dataset characteristics impact the performance on a long-horizon robot manipulation task using a robotic arm. The goal of the paper is to provide guidance on strategically organizing datasets for training agents via data-driven RL.