1. This paper proposes an image-to-image translation HourGlass-based architecture for object pushing policy learning.
2. The architecture combines a predictor of which pushes lead to changes in the environment with a state-action value predictor dedicated to the pushing task.
3. Experiments with a UR5 robot arm demonstrate that the overall architecture helps the DQN learn faster and achieve higher performance in a pushing task involving objects with unknown dynamics.
The article is written by researchers from Idiap Research Institute, Switzerland, and is funded by Swiss National Science Foundation through the HEAP project (Human-Guided Learning and Benchmarking of Robotic Heap Sorting, ERA-net CHIST-ERA). The authors present an efficient image-to-image translation HourGlass-based architecture for object pushing policy learning, which combines a predictor of which pushes lead to changes in the environment with a state-action value predictor dedicated to the pushing task. Experiments with a UR5 robot arm demonstrate that this overall architecture helps the DQN learn faster and achieve higher performance in a pushing task involving objects with unknown dynamics.
The article is well written and provides sufficient evidence for its claims. It presents both sides of the argument equally and does not contain any promotional content or partiality towards any particular point of view. The authors have also provided code on GitHub for further research into their proposed architecture, which adds to its trustworthiness and reliability. Furthermore, they have noted possible risks associated with their approach such as local optima in large state action spaces and reliance on well chosen deep learning architectures and learning paradigms. All these points make this article trustworthy and reliable.