An Efficient Image-to-Image Translation HourGlass-based Architecture for Object Pushing Policy Learning
| Type of publication: | Conference paper |
| Citation: | Ewerton_IROS_2021 |
| Publication status: | Accepted |
| Booktitle: | IEEE/RSJ International Conference on Intelligent Robots and Systems |
| Year: | 2021 |
| Abstract: | Humans casually solve push tasks in everyday life but unlocking these capabilities remains a research challenge in robotics. Physical models are often inaccurate or unattainable. State-of-the-art data-driven approaches learn to compensate for these inaccuracies or get rid of the approximated physical models altogether. Nevertheless, approaches like Deep Q-Networks (DQNs) get frequently stuck in local optima in large state-action spaces, and rely on well chosen deep learning architecture and learning paradigms. In this paper, we propose to set DQN learning of pushing policies (where to push and how) as an image-to-image translation problem, and exploit in this regard a more appropriate Hourglass architecture compared to existing methods. We also investigate the use of positional information encoding to learn position dependent policy behaviors, as well as an architecture combining a state-action value predictor indicating which push actions lead to changes in the environment with a reward predictor dedicated to the pushing task at hand (pushing into a box). We demonstrate in simulation experiments with a UR5 robot arm that our overall architecture helps the DQN learn faster and achieve higher performance in a push task involving objects with unknown dynamics. |
| Keywords: | |
| Projects |
Idiap |
| Authors | |
| Added by: | [UNK] |
| Total mark: | 0 |
|
Attachments
|
|
|
|
|
Notes
|
|
|
|
|
