In this research paper, the authors aim to enhance the policy for navigating a labyrinth using various components. They investigate the impact of different elements, such as low-dimensional observations (vec), using images (img), and data augmentation, on the performance of the policy. The experiments show that incorporating these components significantly improves the policy’s efficiency in navigating the labyrinth.
To evaluate the effectiveness of these components, the authors run their policy on a physical setup 50 times and compare it to the previous fastest recorded completion time. They find that their method outperforms the previous record by an average of 0.22 seconds with a 76% success rate.
The authors also provide a visual representation of the ball trajectory during a successful run of their final policy, which helps to demystify complex concepts by providing a tangible example.
In summary, this research focuses on improving the policy for navigating a labyrinth using various components, such as low-dimensional observations, image use, and data augmentation. The results demonstrate that these components significantly improve the policy’s performance, making it faster and more efficient. By providing a visual representation of the ball trajectory during a successful run, the authors help to clarify complex concepts and make the findings more accessible to readers.
Improving Policy Performance through Data Augmentation and Image Patches in Robot Learning
