Self-Supervised Representation Learning for Vision-Based Reinforcement Learning

The patent applies preliminary action by pre-augmenting training images with various visual distractions (clouds, shadows, light variations) before the model sees them. This prepares the model in advance for real-world conditions, so when actual visual distractions appear during deployment, the model remains robust. The data augmentation process adds distracting elements to training images beforehand, allowing the model to learn invariant features despite the added complexity in training.

The patent uses copying by creating augmented copies of original training images with added visual distractions. Instead of modifying the original images or creating entirely new datasets, the system generates multiple distorted versions (copies) of each training image with different distraction levels and types. These copied images with distractions serve as additional training samples, enabling the model to learn robustness without requiring fundamentally new training approaches.

The patent applies parameter changes by systematically varying parameters of visual distractions in training data, such as opacity levels, spatial positions, sizes, and types of distractions (clouds, shadows, light). By changing these parameters across different training samples, the model learns to generalize across diverse real-world conditions. The augmentation process modifies image parameters to simulate different environmental conditions, enabling the model to adapt to varying real-world scenarios.