Reinforcement Learning Data Augmentation for Asymmetric Packing Tasks

The patent applies asymmetry by recognizing that packing tasks have inherent directional asymmetry (source container to destination container) and designs data augmentation methods that preserve this asymmetry. Instead of using symmetric plane conversions that create unrealistic reversed scenarios, the system generates augmented data that maintains the correct directional relationship while varying other parameters like container positions and object characteristics.

The patent inverts the conventional approach by not simply reversing coordinates, but rather by inverting the augmentation logic itself - instead of asking 'how to reverse the data?', it asks 'how to generate varied but realistic data?'. This leads to augmentation techniques that swap non-critical elements (like container identities or background elements) while preserving the critical source-destination relationship.

The patent applies local quality by differentiating between elements that should be varied (local characteristics like container positions, object attributes, background elements) and elements that must remain consistent (global structure like source-destination relationships, task logic). This selective augmentation approach maintains data diversity while preserving learning effectiveness.

The patent utilizes parameter changes by systematically varying specific parameters in the training data (container positions, object properties, environmental conditions) while keeping the fundamental task parameters (source container, destination container, packing action) intact. This allows diverse training scenarios without compromising the core learning objective.