Robot Learning Evaluation Using Shared Embeddings and Local Caching

The patent segments the learning process into two distinct parts: (1) robots immediately store and use learned information locally without updating models, and (2) the central server periodically aggregates and updates models in batches. This segmentation allows robots to achieve fast local adaptation while the server handles computationally intensive model updates, resolving the contradiction between learning speed and computational time.

The patent introduces a central server as an intermediary between robots and the model update process. The server aggregates learned information from multiple robots, performs batch model updates, and distributes updated models back to robots. This intermediary approach eliminates the need for each robot to perform expensive model updates individually, significantly reducing computational time and energy consumption while maintaining learning accuracy.

The patent implements preliminary action by having robots immediately store learned information in their local databases upon encountering new objects, allowing them to adapt and recognize new objects right away without waiting for model retraining. The computationally intensive model updates are performed preliminarily and periodically by the central server, ensuring robots always have access to improved models while maintaining high adaptation speed.

The patent uses copying by distributing updated machine learning models from the central server to multiple robots. Instead of each robot performing independent retraining, the server creates updated model copies and distributes them fleet-wide. This copying approach maintains consistent recognition capability across all robots while dramatically improving adaptation speed and productivity.

The patent merges the learning processes of multiple robots by having them contribute their learned information to a central server. The server aggregates embeddings and learned objects from all robots, performs unified model updates, and distributes improvements fleet-wide. This merging approach maintains individual learning speed through local caching while dramatically improving fleet-wide learning efficiency through shared knowledge.

The patent implements universality by creating a centralized model that serves all robots in the fleet. The central server develops universal machine learning models that incorporate knowledge from all robots and distribute them fleet-wide. This universal approach allows each robot to benefit from the collective learning of the entire fleet, maximizing learning efficiency while preserving individual adaptation capabilities.