Robot Embedding Cache Sharing for Fast Fleet Object Learning

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current robot systems face challenges in efficiently updating machine learning models, as training or retraining these models is computationally intensive and time-consuming, often requiring large datasets, which can be impractical for individual robots to handle in real-time, especially when encountering new objects.

Innovation Solution

A method where robots can instantly learn new information by storing representations in a local cache and share these representations with other robots, allowing for rapid dissemination of new knowledge without the need for extensive retraining, while a central server updates machine learning models periodically based on aggregated learning from the fleet.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Reliability

If robots retrain machine learning models to learn new information, then the robot's recognition ability is improved, but the time and computational resources required increase significantly

Engineering Contradiction:
Improverecognition abilityVSAvoidtraining time
Core Design Contradiction:
ReliabilityVSLoss of time

Solution Approach 1:

The patent divides the machine learning system into two independent components: a pre-trained machine learning model that remains unchanged, and a local cache that stores learned information about new objects. This segmentation allows the robot to quickly acquire new knowledge by storing representations in the cache without triggering time-consuming model retraining, while still improving recognition ability for encountered objects

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent creates simplified copies of object representations (embeddings) and stores them in a local cache. These copied representations allow the robot to rapidly learn about new objects by storing and retrieving embedding vectors, avoiding the need to copy entire machine learning models for each new object encounter

Inventive Principle:
Principle #26Copying

2Reliability

If robots retrain machine learning models to learn new information, then the robot's recognition ability is improved, but the computational resources required increase significantly

Engineering Contradiction:
Improverecognition abilityVSAvoidcomputational resources
Core Design Contradiction:
ReliabilityVSUse of energy by moving object

Solution Approach 1:

The system separates the computationally intensive machine learning model from the lightweight local cache. The model remains fixed and pre-trained, while new information is stored as simple embedding representations in the cache. This segmentation enables the robot to improve recognition ability through cache lookups that consume minimal computational resources compared to full model retraining

Inventive Principle:
Principle #1Segmentation

Solution Approach 2:

The patent uses inexpensive embedding representations stored in local cache as temporary, disposable knowledge units. These lightweight representations can be quickly created, stored, and discarded without significant computational cost, providing an energy-efficient alternative to maintaining and retraining complex machine learning models

Inventive Principle:
Principle #27Cheap short-living objects (Disposable)

3Adaptability or versatility

If individual robots learn new information independently, then each robot adapts to its specific environment, but the fleet-wide learning efficiency decreases

Engineering Contradiction:
Improveenvironmental adaptationVSAvoidlearning efficiency
Core Design Contradiction:
Adaptability or versatilityVSProductivity

Solution Approach 1:

The patent merges individual robot learning experiences by collecting embeddings from multiple robots and consolidating them into a shared fleet-wide cache. This combining approach allows the fleet to collectively learn from all robots' encounters, improving overall learning efficiency while still enabling each robot to adapt to its specific environment through access to the shared cache

Inventive Principle:
Principle #5Merging (Combining)

Solution Approach 2:

The local cache serves multiple functions: it stores individually learned object representations for immediate robot use, collects embeddings from fleet-wide experiences, and provides rapid lookup capability for recognition tasks. This multi-functionality enables both environmental adaptation and fleet-wide learning efficiency to coexist within a single system

Inventive Principle:
Principle #6Universality (Multi-functionality)

Data Source

PatentEP3713719B1Sharing learned information among robots
Publication Date: 2024.08.28 GOOGLE LLC
  • EP3713719B1 patent drawingFigure 1
  • EP3713719B1 patent drawingFigure 2
  • EP3713719B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on computer storage media, for sharing learned information among robots. In some implementations, a robot obtains sensor data indicating characteristics of an object. The robot determines a classification for the object and generates an embedding for the object using a machine learning model stored by the robot. The robot stores the generated embedding and data indicating the classification for the object. The robot sends the generated embedding and the data indicating the classification to a server system. The robot receives, from the server system, an embedding generated by a second robot and a corresponding classification. The robot stores the received embedding and the corresponding classification in the local cache of the robot. The robot may then use the information in the cache to identify objects.