Spatial Embedding Neural Network for Agent Navigation

Resolve Bottlenecks,
Find Innovative Solutions
Generate Solutions

Solution Overview

Problem

Current machine learning models face challenges in efficiently processing data to generate spatial embeddings that enable an agent to navigate through complex and unfamiliar environments, particularly in real-world scenarios, as they often require extensive training and computational resources.

Innovation Solution

A spatial embedding neural network system that processes data characterizing an agent's motion to generate predictive spatial embeddings, which are used by an action selection neural network to select actions for navigating through environments, allowing the agent to find direct routes and exploit rich spatial information for efficient navigation.

Engineering Contradictions & Design Principles

VSEngineering Contradiction Analysis

1Productivity

If current machine learning models are used to process data and generate spatial embeddings for navigation, then the agent can navigate through environments, but extensive training and computational resources are required

Engineering Contradiction:
Improvenavigation efficiencyVSAvoidtraining time
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The system segments the navigation problem into two distinct components: a spatial embedding network that processes motion data to generate spatial representations, and a separate action selection network that uses these embeddings to choose actions. This segmentation allows each network to be optimized for its specific function, improving overall navigation efficiency while reducing the training burden compared to monolithic models

Inventive Principle:
Principle #1Segmentation

2Productivity

If current machine learning models are used to process data and generate spatial embeddings for navigation, then the agent can navigate through environments, but extensive computational resources are required

Engineering Contradiction:
Improvenavigation efficiencyVSAvoidcomputational resources
Core Design Contradiction:
ProductivityVSUse of energy by moving object

Solution Approach 1:

The system extracts and processes only the essential motion data (position, velocity, acceleration) through the spatial embedding network, separating this from other environmental observations. By taking out only the critical spatial-temporal features needed for navigation and processing them through a dedicated embedding network, the system reduces overall computational resource requirements while maintaining navigation effectiveness

Inventive Principle:
Principle #2Taking out (Extraction)

3Productivity

If the agent uses traditional navigation methods, then it can explore environments, but it cannot find direct routes and must visit many areas sequentially

Engineering Contradiction:
Improvenavigation speedVSAvoidtime to reach goal
Core Design Contradiction:
ProductivityVSLoss of time

Solution Approach 1:

The spatial embedding network performs preliminary processing of motion data to generate compressed spatial representations before the action selection network makes decisions. This preliminary action of creating informative spatial embeddings allows the agent to understand the spatial relationships and plan direct routes to goals, rather than requiring sequential exploration of all areas

Inventive Principle:
Principle #10Preliminary action

4Device complexity

If the spatial embedding network processes only motion data, then it reduces computational load, but it must effectively integrate temporal information to capture spatial relationships

Engineering Contradiction:
Improveprocessing complexityVSAvoidspatial relationship information
Core Design Contradiction:
Device complexityVSLoss of information

Solution Approach 1:

The spatial embedding network transforms the input motion data parameters (position, velocity, acceleration) into a different parameter space through learned embeddings. This parameter transformation allows the network to effectively integrate temporal information and capture spatial relationships in a compressed representation, maintaining information quality while reducing processing complexity

Inventive Principle:
Principle #35Parameter changes

Data Source

PatentEP4107669B1Generating spatial embeddings by integrating agent motion and optimizing a predictive objective
Publication Date: 2024.12.04 DEEPMIND TECH LTD
  • EP4107669B1 patent drawingFigure 1
  • EP4107669B1 patent drawingFigure 2
  • EP4107669B1 patent drawingFigure 3

AI summary

Methods, systems, and apparatus, including computer programs encoded on a computer storage medium, for training a spatial embedding neural network that is configured to process data characterizing motion of an agent that is interacting with an environment to generate spatial embeddings. In one aspect, a method comprises: processing data characterizing the motion of the agent in the environment at the current time step using a spatial embedding neural network to generate a current spatial embedding for the current time step; determining a predicted score and a target score for each of a plurality of slots in an external memory, wherein each slot stores: (i) a representation of an observation characterizing a state of the environment, and (ii) a spatial embedding; and determining an update to values of the set of spatial embedding neural network parameters based on an error between the predicted scores and the target scores.