Embodied Agent Memory via Convergence Divergence Zone
Find Innovative SolutionsGenerate Solutions
Solution Overview
Problem
Current AI systems lack the ability to learn and remember in a human-like manner through real-time embodied interaction with their environment, and existing technologies like Hierarchical Temporal Memory (HTM) fail to provide effective memory for Embodied Agents that can learn from sensorimotor experiences.
Innovation Solution
The implementation of an Experience Memory Store with a Convergence Divergence Zone (CDZ) architecture that allows Embodied Agents to store and retrieve multimodal experiences, enabling real-time learning and memory formation through the use of Self-Organizing Maps (SOMs) and Associative Self-Organizing Maps (ASOMs, which can be populated and modified by both internal experiences and external authoring.
Engineering Contradictions & Design Principles
Engineering Contradiction Analysis
1Measurement precision
If offline learning with prepared data is used, then learning accuracy is improved, but adaptability to real-time embodied interaction deteriorates
Solution Approach 1:
The patent implements dynamic memory registration where sensory inputs are continuously registered in temporal memory structures during real-time interaction. The system transitions from static offline learning to dynamic online learning by maintaining eligibility traces and updating memory representations as the agent interacts with its environment, enabling adaptability while preserving learning accuracy through structured memory consolidation.
Solution Approach 2:
The patent prepares memory structures and eligibility traces in advance before actual learning occurs. By pre-configuring the hierarchical temporal memory architecture with prediction circuits and eligibility trace mechanisms, the system enables rapid real-time learning without sacrificing the structured accuracy of offline learning principles.
2Stability of the object's composition
If Hierarchical Temporal Memory is used, then memory structure is improved, but real-time learning from sensorimotor experience deteriorates
Solution Approach 1:
The patent incorporates prediction circuits that generate predictions about future sensory inputs and compare them with actual inputs in real-time. This feedback mechanism allows the HTM system to continuously learn from sensorimotor discrepancies, maintaining structural stability while enabling real-time learning through error-driven updates to memory representations and eligibility traces.
Solution Approach 2:
The system performs self-directed learning by automatically generating predictions, detecting prediction errors, and updating its own memory structures without external intervention. The eligibility trace mechanism enables the system to self-regulate which memory connections are strengthened based on temporal contiguity and prediction accuracy, maintaining HTM structure while enabling autonomous real-time learning.
3Adaptability or versatility
If multiple modalities are integrated, then learning flexibility is improved, but system complexity deteriorates
Solution Approach 1:
The patent segments the multimodal learning system into distinct sensory modality channels, each with its own eligibility trace and prediction circuits. By dividing the complex multimodal integration task into separate modular processing streams that converge in the hierarchical temporal memory structure, the system achieves learning flexibility across multiple modalities while managing complexity through structured segmentation of processing pathways.
Data Source
AI summary
Computational structures provide Embodied Agents with memory which can be populated in real time from Experience, and/or or authored. Embodied Agents (which may be virtual objects, digital entities or robots) are provided with one or more Experience Memory Stores which influence or direct the behaviour of the Embodied Agents. An Experience Memory Store may include a Convergence Divergence Zone (CDZ), which simulates the ability of human memory to represent external reality in the form of mental imagery or simulation that can be re-experienced during recall. A Memory Database be generated in a simple, authorable way, enabling Experiences to be learned during live operation of the Embodied Agents or authored. Eligibility-Based Learning determines which aspects from streams of multimodal information are stored in the Experience Memory Store.