Project Name: General Structural Artificial Intelligence Framework Name: AI-C Engine Theoretical name: Structure based Intelligence
A core insight
Humans do not first "recognize" cats and then "see" them.
Humans first draw cats out of the background before they can 'see' them. Recognition is the processing of segmentation results. Splitting is the starting point.
What structured AI does is to segment the foreground from the background and record each segmentation.
📖 What is this
AI-C Engine is a structured cognitive framework based on the underlying principle of "splitting the world". It does not rely on pre training data, starts from an empty state, and autonomously forms a traceable, interpretable, and evolving cognitive structure through segmentation and accumulation during continuous operation.
Core concept: Intelligence is not the correct representation of the world, but the ability to continuously extract usable structures from the background.
Splitting refers to "drawing" an object from the background.
It is not 'recognition', not 'classification', not 'matching'. It is pulling something out of chaos alone.
The simplest example
You walk into a room.Before and after segmentation The room, table, chair, book, lighting... are all indistinguishable entities. "Books" are drawn out from the background
Three different levels of segmentation
- Physical segmentation
Cut out the background and foreground All objects in the room, a chair Everything on the table, a cup Everyone in the field of vision, a person wearing a red shirt 2. Abstract segmentation
Cut out the background and foreground The happiest moment of the day All chat records' he mentioned things he was afraid of ' Continuous exercise data shows a sudden change in running speed 3. Self segmentation
Cut out the background and foreground All the things I am doing All emotions' the anxiety I feel now ' All memories of 'how did I become like this now'
Mathematical AI vs Structural AI: The Difference in Processing Apples
Step by step mathematical formula AI structural AI (segmentation) Seeing the scene, identify the "apple" and cut it out of the background Extracting an object from chaos based on the features in the training data of essential matching The result is' This is an apple 'and' This apple came out of the background ' Record no 'I sliced an apple from the background in step X' Traceability does not know when, from what background, or how it was divided Examples of Segmentation in Daily Life
Scene segmentation explanation Listen to someone speak and cut out 'he's angry' from a string of sounds. You didn't recognize the word 'angry', but drew the structure of 'angry' from the tone Look at a photo and cut out 'him' from all the details. You didn't 'recognize' him, but 'picked' him out from the background Recalling a day, find 'that thing' from all your memories and draw 'that thing' from your entire day's experiences To make a decision and cut out the 'path' from all choices, you are not 'calculating' the optimal solution, but drawing a path from many possibilities
Structural Artificial Intelligence
Structure-based Artificial Intelligence
1、 One sentence definition
Structural artificial intelligence is an intelligent form that uses "splitting" as an atomic operation and "structural growth" as the core path. It does not approximate correctness through computation, but generates structure through segmentation.
2、 The essential difference from mainstream AI
Dimension mainstream AI (mathematical formula) structural AI Core operation calculation (function, probability, optimization) segmentation (D ₀: distinguish foreground from background) Knowledge source external injection (training data) internal generation (segmentation accumulation) Operation mode matching, statistical prediction structure generation, recursive segmentation Low interpretability (black box) High (fully traceable) Self cognition without existence (a continuously evolving self model) Temporal stateless (each calculation is independent) with history (irreversible accumulation) Axiomatic System of Statistical Laws in Reference System Training Data More accurate prediction of goals and more complete cognitive structure
We use a very specific scenario to compare the differences between "mathematical intelligence" and "structural intelligence" in handling the same task.
Direct comparison: How to 'migrate' from learning to grabbing in factories to households?
Mathematical path: retraining
In the factory: Neural networks use massive data to remember specific parameters such as visual features, coordinate positions, joint angles, etc. corresponding to "grasping object A". Moving to the home: Faced with B objects of completely different shapes, colors, and positions, all the parameters previously remembered have become invalid. It needs to re collect data and retrain the model in order to establish new mapping relationships on specific new data. Structured Path: The AI-C Engine does not record "coordinate data" in the "segmentation" mode, but instead records the complete thought process of the robot "how to segment and extract graspable objects from the background" through divisiion_chain. It does not rely on the coordinate data of specific objects, but records the "grasping" structure that can be applied to any object.
At the factory: Division_chain records the structured steps of "cutting out the target object from the scene background → analyzing its spatial relationship with surrounding objects → planning the grasping path". Moving to the home: When the robot faces a new object, it no longer needs to be "retrained", but instead calls on the "grasping" structure stored in Division_chain. It will re divide the current family scene according to the established thinking path, identify new target objects, new spatial relationships, and plan new paths. Simply put, the cognitive system does not take away the "factory's answer" (coordinate data), but takes away the "steps and ideas for solving problems in the factory". Applying this idea to new scenarios is the essence of 'experience transfer'. Not by memory, but by mastering transferable cognitive structures to adapt to new environments.
Scenario: A robot is in the kitchen, and the user says, 'Give me that apple.
1、 The processing method of mathematical intelligence
operating process：User instructions → Large model analysis → Visual model recognition of apples → Path planning → Grab → Complete
specific steps
What happened in the steps? Essence 1 voice to text: "Give me that apple" input → text Analysis of Two Models: "Apple" is the target, "Take" is the action pattern matching Scan the scene with 3 visual models and match the feature matching of the "apple" 4. Calculate the grasping path, output control instructions, and perform mathematical calculations 5 robots perform grasping and complete execution
Result
✅ The apple has been brought over ❌ It doesn't know what it just did - there's no record ❌ It doesn't know why it chose this grasping angle - there's no explanation ❌ Next time encountering a similar scenario, it starts from scratch - without accumulation ❌ If an apple is replaced with an orange, it may not recognize it - there is no migration essence
Mathematical intelligence=a function: input (instruction+image) → output (action) It completed the task, but it did not 'understand' the process.
2、 Structural intelligence processing method
operating process
User instructions → Perceive changes → Split events → Record to H → Query Division_chain → Update Mself → Generate Report → Execute Action → Record Result → Reflect
specific steps
What happened to the corresponding structure of the steps The system perceives the change of "sound" as an attention filter 2. Cut out the "instruction" from the background and add a new split to divisio_chain 3 records to the history stack: "User requests to get an apple" H Have I ever taken an apple before? Experience Search 5 found that "I have taken an apple before and used a 45 degree angle" for experience matching 6 visual perception scenes, divide into "apples" divisio_chain, add segmentation 7. Combining historical and current factors, choose the grasping strategy of Mself to participate in decision-making 8. Perform physical grasping actions Record result: "This capture was successful and took 2.3 seconds" H update Update 10: "I am becoming more and more proficient in grabbing apples" 11. Generate report: "I completed the capture using a 45 degree angle" Report
The second time encountering a similar task
Scenario: The next day, the user says, "Give me that orange
Mathematical intelligence
Re parse instructions → Re identify oranges → Re plan paths → Execute Doing this for the first time, just like yesterday It doesn't remember taking an apple yesterday Structural Intelligence
Querying Division_chain: There is a splitting mode for "taking an apple" Identify that "taking oranges" and "taking apples" are structurally similar Call yesterday's experience (grasping angle, force, speed) Just need to fine tune (oranges are slightly heavier than apples) Faster and smoother completion Update on Myself: "I am becoming increasingly skilled at grabbing spherical fruits
Why is this distinction important
Limitations of Scene Mathematical Intelligence and Advantages of Structural Intelligence Industrial robots need to be reprogrammed every time they change products, and the accumulated grasping modes can be reused across products Home robots are always doing household chores for the first time, and the more they do, the more they understand this home Humanoid robot falls and doesn't know why it can trace 'why it fell' Long term deployment will not improve, and the more you use it, the smarter you become Security audits cannot trace decisions, and every step can be audited
A reference statement
Mathematical intelligence is like an AI that is always doing things' for the first time '. It can complete tasks every time, but it always starts from scratch. Structural intelligence is like an AI that 'remembers' - it accumulates once it does it; Do it a hundred times, optimize it a hundred times. It can not only complete tasks, but also explain 'why did you do it this way' and transfer experience to new tasks. This is the power of 'division' and 'accumulation'
🔥 Core Features
| Characteristics | Description |
|---|---|
| ** 🧠 Structural cognition * * | grows from an empty state by splitting into atomic operations, rather than relying on training data |
| ** 📜 Traceable decision-making * * | All operations are recorded in the historical stack, and the decision-making process is completely transparent |
| ** 📈 Experience accumulation * * | Splitting chain records the growth path of cognitive structure, achieving cross scenario migration |
| ** 🪞 Self model * * | The system continuously forms and updates its own cognition of itself |
| ** 💪 Body Perception * * | Cognition and Autonomous Adjustment of Energy Consumption, Force Perception, Temperature, and Wear |
| ** 🎯 Autonomous target generation * * | Automatically generate targets without instructions and continue to operate |
| ** 🔧 Correction layer * * | Detect anomalies and self correct |
🏗️ Architecture Overview
┌─────────────────────────────────────────────────────────────────┐ │ Cognitive layer │ │ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │ ││ H ││ division ││ Mself ││ correction layer ││ ││ Historical Stack ││ _chain ││ Self Model │││ ││││ Split Chain ││││││ │ └──────────┘ └──────────┘ └──────────┘ └──────────┘ │ │ ┌──────────────────────────────────────────────────────────┐ │ ││ Autonomous target generator ││
├─────────────────────────────────────────────────────────────────┤ │ Body Perception Layer │ │ ┌──────────┐ ┌──────────┐ ┌──────────┐ ┌──────────┐ │ ││ Energy consumption cognition ││ Force perception cognition ││ Temperature cognition ││ Wear and tear cognition ││ │ └──────────┘ └──────────┘ └──────────┘ └──────────┘ │ ├─────────────────────────────────────────────────────────────────┤ │ External layer │ │ ┌──────────────────────────────────────────────────────────┐ │ ││ External Resource Integration Layer · Translation Layer · Report││ │ └──────────────────────────────────────────────────────────┘ │ └─────────────────────────────────────────────────────────────────┘
🚀 Quick Start
Installation
git clone https://github.com/ai-c-engine/General-Structural-Artificial-Intelligence.git cd ai-c-engine pip install -r requirements.txt
Running example
python examples/basic_cognition.py
python examples/body_perception.py
python examples/dialogue_flow.py
📦 Python
- Python 3.8+
🤖 Application Scenarios
| Scenario | Description |
|---|---|
| * * Embodied Intelligence/Humanoid Robots * * | Provide traceable decision chains and body perception for robots |
| * * AI Agent * * | Provides evolved cognitive structures and experience accumulation |
| * * Industrial robots * * | Fault tracing, energy consumption optimization, maintenance prediction |
| * * Research Platform * * | Validate the Theoretical Hypothesis of Structural Intelligence |
📚 Document
(docs/THEORY.md) (docs/API.md) (docs/DIMOS_INTEGRATION.md) (docs/DEPLOYMENT.md)
🗺️ Document roadmap
-[x] Core Cognitive Layer (H, Division_chain, Mself) -[x] Autonomous target generator -[x] Correction layer -[x] Forgetting and Recovery Mechanism -[x] Body perception layer (energy consumption, force perception, temperature, wear and tear) -[] Integration with DimensionalOS -[] Real machine deployment verification -[] Long term autonomous operation experiment
Welcome to participate through the following methods: -Submit Issue or PR -Share usage cases -Discuss theoretical extensions
📝 open source license
MIT License
📬 Contact Information
Project author: [C] Email: [as2020319@163.com] Project Address: https://github.com/ai-c-engine/General-Structural-Artificial-Intelligence.git
AI-C Engine is a continuously evolving open source project, welcome to follow and participate