omnitrain 1.1.2

Multimodal AI Framework for Robotics

OmniTrain v2.1.0: Conectoma

Bio-Inspired Sparse Neural Circuits & Formal Safety for Robotics

[!WARNING] This project is currently UNFINISHED and in PRE-BETA phase. Core features and APIs are undergoing active development, testing, and security audits. Use with caution.

---

OmniTrain is a production-grade framework for building Bio-Inspired Conectomas (Hub & Wall architecture). It utilizes Closed-form Continuous-time (CfC) networks and Input Convex Neural Networks (ICNN) to provide sub-millisecond, provably safe robotic control on edge hardware (Jetson/Qualcomm).

---

What's New in v2.1 (Update)

• Training-Serving Parity: Automatic capture and application of Z-Score normalization statistics. No more data degradation in deployment.
• Lagrangian Stability: Stabilized primal-dual safety controller using per-sequence dual updates.
• Unified Fusion: Optimized multi-sensor ingestion in OmniStream to prevent neural double-evolution.
• Kernel Robustness: Enhanced CLI with kernel-level exception handling for 24/7 mission-critical operation.
• Hardware Failsafes: Improved Tier 1 monitoring with worst-case coverage across all sensor dimensions.
• Stabilization: Passed the Integrity 5-Problem Health Audit (v2.1) ensuring zero-leak SHM and RK4 dynamics parity.

---

Quick Start

1. Installation

You can install OmniTrain directly from PyPI:

pip install omnitrain

Alternatively, you can install from source:

Linux/macOS Source Installation

git clone https://github.com/Mrmyms/Omnitrain.git
cd Omnitrain
chmod +x setup.sh
./setup.sh

Windows Source Installation

git clone https://github.com/Mrmyms/Omnitrain.git
cd Omnitrain
python -m venv .venv
.venv\Scripts\activate
pip install -r requirements.txt
pip install -e .

2. Scaffold a New Project

from omnitrain import ProjectManager

# Set up project folders, default config.yaml, and a training dataset
ProjectManager.init_project()

3. Programmatic Training Curriculum

from omnitrain import LiquidTrainer

# Load trainer and fit behavioral and safety policies
trainer = LiquidTrainer("config.yaml")
metrics = trainer.fit("robot_logs.csv", epochs=5)

4. Edge Deployment Compile

from omnitrain import EdgeDeployer, ESP32Exporter

# Option A: Export to Qualcomm Snapdragon (DLC) or TensorRT
deployer = EdgeDeployer("models/bot_brain.omni")
deployer.export(target="tensorrt")

# Option B: Export to Microcontrollers (ESP32) Zero-Copy Binary
exporter = ESP32Exporter(output_dir="esp32_firmware/data")
exporter.export(model, input_dim=8, d_model=128, output_dim=4, filename="bot_brain.omnibit")

5. Multi-process Sensor Runner

from omnitrain import AgentRunner

# Start the real-time TokenBus circular buffer and spawn sensor modalities
runner = AgentRunner("config.yaml")
runner.start()

# Monitor live circular buffer stream telemetry for 10 seconds
runner.run_telemetry(duration=10.0)
runner.stop()

---

Architecture: The Conectoma v2.1

graph TD
    S1[Sensors] -->|Z-Score Normalization| H[BioConectomaHub]
    subgraph "Hub & Wall"
        H -->|Sparse Mask| W[Interneuron Wall]
        W -->|Recurrent| W
        W -->|Decision| C[Command Layer]
    end
    C -->|Safe Action| SG[OmniShield v2.1]
    SG -->|Convex Barrier| M[Motor Output]
    style H fill:#d1f2ff,stroke:#333
    style W fill:#fff4d1,stroke:#333

---

Resources

• Technical Deep Dive & Conectoma Spec: Official architecture specification, CfC cells, and ICNN barriers.
• Theoretical Frameworks: Liquid Networks, ICNNs, and CTMT math.
• Connectivity Guide: Connecting sensors (ROS 2, Isaac Sim, CSV) to TokenBus.
• LNN Research Database: MIT CSAIL papers, benchmarks, and LNN/CfC references.
• Training Pipeline: 5-phase curriculum (Imitation, Safety, Noise, Lagrangian Stability, Pruning).
• ESP32 Edge Port Guide: Zero-Copy C++ Engine deployment instructions for 512KB microcontrollers.

---

OmniTrain Team "Fuse Everything. Trust Nothing. Verify Formally."