Nexum 2.0: A Hybrid Operating System Integrating Robotics, Computer Vision, and Local AI

Nexum 2.0 is a hybrid operating system that integrates robot control, computer vision, and local AI processing into one unified system. Built on Zorin OS, it supports multi-scenario deployment from edge devices to high-performance workstations.

• Original Author/Maintainer: devlucassilvapetris
• Source Platform: GitHub
• Original Project Name: Nexum-2.0
• Original Link: https://github.com/devlucassilvapetris/Nexum-2.0
• Release Date: 2026-06-09

---

Nexum 2.0 is a hybrid operating system for human-machine interaction, integrating robot control, computer vision, and artificial intelligence capabilities into a unified software framework. Unlike traditional robot systems, Nexum 2.0 emphasizes local AI processing—all intelligent computing is done on the device side without relying on external cloud services, which has significant advantages in privacy-sensitive and offline scenarios.

The system is built on Zorin OS, fully leveraging the flexibility and stability of the Linux ecosystem while being deeply optimized for real-time control tasks.

---

Nexum 2.0 adopts a layered architecture design with clear responsibilities for each module, facilitating expansion and maintenance:

The kernel layer is the neural center of the system, responsible for:

• System Management: Resource monitoring and process scheduling
• Hardware Drivers: Unified management of various hardware devices
• Module Communication: Providing inter-process message passing mechanisms

This layer abstracts underlying hardware operations and includes three core subsystems:

Kinematics Control Supports precise control of servo motors and stepper motors, with built-in Kalatec integration solutions. Developers can configure parameters such as maximum speed, acceleration, and torque of motors to achieve smooth motion trajectory planning.

Acoustic System Adopts a bionic ear design and supports high-fidelity audio collection. Configurable parameters include sampling rate (default 44.1kHz), buffer size, and gain, suitable for speech recognition and environmental monitoring scenarios.

Mechanical Structure Handles mechanical control and physical computing, providing a unified physical world abstraction for upper-layer applications.

This is the most distinctive part of Nexum 2.0, integrating a complete AI technology stack:

Local Model Management

• Support hot loading and version switching of models
• Run inference without networking
• Protect user privacy data

Model Fine-tuning

• Efficient fine-tuning based on Unsloth
• 4-bit quantization reduces memory usage by 75%
• Supports LoRA and PEFT parameter-efficient fine-tuning
• Gradient checkpointing allows small VRAM devices to train large models

Model Serving

• vLLM + PagedAttention for high-throughput inference
• Ray Serve provides auto-scaling capabilities
• GPU utilization can reach 60%, leaving room for concurrent tasks

RAG (Retrieval-Augmented Generation)

• LangChain orchestrates RAG pipelines
• Qdrant for rapid prototyping
• Milvus supports production-level deployment (scalable to billions of vectors)
• Instructor ensures structured output

Prompt Optimization

• DSPy framework for programmatic prompts
• Automatically optimize prompt template effectiveness

Model Evaluation

• Ragas provides multi-dimensional RAG evaluation metrics
• Weights & Biases tracks experiment processes
• Evaluation dimensions include: faithfulness, answer relevance, context precision, recall

The vision module provides real-time image processing capabilities:

• Image Processing: Real-time enhancement, noise reduction, contrast adjustment
• Object Recognition: Advanced detection and classification algorithms
• Scene Analysis: Environmental understanding and interaction perception

---