AI-Native Platform Engineering for Physical AI

Read full text Summarise with VIRAT
Authors: Petteri Teikari
Cited by: 0

Summary

No summary yet. VIRAT can read this paper and write a plain-language summary of its aim, methods, key findings and limitations.

Summarise with VIRAT Free to try with your Science Catalogs account.

Abstract

This presentation provides a comprehensive architecture specification for an AI-native internal platform designed to support Physical AI applications — autonomous unmanned aerial systems (UAS), ground robots, and heterogeneous sensor networks — in defense technology, law enforcement, border patrol, and search-and-rescue domains. The deck synthesizes industry and academic sources into a unified platform design organized around five horizontal capabilities: simulation and digital twin (Isaac Sim, Gazebo, Pegasus), multimodal data management and replay (Rerun Cloud, MCAP, ROS2), intelligence fusion (OSINT and CLOSINT aggregation into a common operating picture), browser-based 3D/4D visualization (React, Rust/WASM, CesiumJS, MapLibre GL, 3D Gaussian Splatting), and agentic developer experience (Claude Code, Model Context Protocol, curated skills). Orchestration follows a proven two-tier pattern: Prefect 3.x for deterministic macro-orchestration of pipeline DAGs and Pydantic AI for optional LLM-assisted micro-orchestration within flows. The presentation makes three novel contributions. First, it demonstrates that Physical AI platform engineering — the intersection of platform engineering, robotics/autonomous systems, and defense — is a research gap with essentially no academic literature, despite significant industry investment. Second, it proposes a concrete OpenSpec methodology for specification-driven development (SDD) of such platforms, enabling parallel team execution without merge conflicts. Third, it introduces a three-layer governance architecture (Governance-as-Code via WASM policy modules, Governance-as-a-Service with Trust Factor scoring, and Content-Based Access Control for contextual information flow) that addresses the unique compliance requirements of classified defense environments. The slides are organized in six acts: Act 0 (Opening, slides 1-2) provides the executive summary and graphical abstract. Act I (Motivation & Landscape, slides 3-16) establishes why Physical AI needs dedicated platform engineering, surveys the competitive landscape, and identifies the gap. Act II (The Architecture, slides 17-32) details the five-layer platform architecture, from orchestration through intelligence fusion. Act III (Platform Tasks, slides 33-44) covers recording, annotation, sensor QA, calibration, and 3DGS sim-to-real. Act IV (Advanced Topics, slides 45-52) addresses edge compute continua, heterogeneous swarms, and RL infrastructure. Act V (Security & Governance, slides 53-57) presents MLSecOps, zero-trust, and the three-layer governance architecture. Act VI (Roadmap & Closing, slides 58-64) delivers the phased implementation plan, team structure, technology stack, TCO analysis, risk register, and the research opportunity.

Fields of Study & Subjects

Robotics and Automated Systems Military Strategy and Technology Artificial Intelligence Applications

Publication details

Journal / source
Figshare
Publication date
27 March 2026
Language
EN