Executive Summary

Introduction — The Data Bottleneck of Physical AI

The Physical AI market is poised for explosive growth as AI technologies that interact with the physical world spread across manufacturing, robotics, defense, and shipbuilding. But behind that innovation lies a critical bottleneck — Data Famine.

Rare data such as defects, edge cases, and disasters — the very situations that determine AI model robustness — is nearly impossible to collect intentionally in the real world. Welding-line defects that occur at less than 0.001% in a smart factory, extreme weather and sudden accidents that an autonomous vehicle may face: this kind of scarcity is the single largest barrier to Physical AI adoption.

PebbloSim is the strategic application designed to solve this bottleneck. Within the Pebblous Data Greenhouse ecosystem, it operates as the core execution engine of the Action layer — completing the autonomous cycle of Observe, Judge, Act, and Prove. It is the most powerful proof that Data Greenhouse is not merely an observation system but an operating framework that takes responsibility for data quality and lifecycle.

PebbloSim accelerates the development of industry-specific AI models — a key goal of the AADS (Agentic AI Data Scientist) Phase 2 program — and helps secure data sovereignty in the Physical AI market. This document defines PebbloSim's reason for existence (Why) and ultimate goal (What), then lays out the architecture and development strategy that realize them.

PebbloSim is the core execution engine of the Action layer within the Pebblous Data Greenhouse ecosystem, which autonomously cycles through Observe, Judge, Act, and Prove operations on data.

1.1Core Concept & Objectives

PebbloSim is defined as "a digital twin–based simulation and synthetic-data generation platform for Physical AI training data." Beyond simply building virtual environments, it is organically integrated with the Data Greenhouse to autonomously produce high-quality, AI-ready data.

Where conventional generative AI (Sora, Stable Diffusion, and similar models) relies on probabilistic correlations and frequently violates physical laws — a phenomenon known as Physical Hallucination — PebbloSim grounds neural rendering in the physical consistency guaranteed by a digital twin engine. This Zero Physical Hallucination principle is the foundation of PebbloSim's technical credibility and the precondition for synthetic data to substitute for real process data on industrial floors.

1.2Key Application Domains

PebbloSim targets four key domains where Data Famine is most severe. They share a common structural problem — collecting defect, accident, and edge-case data is practically impossible — and they show high demand for high-quality synthetic data with guaranteed physical consistency.

1.3Key Differentiators

The synthetic-data market already has many players. PebbloSim secures structural differentiation through three things — recording the generation process itself as auditable Operational Evidence, guaranteeing physical consistency through a Neuro-Symbolic World Model, and creating a self-reinforcing Data Flywheel whose diagnostic and generative accuracy improves with use.

Differentiator 1 — PebbloSim as Operational Evidence

Functioning as core Operational Evidence means PebbloSim is not merely a tool that produces synthetic data files (.jpg, .mp4, etc.) but a system that generates an Audit Trail proving "why this data was generated, and through what process its quality improved."

Conventional simulators deliver the data you requested — and nothing more. PebbloSim delivers a record of how Data Greenhouse autonomously solved the underlying problem. It does not merely generate "an image of a rainy day"; it records the full causal chain — Data Clinic diagnosed insufficient rainy-weather data (Why), AADS configured precipitation 10 mm and illuminance 50 lux (How), and PebbloSim generated the data (Action). The Operational Evidence Package combines three elements:

Differentiator 2 — Neuro-Symbolic Hybrid World Model

To overcome the limitations of generative AI that ignores physical laws, PebbloSim combines the logical consistency of symbolic simulation with the visual expressiveness of neural generative models. Video-generation models such as Sora and Stable Diffusion learn only probabilistic correlations between pixels, producing physical hallucinations — cars floating in midair, shadows misaligned with light sources. PebbloSim instead first builds a World Model governed by gravity, friction, and optics, and then dresses that skeleton in generative-AI skin.

Differentiator 3 — Self-Reinforcing Cycle (Data Flywheel)

A structure that gets smarter with use. Data production strengthens AI intelligence, and stronger intelligence produces more sophisticated data — a self-reinforcing virtuous cycle that forms a technological moat competitors cannot easily replicate. The cycle operates through three mechanisms.

First, internalizing intelligence. The high-quality synthetic data (Curriculum Data) PebbloSim generates is not merely shipped to customers — it is also used to retrain the core AI models that serve as the system's brain. Through this self-training loop, the system evolves over time to understand more complex physical situations and to design more sophisticated scenarios.

Second, reinforcement between diagnosis and generation. When the system spots a data void, simulation fills it; as the model improves on that data, it begins to detect even subtler defects that were previously invisible. This infinite loop creates a Data Flywheel that exponentially compounds the value of an enterprise's data assets.

Third, asset appreciation. While conventional software becomes obsolete over time, the Data Greenhouse powered by PebbloSim is an appreciating asset whose diagnostic and generative accuracy improves as data accumulates. This is a moat that competitors cannot replicate in the short term.

1.4Regulatory Compliance & Business Value

PebbloSim's operational evidence is the key to solving the AI regulatory and reliability challenges that enterprises face.

PebbloSim is a core application that runs on top of Data Greenhouse, an AI data operating system (Data OS). Its workflow is captured by a single equation — Engine (infrastructure) + Scenario (blueprint) = Simulator Instance (GenSim).

2.1Four-Stage Workflow

PebbloSim's workflow is not the operation of a stand-alone simulator but the execution cycle of an application running on a data operating system. It is dedicated to Data Bulk-up — using physics simulation to reinforce sparse regions of the data distribution — while staying fully synchronized with the Data Greenhouse's Observe-Judge-Act-Prove loop. This is what lets PebbloSim function not as an individual tool but as scalable platform infrastructure.

The four stages are Twin (Base Class) → Design (Architect) → Generate (Action) → Verify (Director). The Twin stage prepares a digital twin environment that respects physical law. The Design stage translates abstract commands into concrete scenario scripts. The Generate stage actively produces multimodal data, and the Verify stage uses PebbloScope to surface results in a form a human reviewer can audit. The cycle runs without breaks, keeping each GenSim instance autonomous.

2.2Core Modules

2.3Greenhouse Integration

The business value of PebbloSim does not come from running it standalone — it comes from the organic integration with the Greenhouse ecosystem. As the pipeline from diagnosis to generation to verification is automated, the customer no longer has to build separate manual pipelines to improve data quality. The repetitive work of data engineers is reduced, and the cost and time of quality improvement is structurally compressed.

We chose automotive painting as the first validation scenario for a clear reason — it is a textbook case where the absence of rare defect data directly bottlenecks AI model performance. Micro-scratches occur at less than 0.1% in real production lines, making it hard to assemble enough training data, and detection difficulty changes drastically with lighting conditions on the floor. The scenario also lets us quantify Sim-to-Real transfer effects, making it the industry reference that most directly demonstrates PebbloSim's ROI to both investors and customers.

3.1Five-Step Execution

A complex, ambitious platform like PebbloSim cannot be built big-bang. Pebblous adopts an incremental, iterative strategy — four quarterly PoC cycles over one year that progressively complete PebbloSim.

The core of this approach is a Wedge Use Case strategy. We start with the most urgent, ROI-clear automotive process (PoC 1) to prove immediate value, then expand into defense (PoC 2) and shipbuilding (PoC 3), and finally complete a fully autonomous platform (PoC 4). Each stage deepens integration with Data Greenhouse and is designed to contribute directly to the quantitative goals of the AADS Phase 2 government program — a progressive deepening strategy moving from Closing the Loop → Sovereignty → Depth → Autonomy.

The four-cycle strategy is also a risk-management framework. Each PoC depends on outputs from the previous stage, so technical uncertainties surfaced early flow immediately into later designs. And because every cycle's deliverable feeds directly into the AADS Phase 2 technical report, technology validation and project execution stay synchronized on a single timeline.

4.1Four-Cycle PoC Roadmap

5.1The Paradigm Shift

We declare a transition — from an era of collecting data we happen to find in the real world, to an era of cultivating data we deliberately design. PebbloSim is the cornucopia that supplies AI-ready data without the constraints of physical risk, cost, or time.

5.2Business Impact

In the era of Physical AI, collecting data is no longer enough. Cultivating it — and accumulating every piece of evidence about that cultivation in auditable form — is becoming the precondition for entering the market. PebbloSim is the center of that paradigm shift, the system that closes Data Greenhouse's diagnose-judge-generate-verify pipeline into a fully automated operating framework.