Journal of Advanced Robotics, Autonomous Systems and Human-Machine Interaction
ARYA: A Physics-Constrained Composable & Deterministic World Model Architecture
Abstract
Seth Dobrin and Lukasz Chmiel
This paper presents ARYA, a composable, physics-constrained, deterministic world model architecture built on five foundational design principles: nano models, composability, causal reasoning, determinism, and architectural AI safety. We demonstrate that the platform satisfies the formal requirements established by the world model research community, including state representation, dynamic prediction, causal and physical awareness, temporal consistency, generalization, learnability, and applicability to planning and control. Unlike monolithic large language models, ARYA implements these capabilities through a hierarchical system-of-system-of-systems of specialized nano models orchestrated by AARA (ARYA’s Autonomous Research Agent), an always-on cognitive daemon that operates a continuous sense-decide- act-learn loop. The nano model architecture provides linear scaling, sparse activation (invoking only task-relevant models), selective untraining, and sub-20-second training cycles. Combined, these properties resolve the traditional tension between capability and computational efficiency. A central contribution is the “Unfireable Safety Kernel”, an architecturally immutable safety boundary that cannot be turned off, bypassed, or circumvented by any component of the system, including its own self-improvement engine. This layer is not a statement on social or ethical alignment; rather, it is a technical framework for ensuring that human control and governance are maintained as the system’s autonomy increases. Safety is not a policy layer applied after the fact; it is an architectural constraint that governs every operation the system performs. We present the formal alignment between ARYA’s architecture and the canonical world model requirements, summarize its state-of-the-art performance across 6 of 9 competitive benchmarks, and describe its deployment across seven active industry domain nodes (aerospace, pharma manufacturing, oil & gas, smart cities, biotech, defense, and medical devices), and report empirical evaluation on nine external benchmarks where AARA achieves state-of-the-art on six—including causal reasoning, physics reasoning, PhD-level science, enterprise workflows, embodied planning, and AI safety—with zero neural network parameters).