TLDR: This research presents a model-driven method to automatically generate AI planning descriptions (PDDL files) from SysML-based engineering models. It uses a dedicated SysML profile to embed planning semantics directly into engineering designs and a transformation algorithm to convert these enriched models into planning artifacts. A four-phase workflow guides the process from model analysis to automated generation. Demonstrated through an aircraft assembly case study, the method enables consistent validation of automated production system variants, optimizing task sequences and reducing manual effort by bridging the gap between detailed engineering and symbolic AI planning.
The field of automated production systems is constantly evolving, requiring sophisticated methods to design, validate, and optimize complex processes. A significant challenge lies in bridging the gap between detailed engineering models, often created using Model-Based Systems Engineering (MBSE) and Systems Modeling Language (SysML), and the symbolic reasoning capabilities of Artificial Intelligence (AI) planning.
Engineering models excel at describing system structure and behavior, but they typically lack the specific planning semantics—such as action preconditions, effects, and resource constraints—that AI planners need to evaluate task feasibility and efficiency. This disconnect often leads to manual, time-consuming, and error-prone transformations between engineering designs and planning descriptions.
A Novel Model-Driven Approach
A recent research paper, “Bridging Engineering and AI Planning through Model-Based Knowledge Transformation for the Validation of Automated Production System Variants,” introduces a novel model-driven method to address this challenge. The approach enables the automated generation of symbolic planning artifacts, specifically PDDL (Planning Domain Definition Language) domain and problem files, directly from SysML-based engineering models. Unlike previous methods that rely on manual transformations or external capability models, this approach integrates planning semantics natively within the engineering models, ensuring consistency and traceability.
The Four-Phase Workflow
The method is structured around a four-phase workflow:
Phase I: System Model Analysis and Scoping
This initial phase involves analyzing existing engineering models to understand the system architecture and identify the specific components, resources, and relationships relevant to the planning task. Since engineering models can be very large and complex, a crucial step is to define a clear scope, focusing only on the information needed for the planning domain.
Phase II: System Model Enrichment with Planning Semantics
In this phase, the relevant parts of the system model are extended with planning semantics using a dedicated SysML profile. This profile introduces reusable stereotypes for core PDDL constructs, such as object types, predicates (logical conditions), and actions. By applying these stereotypes, engineers can embed planning knowledge directly into their SysML models, defining how system elements behave in terms of parameters, preconditions, and effects.
Phase III: Product Model Integration
While the system model defines general capabilities, the product model contains instance-level data specific to a particular planning task, such as the positions and types of components. This phase involves extracting relevant product information and transferring it into the MBSE environment. This data is then annotated using the same planning concepts defined in Phase II, ensuring consistency between the domain (general rules) and the problem (specific scenario).
Phase IV: Automated PDDL Generation
The final phase leverages a transformation algorithm to automatically generate the PDDL domain and problem files from the enriched system and product models. This algorithm systematically extracts the stereotype-annotated model content and converts it into syntactically correct PDDL code. The generated files can then be fed into standard PDDL solvers to compute optimized plans.
The SysML Profile and Transformation Algorithm
The core of this method lies in its dedicated SysML profile, which maps PDDL concepts like domain definitions, object types, logical conditions, and action specifications to standard SysML elements. For instance, a PDDL object type might be represented by a SysML Class stereotyped as <<PDDL Type>>, and a planning action by a CallBehaviorAction stereotyped as <<PDDL Action>>. The profile also includes Object Constraint Language (OCL) constraints to validate the consistency of the planning semantics within the model.
The transformation algorithm, implemented using the Apache Velocity Engine, processes these annotated models. It identifies planning-relevant elements through stereotype-specific queries and converts them into PDDL syntax using Velocity Template Language (VTL) templates. This ensures that the generated PDDL code accurately reflects the logic embedded in the SysML model, maintaining traceability.
Case Study: Aircraft Assembly
The applicability of this method was demonstrated through a practical case study in aircraft structure assembly. The scenario involved the automated planning of collar screwing operations inside an aircraft fuselage using a robotic arm with interchangeable tools. The objective was to generate an optimized plan that minimizes cycle time and reconfiguration effort.
The workflow was applied to an existing SysML model of the Collar Screwing System. Planning semantics were embedded, product data (rivet types and positions) was integrated, and PDDL files were automatically generated. These files were then validated and used with a PDDL planner (Delfi) to compute an optimized sequence of operations. The resulting plan effectively grouped operations by tool type, reducing unnecessary tool changes and travel distances, thereby optimizing the overall process time.
Also Read:
- Unifying Autoformalization: A Framework for Bridging Informal and Formal AI Reasoning
- Making Ontologies Active: The boldsea Approach to Event-Driven Processes
Conclusion and Future Directions
This model-driven approach offers a significant advancement in integrating AI planning with engineering workflows. It provides a traceable and automated way to generate consistent planning artifacts directly from verified engineering models, supporting efficient variant analysis and task-oriented system planning. The authors, including Hamied Nabizada, Lasse Beers, Alain Chahine, Felix Gehlhoff, Oliver Niggemann, and Alexander Fay, highlight that this method is domain-independent and applicable to various engineering contexts requiring symbolic reasoning.
Future work aims to further simplify the modeling process, potentially through interactive annotation assistants and semi-automated suggestions. Improved integration between product data tools (like Dassault Systèmes 3DExperience) and MBSE environments (like Magic Systems of Systems Architect) is also planned to reduce manual effort. Additionally, compatibility with the Unified Planning Library and migration to SysML v2 are being considered to ensure the method’s continued relevance and broader applicability. You can read the full research paper here.
