SAM³: Concept-Driven Framework

• SAM³ is a systematic, multi-level metamodel that defines and organizes scenario variables across method, suite, scenario, and event layers.
• It partitions scenario variables with formal definitions and constraints to ensure agile and precise scenario composition, especially in automated driving.
• The framework integrates structured metric computation and safety analysis, supporting both deductive and inductive scenario engineering workflows.

The SAM³ Concept-Driven Framework denotes a systematic, multi-level metamodel for scenario engineering, rigorously defined via a set of scenario variables (SVs) that provide the semantic backbone for structured scenario definition, management, and analysis. Synthesized from the Conceptual Scenario Model (CSM) literature, SAM³ explicitly subsumes the full scenario engineering lifecycle—method, suite, scenario, and event—through a comprehensive taxonomy of variables, formal object definitions, type-level relations, and formalized constraints, thus enabling precise, agile, and extensible scenario-centric development for software and systems domains, and particularly for automated driving systems (ADS) (Baek et al., 2022).

1. Scenario Variable Taxonomy and CSM Layering

SAM³ is structured by the formal partitioning of scenario variables (SVs) into four layers, each representing a distinct abstraction in scenario-based engineering:

• Method-level SVs formalize the overarching purpose, specification, and execution characteristics. Examples include MethodPurpose, decomposing into TargetProblemDomain, TargetVisionIntent, Hypothesis, StrategyTactic, and Assumption.
• Suite-level SVs structure scenario collections, providing SVs such as SuiteMetaData, SuiteViewpoint (SuiteGoalValue, SuitePerspective, SuiteBaselineScenario), and SuiteOntology (SystemOntology, InfraOntology, EnvOntology, SituationOntology, IntangibleFactorOntology). Suite-level composition operators (InterScenarioAssociation, InterScenarioConcurrency, InterScenarioCausality) govern scenario orchestration.
• Scenario-level SVs capture the semantics of individual scenarios: ScenarioMetaData (ScenarioType, ScenarioUnit), ScenarioElements (Participants, Events, InterEventRel, TransitionPaths, EventPatterns), and functional links to inputs, outputs, temporal, geospatial, condition, change, external interaction, and uncertainty variables.
• Event-level SVs resolve the lowest granularity, including EventMetaData, EventInput/Output, EventCondition, EventBehavioral (EventActor, EventAction, EventOccurrenceMeasure, EventHandler), EventInteraction, EventTemporal, EventGeospatial, and EventUncertainty.

This SV taxonomy supplies the essential conceptual scaffolding for integrating qualitative (ontological) and quantitative (metric, probabilistic) scenario attributes at every modeling layer (Baek et al., 2022).

2. Formal Definitions and Typing

Each SV is formally typed via LaTeX-style signatures.

Let $M_\text{SVs}$, $St_\text{SVs}$, $Scn_\text{SVs}$, and $Evn_\text{SVs}$ be disjoint sets of Method-, Suite-, Scenario-, and Event-level variables, respectively. For any $$SV \in (M_\text{SVs} \cup St_\text{SVs} \cup Scn_\text{SVs} \cup Evn_\text{SVs})$$:

• Assignments include, e.g.,

$$\text{MethodPurpose} \equiv \langle \text{TargetProblemDomain}: \text{String},\, \text{TargetVisionIntent}: \text{String},\, \text{Hypothesis}: \text{String},\, \text{StrategyTactic}: \text{String},\, \text{Assumption}: \text{String} \rangle$$

• Relations and functions specify cross-level behavior:
  • $$f_\text{behavior}: \text{Entity} \times \text{Context} \to \text{EventAction}$$
  • $$g_\text{output}: \text{EventInput} \times \text{EventBehavioral} \to \text{EventOutputData}$$
  • $$constraint: \text{ScenarioCondition} \implies \text{Boolean}$$
  • $uncertainty: \text{EventUncertainty} \to [0,1]$

Each SV is explicitly associated to its layer via $$BelongsToLevel: SV \to \{\text{Method}, \text{Suite}, \text{Scenario}, \text{Event}\}$$.

3. Metamodel Architecture and Structural Constraints

SAM³'s metamodel formalizes the relationships across the four SV layers through class- and attribute-level structure (UML-style):

• ScenarioMethod: encapsulates MethodPurpose, specification, and execution, linking to a variable set of ScenarioSuites.
• ScenarioSuite: defines suite-wide metadata, viewpoint, ontology, scenario pool, input and scenario composition.
• Scenario: partitions into metadata, elements, targets, I/O, temporal, geospatial, change, external interaction, uncertainty; aggregates 1..* Events.
• Event: atomic level; encodes meta, input, output, condition, behavioral, interaction, temporal, geospatial, and uncertainty fields.

Key formal constraints include:

• $$\sum_{s \in \text{ScenarioSuite}} |s.\text{scenarios}| \ge 1$$
• $$\forall scn \in \text{Scenario}, \forall ev \in scn.\text{events}: ev.\text{meta}.\text{EventType} \in scn.\text{meta}.\text{AllowedTypes}$$
• $$\text{ScenarioCondition}(scn).\text{precondition} \implies \forall ev \in scn.\text{events}: \text{eventCondition}(ev).\text{startTrigger} = \text{True}$$
• Temporal and spatial relations are explicitly constrained via $t_\text{start} < t_\text{end}$ and allowed $\text{georep}$ levels.

This formal architecture enables compositional reasoning, inheritance of suite-wide parameters, and agile extension to new scenario or event types (Baek et al., 2022).

4. Specialization for Automated Driving and Safety Analysis

For automated driving systems (ADS), SAM³'s SuiteOntology is directly extended to encompass ODD (Operational Design Domain) variants, e.g., SAE_ODD_Level and WeatherOntology. ScenarioTarget is specialized with ScenarioTargetCriticality (values: Normal, Warning, Critical) and ScenarioIndicatorOracle (Boolean pass/fail). ScenarioOutputData is expanded to include first-class safety metrics such as SafetyDistance, ReactionTime, and Time-to-Collision (TTC), all parameterized as real-valued variables:

• Safety metrics definitions:

$$\text{SafetyDistance} = v_{\text{ego}}\cdot t_{\text{react}} + \frac{v_{\text{ego}}^2}{2a_{\text{brake}}}$$

Parameters and results are linked to event-level inputs (e.g., velocity, lead vehicle state) and evaluated against thresholds at runtime.

EventBehavioral ties scenario logic to physics-based handlers, allowing for faithful simulation of control/action sequences in driving scenes.

SAM³ explicitly supports both top-down (MethodPurpose $\to$ SuiteViewpoint $\to$ Scenarios $\to$ Events) and bottom-up (EventPool composition $\to$ Scenario construction $\to$ Suite grouping) scenario development workflows.

5. Exemplary Instantiation and Metrics Computation

Consider the "Cut-In" scenario for ADS:

• Suite-level: SuiteGoalValue = maximize safety margin; SuiteOntology includes lanes and highway semantics.
• Scenario: ScenarioType = CutIn, ScenarioInputParameter = $\{v_{ego}=25$ m/s$, gap_{init}=50$ m$\}$.
• Condition: Precondition: $gap_{init} \ge 40$ m; Temporal span 0–10 s.
• Events: LeadVehicleCutIn and BrakeCommand with associated uncertainty values.
• Safety computation:

$$d_{safe} = 25*0.5 + \frac{25^2}{2*5} = 75\,\text{m};\quad \text{new\_gap} = 50-3 = 47\,\text{m} < 75\,\text{m}$$

• ScenarioOutputData: SafetyDistance = 75 m, ScenarioIndicatorOracle = False (fail).

This mechanism operationalizes metrics computation and safety assessment as intrinsic semantic outputs at the scenario and event levels (Baek et al., 2022).

6. Comparative Analysis with Existing Frameworks

SAM³ provides full four-level (Method→Suite→Scenario→Event) coverage; explicit SuiteComposition and SuiteOntology handling (including ODD); support for agile composition and iterative refinement; and direct linkage of scenario logic to physics-based models and quantitative uncertainty attributes. In contrast, alternative methods such as ASAM OpenSCENARIO, PEGASUS, Schütt’s SceML, and scene-centric DSLs (Bach et al., Scenic) are each limited in compositional coverage, generic uncertainty handling, or metric formalization.

SAM³’s unification of SV taxonomies, architectural layering, and metric-driven outputs position it as a rigorous, extensible platform for scenario-based engineering in domains of high system complexity, safety criticality, and semantic heterogeneity (Baek et al., 2022).

7. Significance and Extensibility

SAM³, as the instantiation of the CSM with full SV coverage, provides a universal metamodel for scenario definition, composition, and evaluation. Its methodological rigor is domain-agnostic, but the formal SV set and extensible ontology enable rapid specialization to application-specific ODD, metrics, and event/actor ontologies—while supporting both deductive (top-down) and inductive (bottom-up) scenario engineering workflows. A plausible implication is that SAM³'s formalism supports not only scenario library design and simulation-based validation, but also comparative scenario-method analysis, safety-criteria traceability, and continuous scenario complexity management (Baek et al., 2022).