Overview
Object classification using AI/ML was developed to achieve a SIL-rated detection capability capable of distinguishing between different safety-relevant object classes and triggering appropriate system actions. The solution leverages a multi-sensor architecture combining radar-based detection (RBD), time-of-flight (ToF) sensing, and other sensing modalities. This work represents a novel application of AI in functional safety for human detection within safety-critical environments. 

Challenges
The system was developed in the absence of a published AI functional safety standard, with additional gaps in how IEC 61496 addresses testing of AI-based perception functions. This created uncertainty in how to define, validate, and certify safety-related AI behavior. 

A further challenge was the sensor-agnostic architecture, which made traditional verification approaches significantly more complex, as system-level behavior could not be validated using conventional sensor-specific methods. 

Considerations included balancing system safety behavior between two extremes: 

• A highly conservative, risk-averse system that becomes “too safe,” stopping for every detected object and negatively impacting operational workflow 

• A more generalized system that improves efficiency but risks missing critical detections, potentially leading to injury or fatal outcomes 

In addition, there were no established methodologies for verifying AI-based perception models in safety-critical applications. 

Approach
Several key design and engineering mitigations were implemented: 

• Definition and formalization of the Operational Design Domain (ODD) to bound and constrain safe system behavior 

• Structuring verification and validation to combine classical functional safety methods with AI-specific evaluation techniques 

• Implementation of diagnostic strategies to ensure out-of-distribution (OOD) conditions do not lead to hazardous behavior, while enabling safe fallback states 

• Development of a layered V&V approach integrating deterministic testing with AI performance validation 

• Extension of classical functional safety methods to AI-based perception systems, enabling structured evaluation of safety-relevant machine learning behavior 

Key Insights
The project highlighted several important findings:

• Proper definition and characterization of the ODD is essential to constrain AI safety behavior 

• Effective AI validation requires addressing both in-distribution performance and out-of-distribution scenarios 

• Safety architectures must explicitly handle uncertainty rather than treating it as an exception case 

• Classical functional safety methods can be extended to AI systems, but require additional structure to remain effective 

• AI-enabled safety systems require explicit consideration of human interaction use cases and real-world operational variability 

Outcome
A novel verification and validation framework for AI/ML-based safety systems was developed. The approach is still under development but has been reviewed and approved by an independent TIC assessment body, demonstrating feasibility for structured certification pathways of AI-enabled safety functions.