Explainable AI and Neural Network Verification for SOTIF Compliance

Demonstrating that AI perception models behave predictably and safely at the boundary of their operational design domain (ODD) is the foundational challenge in SOTIF validation for highway pilot — and explainable AI methods have emerged as the primary technical lever for addressing it. Analysis of over 60 patent filings across Korea, Japan, Taiwan, Europe, and the United States confirms that XAI-based verification is the single most active area of IP development in this space, led by StradVision Inc. with at least five active patents spanning explainable AI verification, multi-source sensor fusion, and trajectory-based mode-switching safety networks.

StradVision’s approach, described in their 2020 patent on functional safety warnings using explainable AI, centres on a verification learning device that extracts attribute information from training images to generate a quality vector. This quality vector is passed through a dedicated neural network to produce predicted safety information. Backpropagation through a loss module trains the verification network to detect when image quality or scene characteristics deviate from conditions under which the primary detection network was validated. A subsequent active Korean grant from 2022 extends this framework by formalizing the extraction of attribute information as a systematic quality-gating mechanism applicable to the inference pipeline of autonomous driving neural networks, with explicit application to functional safety warnings issued to the driver.

By constructing a parallel explainability network that monitors quality vectors at inference time, this architecture provides a runtime mechanism to detect when the primary perception system may be operating outside its validated performance envelope. The Japanese-jurisdictional counterpart filed in 2021 confirms the international deployment of this verification strategy, indicating a deliberate multi-jurisdiction IP protection approach across Korea, Japan, and the US. According to ISO, the SOTIF standard was published in 2022 as ISO 21448:2022, providing the formal framework within which these patent claims operate.

Sensor Fusion, Motion Estimation, and Multi-Source Perception Validation

Highway pilot systems depend on the continuous, reliable estimation of surrounding vehicle states — position, velocity, and trajectory — over extended time horizons at highway speeds, and SOTIF validation requires that these estimation pipelines be tested across the full range of sensor degradation scenarios, including adverse weather, sensor occlusion, and communication dropout. The patent record shows a technically mature set of approaches for achieving this, with StradVision’s four-neural-network fusion architecture and Autonomous A2Z’s multi-modal sensing extensions representing the current frontier.

StradVision’s 2021 patent on learning methods for safe autonomous driving describes a four-neural-network architecture that separately processes image-based and V2X communication-based feature maps, combines them through an integration network, and generates predicted surrounding motion information. The training framework uses a loss layer to jointly optimize all four networks, ensuring that the integrated perception output is robust to single-source dropout. The earlier 2020 Korean filing confirms the foundational architecture, while the Japanese counterpart demonstrates the cross-jurisdictional IP strategy being pursued.

“A four-neural-network architecture that jointly optimises image-based and V2X feature maps through a shared loss layer ensures that integrated perception output remains robust to single-source dropout — directly addressing SOTIF’s sensor degradation coverage requirement.”

For vehicle type classification — a prerequisite for accurate motion prediction in highway scenarios — Autonomous A2Z’s 2025 Korean patent describes a shared encoder architecture that produces feature maps for simultaneous car-type estimation, orientation prediction, and distance estimation from detection images. A companion patent extends the sensing modality to acoustic cues through time-series sound data fusion, broadening the SOTIF-relevant coverage of perception edge cases beyond what camera and V2X alone can provide.

From a vehicle control safety perspective, SOKEN’s 2025 patent describes a diagnostic device with an abnormality detection unit that monitors the output signal of surrounding monitoring sensors and recognition results sequentially in time. When an anomaly is detected in the recognition system, the device triggers a Minimal Risk Maneuver (MRM) — a direct implementation of SOTIF’s requirement for safe state transitions when the AI perception system’s reliability cannot be guaranteed. This time-sequential monitoring architecture represents a key implementation pattern for highway pilot systems operating under SAE Level 3 conditions, where the driver is not continuously monitoring the road. Standards bodies including SAE International and ISO have established the Level 3 automation context within which MRM requirements apply.

Field Performance Monitoring and Ground Truth Generation

Establishing ground truth for AI perception performance in real-world operational conditions is one of the most operationally challenging aspects of SOTIF validation — and the patent record reveals three distinct architectural approaches to solving it: retrospective time-window ground truth, closed-loop self-updating scene databases, and bidirectional simulation feedback loops. Each addresses a different aspect of SOTIF’s requirement to continuously narrow the unknown-unsafe operational space.

ZENUITY AB’s active 2025 European patent presents a field monitor system that stores a range of perception data over a defined time window and then uses post-processing of a sub-window of recent data to generate ground truth intermediate data at an intermediate time point. This retrospective ground truth reflects the system’s estimates of vehicle surroundings conditioned on subsequent observations — a technically sound approach for validating whether the perception system’s real-time estimates were accurate without requiring simultaneous external reference sensors. This architecture is well-suited to supporting SOTIF’s field monitoring requirements and enabling statistical accumulation of performance evidence across operational miles.

The Taiwan-jurisdictional 2023 patent from Vehicle Research and Test Center explicitly invokes SOTIF terminology — making it the only organisation with an explicit SOTIF label in a patent title in the dataset. It describes a system that classifies sensor and controller output data as either unintended dissociation data or accident scene data, builds a scene database from these events, and uses the database to correct sensor and controller algorithms. A parallel computation then verifies the corrected outputs by comparing them against a dissociation control command benchmark. This closed-loop self-update mechanism directly addresses SOTIF’s requirement for continuous improvement of the system’s ability to handle previously unknown triggering conditions.

IBM Corporation’s 2025 Japanese pending patent formalizes the distinction between intentional test space (quantifiable safety criteria) and unintentional test space (characteristics that cannot be pre-specified), using a coupled simulation of surrounding and autonomous vehicles to feed unintentional test outcomes back into the intentional test framework. This bidirectional feedback loop maps directly to SOTIF’s process of iteratively narrowing the unknown-unsafe operational space, and represents the most formally structured approach to SOTIF test space partitioning identified in the dataset. Research published by Nature on AI safety evaluation methodologies corroborates the importance of explicitly modelling the boundary between quantifiable and unquantifiable test characteristics in safety-critical AI systems.

SOTIF Residual Risk Quantification and Reasonably Foreseeable Misuse

Quantifying residual risk arising from reasonably foreseeable indirect misuse (RFIM) events is a critical SOTIF process step for highway pilot features — and Robert Bosch GmbH is the only identified assignee in the dataset with an explicit algorithmic SOTIF residual risk computation method. This is a significant IP differentiator in a landscape where most patent activity addresses detection and monitoring rather than formal risk acceptance.

Robert Bosch GmbH’s 2023 Japanese pending patent explicitly addresses RFIM. The patent describes receiving signals on driver condition to determine RFIM events — including driver over-reliance, improper mode engagement, and failure to resume manual control upon system request — combining these with specific system events or failures, and calculating a failure rate as a function of event severity and RFIM frequency. The method then determines whether the residual risk related to safety (SOTIF) of the intended function is acceptable based on this combined failure rate. This is among the most explicit SOTIF-aligned patent claims found in the dataset, and represents a direct algorithmic instantiation of SOTIF’s residual risk acceptance criteria.

Continental Teves AG’s 2014 patent establishes an earlier foundational concept: placing the assistance system in a safe state when image data reveals a restriction — effectively a predecessor to SOTIF’s safe-state transition concept triggered by perception system limitation detection. Autoliv Development AB’s 2018 and 2019 patents extend this by comparing real-time surroundings images against a library of pre-stored images, each tagged with hazard ratings and safety mechanism activation decisions. Vehicle behavior detected in association with matching images informs whether to adjust safety mechanism operation, providing a mechanism to validate that safety mechanisms are appropriately calibrated across the range of triggering conditions relevant to SOTIF. The UNECE regulatory framework for automated vehicles, including WP.29 Working Party guidelines, provides the broader regulatory context within which these SOTIF residual risk acceptance methods must demonstrate compliance.

“Robert Bosch GmbH’s patent is the only identified filing with an explicit algorithmic SOTIF residual risk computation method — combining RFIM event frequency with system failure severity to evaluate whether residual risk is acceptable. This is a significant IP differentiator in the SOTIF validation landscape.”

Scenario-Based Test Evaluation Systems

Structured scenario-based testing is a cornerstone of SOTIF validation evidence generation, requiring that the autonomous system demonstrate acceptable performance across a catalog of triggering conditions — and CNBIS, a Korean firm, has filed the most detailed suite of patents formalizing this infrastructure, with road geometry specifications precise enough to be directly implemented in physical test facilities.

CNBIS’s 2025 Korean patent on test evaluation for autonomous car driving based on other car access on traffic yellow signal lamp at a crossroads describes a road structure with precisely defined lane widths of 3.00–3.50 m, minimum intersection entry distances of 100 m from stop line, and paired vehicle detection sensors that provide triggering and evaluation signals synchronized to the behavior of a control vehicle executing a defined event sequence. Similar architectures appear in their left-turn yellow signal lamp test patent and their bicycle access at a crossroads without traffic light test patent, both also filed in 2025.

These structured test systems address the SOTIF requirement for systematic generation of test scenarios that cover both the intended behavior and its triggering conditions. The use of control vehicles executing scripted events, detected by road-mounted sensors, allows repeatable, parameterized scenario execution — critical for statistical accumulation of test evidence across a scenario catalog. This industrialization of SOTIF’s scenario catalog-based test methodology represents a significant development in the IP landscape, as it moves SOTIF validation from a process description into a hardware-software system that can be commercialized and standardized.

Baidu’s approach to coping capability boundary determination uses historical driving behavior data in preset driving scenes, combined with indicator combinations including expected value ranges, to establish coping capability boundary information. This boundary mapping, described in their 2022 Japanese patent, is directly analogous to SOTIF’s process of characterizing the performance limits of the intended functionality across the ODD — and represents the most systematic approach to ODD boundary quantification identified in the dataset from a Chinese assignee.

Key Players, IP Concentration, and What the Landscape Means for Safety Engineers

The SOTIF validation patent landscape is concentrated but not dominated by any single player across all technical dimensions — creating both freedom-to-operate opportunities and clear areas of IP risk for organisations developing highway pilot systems. Understanding the IP positions of the key assignees is essential for R&D leads and patent attorneys navigating this space.

StradVision Inc. — Broadest Portfolio in XAI and Perception Safety

StradVision is the most active assignee in the dataset for AI perception functional safety, with at least five active patents spanning explainable AI verification, multi-source sensor fusion, and trajectory-based mode-switching safety networks. Their 2022 Korean patent on trajectory-plus-state-vector collision probability networks introduces a mode-transition gating mechanism that evaluates whether switching from manual to autonomous driving is safe based on surrounding vehicle trajectories and state vectors — a direct implementation of SOTIF’s ODD entry validation requirement. The multi-jurisdiction filing strategy across Korea, Japan, and the US suggests a deliberate IP protection programme rather than opportunistic filing.

ZENUITY AB — Technical Maturity in Fleet-Scale Field Monitoring

ZENUITY AB’s retrospective ground truth generation approach, active in the European Patent Office as of 2025, is technically sophisticated and operationally practical. By eliminating the need for real-time external reference sensors, it enables statistical accumulation of SOTIF performance evidence at scale — a requirement that becomes increasingly important as regulators move toward evidence-based type approval for Level 3 and Level 4 systems. The EASA and European automotive regulatory framework increasingly demands fleet-scale performance evidence for advanced automated systems, making ZENUITY’s approach strategically well-positioned.

Robert Bosch GmbH — Unique Position in Residual Risk Computation

Bosch’s explicit SOTIF residual risk computation patent, combined with their active grant on driver state communication during autonomous operation (2024), positions them as the only identified assignee with both the risk quantification method and the human-machine interface notification mechanism required for a complete SOTIF compliance architecture. This combination of IP assets is strategically significant for Tier 1 suppliers and OEMs seeking a licensable SOTIF compliance stack.

Emerging Trends: Audio Fusion, Formal Test Space Partitioning, and Self-Updating Databases

Three emerging trends in the dataset deserve particular attention from R&D leads. First, Autonomous A2Z’s 2025 incorporation of time-series sound data fusion into the perception pipeline extends SOTIF-relevant edge case coverage beyond camera and V2X — a relatively unexplored sensing modality in the current IP landscape. Second, IBM’s formal partitioning of intentional and unintentional test spaces provides a methodological framework for incorporating non-quantifiable characteristics into SOTIF safety evaluations — a gap that existing standards guidance does not fully address. Third, Taiwan VRTC’s self-updating scene database demonstrates how regulatory test centres are beginning to build SOTIF-specific IP, potentially influencing future standardisation and type approval workflows. The WIPO patent database confirms the cross-jurisdictional nature of this filing activity, with Korea, Japan, Europe, and Taiwan all showing active SOTIF-adjacent prosecution.