Use cases 

Development and validation of autonomous vehicles equipped with multiple sensors in large-scale and rare edge scenarios is critical to ensure safety of autonomous driving systems. However, collecting, annotating scenarios, and compliance with safety regulations are technically and economically demanding.

Multimodal GenAI ecosystem creates a comprehensive benchmark of scenarios with a humancentric approach to ensure fairness and diversity of scenarios with a safety assessment across user groups.

• World model construction
• Specification generation
• Scenario validation and augmentation
• Consistency assurance and output validation

Service robots play an essential role in enhancing operability and efficiency in mixed human-robot environments. Existing robots from factory default are not equipped with GenAI as the “brain”. They quickly max out on their capabilities, causing frustration from the patients. Their upscaling is necessary to improve their coexistence and interaction with patients in nursing homes.

GenAI technology is applied to a low-speed service robot developed to understand the patients and behave harmoniously with them.

The experiments are carried out in nursing homes in Norway. To assess the operational robustness, the operational environment is migrated to a university kitchen in Sweden.

• The robot makes use of Large Language Models (LLM) to act as the “intelligence” of the robot by retrieving the information from the database via Retrieval Augmented Generation (RAG).
• The robot is aware of the knowledge that it shall answer and does not respond to a user with hate or discriminatory speech.
• The robot may guide the user to a place using LLM-generated high-level action sequence plans. Classic safety mechanisms override all actuation functions by signalling an emergency stop.

Cybersecurity is indispensable for HCPS applications. Predicting, analysing, and responding to cybersecurity risks require active and time-consuming engagement of highly skilled cybersecurity analysts.

Automation of correct and timely operations of a Security Operation Centre with human-centric view by using Large Language Models (LLM) with four activities:

• Threat information synthesis
• Intrusion detection rule generation
• Incident analysis reporting
• Response plan generation

• Upon threat discovery, a security analyst prompts the use case engine providing relevant sources of information about the threat.
• A Retrieval Augmented Generation (RAG) mechanism based on LLM enables the retrieval of qualified existing threat synthesis reports and generates a synthesis according to the specified format.
• Whenever vulnerability knowledge becomes available, an analyst prompts the system with enriched information such as Common Vulnerability Exposure (CVE) bulletins, asking the LLM to generate an intrusion detection rule.