AI Safety and Explainability for Robots and Intelligent Systems at MSSiS 2025
- Aug 2
- 2 min read
Dr Koorosh Aslansefat delivered a keynote on “AI Safety and Explainability for Robots and Intelligent Systems” at the VII Workshop on Modelling and Simulation of Software-Intensive Systems (MSSiS 2025).
Dr Aslansefat is Co-Lead of the Dependable Intelligent Systems (DEIS) Research Centre and Deputy Lead of the Centre for Responsible Artificial Intelligence at Hull. His keynote presented recent advances from SafeML and SMILE, two complementary frameworks developed to strengthen confidence in intelligent and robotic systems.
Machine-learning models are normally trained and tested using a defined collection of data. Once deployed, however, they can encounter changing environments, unfamiliar inputs and new patterns of behaviour. These changes can make a model’s earlier performance evidence less representative of what it is doing now.
SafeML addresses this problem through runtime safety monitoring. It compares operational data with the conditions represented during development and helps detect concept or distribution shifts. An early warning that a system is moving outside its expected operating conditions can support intervention before unreliable behaviour develops into an unsafe outcome.
Monitoring tells engineers that something may have changed; explanation helps them understand the resulting decision. SMILE provides model-agnostic explainability, allowing its methods to be applied without depending on the internal structure of a particular machine-learning model. This helps engineers investigate which factors influenced an output and assess whether the reasoning is acceptable in context.
For robots and other intelligent systems, these capabilities are closely connected. A dependable system needs to recognise when its operating environment has changed, identify potentially unsafe behaviour and provide evidence that humans can interpret. Runtime monitoring and explainability therefore contribute different but mutually reinforcing forms of assurance.
The keynote also demonstrated the value of bringing safety-engineering research into the modelling and simulation community. Models and simulations help teams explore system behaviour before and during deployment, while monitoring and explanation connect those analyses with evidence from real operation.
DEIS thanks Dr André Luiz de Oliveira for chairing the session and the MSSiS 2025 organisers for creating a valuable forum for discussion.
Original LinkedIn update:
Comments