1st Human-Centric AI Webinar: Advances in Constraints Abiding Deep Reinforcement Learning

Date: 25 February 2026, 14:00 – 15:00 (CET)
Location: Online (Teams)
Participation is free. Secure your spot here (registration closed).

IThis webinar is part of the ENFIELD Human-Centric AI webinar series, which showcases research from projects on improving human-AI interaction, transparency, and oversight in AI-based decision-making. In this first session, researchers from the CARE@ENFIELD project present their work on safe deep reinforcement learning. Their approach addresses a core challenge in deploying AI for real-world decision support: how to ensure that AI agents reliably satisfy operational constraints while remaining transparent about how individual decisions contribute to constraint adherence. The talk includes benchmarking results against state-of-the-art methods and a demonstration on unmanned surface vessels.

Title: Advances in Constraints Abiding Deep Reinforcement Learning

Abstract: This talk will present advances in Deep Reinforcement Learning (DRL) towards solving Constrained Markovian Decision Processes (CMDPs). Specifically, with the aim to increase the ability of safe DRL methods to satisfy constraints in expectation, we will present a stepwise formulation of the objective, that in a greater extent than other works aims to assess the criticality of each step towards adhering to constraints. This objective results into a generic framework for stepwise optimization, providing benefits for operational transparency. We will present comparative results of an instantiation of that framework with state of the art safe DRL algorithms, as well as results from Unmanned Surface Vessels showing the strengths of our proposal. The work has been completed in the context of the CARE@ENFIELD project. ence is being used to make energy systems more efficient, transparent and resilient, while supporting the transition towards sustainable mobility.

Participation is free!
When: 25 February 2026, 14:00-15:00 (CET)
Where: Online
To register, please use the following page.

The webinar is organized in the framework of the Horizon Project ENFIELD - European Lighthouse to Manifest Trustworthy and Green AI – and it will take place on the Teams platform.

Registration Closed

Target audience

The target audience includes AI researchers, practitioners, and students with an interest in reinforcement learning, safe AI, trustworthy AI, or human-centric AI. Also relevant for professionals working on autonomous systems, robotics, or safety-critical AI applications.

Speakers