Project: CTRL schemes merged with eXplainable AI for t(h)rustworthy control of physical dynamic systems (CTRLxAI=T(H)RUST)
2022-10-01 – 2026-09-30
- Abstract
As today’s industrial processes become more complex, controllers used in drivetrains for vehicles, machines, robots, process facilities, and other physical dynamic systems face increasing challenges with respect to e.g. efficiency and quality. In an industry 4.0 setting, a higher level of adaptivity and automation is required. Meanwhile, artificial intelligence (AI) is a promising enabling technology. However, examples wherein AI techniques such as reinforcement learning (RL) are directly in control of (high) power (up to kW) and (highly) dynamic (down to (m)s)) physical systems to improve energy efficiency and performance remain very limited. Going beyond the fixed but safe structure of classical controllers and embracing the RL framework provides the ability to learn and adapt. While doing so, expensive trials and unsafe experimentation on real
systems as is common in RL need to be avoided. We therefore propose a fundamentally new approach residing at the intersection of classical control and RL (CTRLxAI). Besides offering increased efficiency and performance (thrust) of the adaptive and autonomous controllers; we will strengthen the trustworthiness (trust) in terms of sample-efficiency, robustness, safety and explainability; critical capabilities for industrial widespread adoption. As such, we will realise our vision CTRLxAI=T (H)RUST. CTRLxAI will focus profoundly on pioneering concepts going beyond the scientific state of the art tackling relevant challenges inspired by the companies in the industrial advisory board. To enable this future utilisation the pioneering CTRLxAI results will be validated up to TRL4-5. Utilisation of the project results by Flemish companies will enable them to increase their competitiveness as well as lower i.a. their production footprint, lowering CO2 and other greenhouse gas emissions per capita contributing to SDG13.
-
A probabilistic treatment of (PO)MDPs with multiplicative reward structure
-
- Journal Article
- A1
- open access
Probabilistic control and majorisation of optimal control
-
- Journal Article
- A1
- open access
Integrated barometric pressure sensors on legged robots for enhanced tactile exploration of edges
-
Dual regularized policy updating and shiftpoint detection for automated deployment of reinforcement learning controllers on industrial mechatronic systems
-
- Conference Paper
- C1
- open access
Probabilistic majorization of partially observable markov decision processes
-
- Journal Article
- A1
- open access
A posteriori control densities : imitation learning from partial observations
-
- Journal Article
- A1
- open access
Optimizing cascaded control of mechatronic systems through constrained residual reinforcement learning
-
- Journal Article
- A1
- open access
Information-theoretic policy learning from partial observations with fully informed decision makers