Intuitive learning: A new paradigm in AI for decision making in intelligent mobility

Decision making in autonomous systems is crucial for their safe operation. Humans are highly adept at making decisions especially in situations which they have not experienced before or for which outcomes are highly unpredictable. The intuition of humans is built up of multiple levels of abstraction of massive amounts of heterogenous personal and social experiences. This project investigates a new paradigm in AI, named intuitive learning, which attempts to develop models of intuitive decision making. We build on the foundations of deep learning, one-short learning, and meta-learning paradigms to develop intuition algorithms. The application of algorithms is planned for demonstrating for localisation and navigation of robots (using our in-house built UK’s first autonomous quadbike) and multi-agent simulation of accident avoidance.

Investigations initiated by developing deep learning models for visual understanding of robot sensor data, which is the foundation of robot navigation. An industry partner, the PTV Group, is involved for providing access to their simulation platform VSSIM in this project which perfectly aligns with two grand challenges of UK industrial strategy: “to put the UK at the forefront of the artificial intelligence and data revolution” and “to become world leader in shaping the future of mobility”. As such, the proposed research promises a significant advancement in AI that will have large benefits on intelligent mobility applications.

This 4-year project is funded by the UKRI EPSRC’s National Productivity Improvement Fund for a total sum of £89,000 and started in 2018, and is aligned with the Artificial Intelligence, Machine Learning and Data Analytics research theme of the IDT.

As for the notable outcomes of the project to date, we have developed a driverless electric quad bike to test the algorithms, collected data through multimodal sensors, and developed algorithms for scene content recognition using variational autoencoders, multimodal localisation algorithms, and are currently progressing on developing path planning algorithms that combine the advantages of neural networks and symbolic reasoning.

For further info, please contact Dr Varuna De Silva.