Kingston AI Symposium 2026

Professor Dana Kulić, Monash University

As robots and other embodied AI systems become more capable and autonomous, how do we determine the most appropriate roles for humans and robots during human-robot teaming?  With the advent of deep learning, large language and foundation models, human-robot collaboration has the potential to become much easier, more natural and detailed, but may also introduce new sources of misunderstanding and error.  In this talk, Professor Dana Kulić will describe her research analysing human-robot teams and employing deep learning and foundation models to facilitate human-robot teaming while enhancing human capabilities and reducing errors.

Professor Anton van den Hengel, AIML, Adelaide University

Transformers, like so many neural networks, are commonly initialised at random because we don't know what structure we actually want in the initial weights. Random initialisation is easy, and safe to the extent that it is unlikely to introduce unwanted bias.  It represents a maximally unstructured starting point for gradient descent's search for the computational scaffolding it needs to begin learning semantics. This talk describes work showing that brief exposure to specific, abstract, procedurally generated data injects beneficial structural inductive biases directly into transformer weights before any semantic training begins.

These biases are modular to the extent that they reside in identifiable architectural components, can be transferred independently, and compose across pretraining rules to jointly improve multiple capabilities. This process is modality-agnostic: a vision transformer warm-started on sequences of balanced brackets, and no images, gains +1.7% final performance on ImageNet-1k, outperforming baselines trained on visual synthetic data. We show that front-loading as little as 0.1% procedural tokens reduces the language modelling data budget by up to 45%. Pretraining on procedural data is thus far more effective than training on real data.