On Wednesday, 16 September at 16:15, Arun Kumar Singh, Professor of Motion Planning, at the University of Tartu, will give his inaugural lecture titled “Beyond Scaling: Building Safe and Data-Efficient Learning Robots” in the White Hall of the UT Museum. The lecture will be in English and the event will be streamed live on UTTV.
Robots are beginning to do things that, until recently, seemed firmly in the realm of science fiction. Humanoid robots can load dishwashers, mobile manipulators can prepare coffee, and autonomous cars can navigate busy streets. Behind many of these advances is a major shift in how robots are built: instead of relying entirely on carefully engineered components for perception, planning, and control, researchers are increasingly training neural networks to learn these behaviours directly from demonstrations and experience. The results can be remarkable, but there is an important difference between succeeding in a short demonstration and operating reliably for hours or days in the real world. Robots in homes, warehouses, factories, or on public roads will inevitably encounter situations that were rare or absent during training, and our own work in navigation and manipulation has shown how quickly learned systems can deteriorate as environments become more cluttered, dynamic, or unfamiliar.
One increasingly popular response is simply to scale up: collect more data, train larger models, and rely on experience to cover more situations. This strategy has transformed language and vision, but physical robots face different constraints. Robot data are expensive and sometimes dangerous to collect, simulations cannot perfectly reproduce the real world, large models can be too slow for real-time control, and increasingly complex policies are difficult to understand and verify when something goes wrong. This talk asks whether robotics needs an alternative to the idea that more data and larger models will solve everything. Professor Arun Kumar Ssingh discusses how the flexibility of modern machine learning can be combined with physical models, optimisation, and explicit safety constraints to build robots that learn and adapt while remaining data-efficient, computationally practical, and reliable in the real world.
Arun Kumar Singh is a Professor of Motion Planning at the Institute of Technology, University of Tartu. He leads the Autonomous and Kooperative Systems (AKS) group and his core focus is on algorithmic foundations of robotics, specifically in regard to motion planning, control and machine learning. Professor Singh and his lab have made impactful contributions in developing optimisation algorithms and learning-based robot planning and control in safety-critical applications using implicit layers of embedded neural networks, and planning under uncertainty. He has been the principal investigator for several national and EU projects with the most recent being XSCAVE (xscave.eu), which seeks to develop deployable AI for heavy machinery.
