PhD-Students

The tools available at our disposal for solving some of the really hard problems our society is facing today are old, inadequate, and need to be deprecated. Everyone is looking to Artificial Intelligence based solutions like it’s the next gold rush, without understanding how Machine Learning actually works. We’re pumping massive amounts of data into these systems, without realizing that this data came before Artificial Intelligence was a thing, and even before the Internet or computers existed: 2D images are actually just digital representations of something that was available before in analog form. Someone needs to pause, zoom out, and take a look first and foremost at the problems that we need to solve, identify and analyze them, and only then derive complete technical solutions that might or might not involve the current generation of Machine Learning en vogue, but most importantly, might indicate that new types and formats of data, whether visual or otherwise, need to be created. And the field that made significant progress there is robotics, where mapping and identifying the world with high accuracy was an absolute requirement for the stability and performance of a machine moving into our world. However, these concepts such as 3D visual representations have not yet been translated fully to scale and made de facto standards for other more common applications. In this talk, we’re taking a trip down memory lane at some of these concepts, and discussing how open source platforms such as the Point Cloud Library (PCL) have contributed to the proliferation of new visual understanding technologies. We’re also taking a look at Fyusion, which attempts to redefine the meaning of “scalable 3D visual formats”, and has created the first comprehensive and scalable technology stack for capturing photorealistic 3D spatial models of the real world using a single camera, built with Visual Understanding in mind.

In this talk, I review the Soar cognitive architecture, including the motivation for cognitive architecture, the history of Soar, the applications it has been applied to, and our current research on Interactive Task Learning. I then discuss Soar from the standpoint of an open-source research project: positives, negatives, and challenges.

In his captivating lecture, Frank Guerin examines the question about the integration between robot vision and a deeper understanding of tasks. As robot grasping is still an unsolved problem, he explores why and how human perception of objects is relevant to manipulation and explains what "transferrable toddler skills" are. The lecture is suitable for beginners.

From small to complex, from robot vacuum cleaner to self-driving car: every robotic system needs some sort of perceptual capabilities in order to perceive information from its environment and to understand how it can manipulate it. Perception can come in many forms. Tim Patten gives a highly interesting introduction on how robots deal with object identification: what is it? (recognition), what type is it? (classification), where is it? (object detection), and how do I manipulate it? (grasping). The talk is suitable for beginners.

Part 2: In his equally interesting follow-up lecture, Animesh Garg continues to explore compositional planning and multi-step reasoning, i.e. when a robot is supposed to do multiple tasks in a certain structure. He also examines robot perception via structured learning through instruction videos, and tackles the question of how to collect the data required for robot learning.