From delivering food to your door, serving us coffee and even removing cancerous tumors, robots can already complete a range of impressive tasks. But now a robot has taken on the golf course, being able to navigate itself to a ball and even sink a putt. Thanks to a 3D camera, the impressive robot dubbed Golfi can find golf balls and wheel itself into place before taking a shot. The camera uses an algorithm to detect hard-coded objects, scan the area and find the ball. Golfi (pictured) was created by the Paderborn University in Germany.

A robot called Golfi is the first to be able to autonomously spot and travel to a golf ball anywhere on a green and sink a putt. Golf-playing robots have been developed before, but they have needed humans to set them up in front of a ball and program them to make the correct swing. The most famous is LDRIC, a robot that hit a lengthy hole-in-one at Arizona's TPC Scottsdale golf course in 2016. In contrast, Golfi, engineered by Annika Junker at Paderborn University in Germany and her colleagues, can find golf balls and wheel itself into place thanks to input from a 3D camera that looks down on a green from above. The camera scans the green and an algorithm then approximates the surface before simulating 3000 golf swings towards the hole from random points, taking into account factors such as the speed and weight of the ball and the friction of the green, which are described by physics-based equations.

A robot called Golfi is the first to be able to autonomously spot and travel to a golf ball anywhere on a green and sink a putt. Golf-playing robots have been developed before, but they have needed humans to set them up in front of a ball and program them to make the correct swing. The most famous is LDRIC, a robot that hit a lengthy hole-in-one at Arizona's TPC Scottsdale golf course in 2016. In contrast, Golfi, engineered by Annika Junker at Paderborn University in Germany and her colleagues, can find golf balls and wheel itself into place thanks to input from a 3D camera that looks down on a green from above. The camera scans the green and an algorithm then approximates the surface before simulating 3000 golf swings towards the hole from random points, taking into account factors such as the speed and weight of the ball and the friction of the green, which are described by physics-based equations.

Robots that can whack a golf ball down a fairway aren't exactly new, but building one that can play the nuanced short game is a more complex problem. Researchers at Paderborn University in Germany have done just that with Golfi, a machine that uses a neural network to figure out how to line up a putt and how hard to hit the ball to get it into the hole from anywhere on the green. The robot takes a snapshot of the green with a Microsoft Kinect 3D camera and it simulates thousands of random shots taken from different positions. It takes factors like the turf's rolling resistance, the ball's weight and the starting velocity into account. Paderborn doctoral student Annika Junker told IEEE Research that training Golfi on simulated golf shots takes five minutes, compared with 30-40 hours were the team to feed data from real-life shots into the system.

Voice assistants, smart homes, or industrial 4.0 systems: artificial intelligence (AI) is increasingly automating processes in a wide variety of living and working environments. However, AI systems often prove to be not particularly competent because they lack either background or contextual knowledge, are unable to assess the scope and implications of assumptions and decisions, and cannot explain their actions. In the Joint Artificial Intelligence Institute (JAII), the two universities at Bielefeld and Paderborn are combining their research competencies in this field of research. The universities jointly founded the institute on July 14, 2020. In the JAII, future research will address the fundamentals of AI systems designed to focus on people.

Information-rich representations of text often decrease sample complexity when an natural language processing (NLP) system is trained on a task. One effective way of producing such representations is the traditional NLP pipeline: tokenization, tagging, parsing etc. An alternative are so-called embeddings that represent text in a high-dimensional real-valued space that is smooth and thereby supports generalization. Most commonly, words are represented as embeddings, but more recently contextualized embeddings like ELMo have been proposed. I will address two challenges for embeddings in this talk.