Die künstliche Intelligenz hat in letzter Zeit gewaltige Fortschritte gemacht. Sie bietet immer neue Möglichkeiten in der Analyse und Generierung natürlicher Sprache und von Bildern und Video, in der Robotik, sowie in ihren Kombinationen. Im ersten Teil des Vortrags werde ich informell Kerntechnologien aktueller KI vorstellen, um ein Gespür dafür zu vermitteln, wofür sich KI in der Robotik mittelfristig einsetzen lassen wird. Anschließend werde ich Grenzen der aktuellen KI charakterisieren und einige Beispiele aus der Forschung beschreiben, die darauf abzielen, diese Grenzen zu erweitern.

Maschinelles Lernen hat vielfältige Anwendungen in der Robotik. Roboter können lernen, Trajektorien abzufahren, auf Effekte und äußere Ereignisse zu reagieren, Objekte zu erkennen und zu manipulieren. In die Praxis halten diese Entwicklungen jedoch nur langsam Einzug. Warum ist das so? Ich argumentiere, dass Robotern - und heutigen KI-Systemen allgemein - echtes Verständnis für ihre Umgebung fehlt, mit der sie über Datenaustausch interagieren. Dies werde ich anhand einiger Beispiele aus unserer Forschung illustrieren, mit der wir versuchen, das Verständnis von Robotern für ihre Umgebung zu erhöhen.

The title of this talk is the motto of the Intelligent and Interactive Systems group at the University of Innsbruck. Our long-term objective is to improve abstraction, generalization, robustness, and ultimately explainability of robot perception and action. Motivated by shortcomings of current machine-learning methods, I will discuss examples of our recent work on learning visual relational concepts, extrapolation of learned movements beyond the training distribution, and learning of symbolic concepts and rules from sensorimotor experience. As an added bonus, I will show how to train neural networks of a thousand layers without skip connections, batch normalization, dropout, or any other architectural tweaks, using our so-called batch-entropy regularization.

Justus Piater lehrt seinem Roboter das Lernen – mit Erfolg. So lernte der Roboter, wie sich ­ Objekte bewegen, wenn er sie anschubst, und wie er diese stapeln kann. Nun soll er sich aus gelernten ­Bewegungen neue erschließen und mit dem Menschen interagieren lernen.

Nachhaltigkeit - Wenn Roboter in menschlichen Umgebungen arbeiten, dann brauchen Sie soziale Fähigkeiten. Forscher wollen es ihnen beibringen und dabei das tatsächliche Lernverhalten von Menschen nachbilden.

The motto of my research group is "Making robots learn to perceive and act with understanding". Our long-term objective is to improve abstraction, generalization, robustness, and ultimately explainability of robot perception and action. I will first present three of our recent lines of research: continual learning from demonstration, how to train 1000-layer neural networks with just a simple regularizer, and autonomous learning of symbolic concepts and rules from sensorimotor interaction. In the second half of my talk I will argue that "understanding" is a meaningful notion in AI that goes beyond prediction and control, and hypothesize that causlity theory is adequate to formalize it. My goal is to launch a discussion on how to leverage causality theory to raise the level of understanding that robots can acquire by interacting with their environment.

The title of the talk is the motto of the Intelligent and Interactive Systems group at the University of Innsbruck. The long-term objective of the group is to improve abstraction, generalization, robustness, and ultimately explainability of robot perception and action. Motivated by shortcomings of current machine-learning methods, Prof. Piater will discuss examples of the group's recent work on learning visual relational concepts, extrapolation of learned movements beyond the training distribution, and learning of symbolic concepts and rules from sensorimotor experience. As an added bonus, Prof. Piater will show how to train neural networks of many hundreds of layers without skip connections, batch normalization, dropout, or any other architectural tweaks, using the so-called batch-entropy regularization.

Recycling obsolete electronic devices (E-waste) is a dangerous task for human workers. Automated E-waste recycling is an area of great interest but challenging for current robotic applications. We focus on the problem of segmenting inner parts of E-waste devices into manipulable elements. First, we extend a dataset of hard-drive disk (HDD) components with labelled occluded and non-occluded points of view of the parts, in order to increase the diversity and the quality of the learning data with different angles. We then perform an extensive evaluation with three different state-of-the-art models, namely CenterMask, BlendMask and SOLOv2 (including variants) and two types of metrics: the average precision as well as the frame rate. Our results show that instance segmentation using state-of-the-art deep learning methods can precisely detect complex shapes along with their boundaries, as well as being suited for fast tracking of parts in a robotic recycling system. (Rojas et al. 2022)

Künstliche Intelligenz (KI) ist bereits heute ein täglicher Begleiter, doch vom universellen Haushaltsroboter sind wir noch weit entfernt. Automatisierte Systeme werden zwar Jobs übernehmen, Maschinen werden aber noch lange nicht die Weltherrschaft an sich reißen, falls überhaupt. Zur Langen Nacht der Forschung spricht Martin Kubin im #derWoche „Maschinenzukunft“ mit Prof. Justus Piater vom Institut für Informatik über Chancen und Gefahren von KI. Er weiß auch, warum KI etwa die Justiz unterstützen kann, aber ein KI-Richter niemals objektiv urteilen könnte.

In the design of "Cognitive Robots" software architectures, researchers had to face many organizational problems due to the particular nature of robots. Not only should they manage to achieve complex tasks with long-term goals, but they must also be reactive to their surroundings to preserve their environment. We will go over several layered control paradigms developed over the history of robotics and see how these can relate to the knowledge about the Octopus' brain organization, and discuss how the Octopus itself can inspire cognitive robotic research.

Wouldn't it be great to have a helper robot that we can task with our annoying chores? How would we teach the robot to perform these tasks? Can we build such robots by equipping them with artificial intelligence? We will discuss how human and machine learning differ and how this impacts what skills, including for communication, our helper robot will be able to learn.

Humans constantly engage in creative activity such as finding explanations, improving technology, or solving novel problems in everyday life. Even imitating other humans, e.g. while learning a new skill, is creative in that the imitator has to infer the conceptual structure underlying the observed manifestation and to regenerate the latter from the former. In contrast, current AI-enabled agents are unable to generate and draw upon such conceptual structures. Their learning is mostly driven by statistics, and any ability to produce novelty is largely limited to trial and error. I will try to characterize this fundamental difference between learning in human and artificial agents, and will discuss possible technological approaches that might help close this gap.

Growing data processing capacities and progress in analytical methods and artificial intelligence are motivating entire branches of science to raise new questions and to develop new approaches. This digital transformation requires progressive, interdisciplinary synergies between the computational sciences and all other scientific disciplines. The Digital Science Center (DiSC) at the University of Innsbruck meets this challenge by integrating and promoting digitalisation of scientific research. In this talk I will present some examples of interdisciplinary research by DiSC staff researchers who are experts in specific aspects of digital science.

With every spectacular achievement of a machine learning system, the long-elusive AI breakthrough is popularly proclaimed to be just around the corner. Most recent successes have been due in large part to massive data and computation, in particular using deep artificial neural networks. But can artificial cognition really be achieved just by further scaling up existing machine-learning techniques? I discuss examples of simple, perceptual problems that are easily solved by humans but very difficult for today's machine learning methods. These problems reflect how humans conceptualize their world. Their mastery is thus likely to be an essential prerequisite for autonomous robots to attain higher levels of cognitive abilities. To get there, a few core issues can be identified that should drive research in cognitive robotics.

Robot learning is all the rage, but what does it take to get robots to learn what we ultimately want them to do? I will give a high-level characterization of some strengths and weaknesses of current machine learning for robotics, and will try to identify areas that require particular attention.

Machine Learning increasingly equips robots with learning capabilities and flexibility. This will enable them to act purposefully in unstructured environments and to react to unforeseen events. People can teach them intuitively to perform tasks instead of having to program them in painstaking ways. Robots can learn from experience and can improve their behavior over time. In this talk I will give an overview of methods, opportunities, and challenges of machine learning in robotics.

Conditional Neural Movement Primitives (CNMP) constitute a novel framework for robot programming by demonstration based on Conditional Neural Processes (CNP). Like Bayesian methods such as Gaussian Processes (GP), CNP learn how target distributions depend on data, and can be conditioned on specific data points to infer new target distributions at test time. Unlike GP that are expensive to train and scale poorly to high dimensions, CNP are neural networks and are trained by gradient descent. CNMP leverage CNP to represent motion trajectories that can be conditioned, at test time, on task paramters such as goal locations, via points, and/or force readings. Moreover, CNMP are conditioned on sensor readings during execution, resulting in robust, reactive behavior. This talk will present an overview of how CNMP work and how they can be used in various robot applications.

With every spectacular achievement of a machine learning system, the long-elusive AI breakthrough is popularly proclaimed to be just around the corner. Most recent successes have been due in large part to massive data and computation, in particular using deep artificial neural networks. But can artificial cognition really be achieved just by further scaling up existing machine-learning techniques? I discuss examples of simple, perceptual problems that are easily solved by humans but very difficult for today's machine learning methods. These problems reflect how humans conceptualize their world. Their mastery is thus likely to be an essential prerequisite for autonomous robots to attain higher levels of cognitive abilities. To get there, a few core issues can be identified that should drive research in cognitive robotics.

Massive availability of data and computing power are promoting data-driven methods in all areas of science and technology. I will describe how the University of Innsbruck supports this via its new Digital Science Center, and will give a flavor of machine learning for data analysis.

A fundamental task of sensory processing is to detect and integrate feature items to group them into perceptual units segregating them from other objects and the background. A framework is discussed which explains how perceptual grouping at early as well as higher-level cognitive stages may be implemented in cortex. Different grouping mechanisms are implemented which are attuned to basic features and feature combinations and mainly evaluated along the forward sweep of stimulus processing. However, due to limitations of local feature detection mechanisms and inherent ambiguities, top-down feedback is required to deliver contextual information helping to disambiguate initial measurements. Feedback of contextual information is demonstrated to improve object recognition performance, stabilize learning of object categories, and integrate multi-sensory representations. The canonical principles of neural computation define a set of core operations to implement above-mentioned mechanisms of perceptual and cognitive inference. These operations can be mapped, in a simplified form, onto neuromorphic platforms to emulate brain-like computation. It is demonstrated that an architecture composed of canonical circuit mechanisms can be mapped onto neuromorphic chip technology facilitating low-energy non-von Neumann computation.

Humanoid robotics plays a central role in robotics research as well as in understanding intelligence. Engineering humanoid robots that are able to learn from humans and sensorimotor experience, to predict the consequences of actions and exploit the interaction with the world to extend their cognitive horizon remains a research grand challenge. Currently, we are experiencing AI systems with superhuman performance in games, image and speech processing. However, the generation of robot behaviors with human-like motion intelligence and performance has yet to be achieved. In this talk, I will present recent progress towards engineering 24/7 humanoid robots that link perception and action to generate intelligent behavior. I will show the ARMAR humanoid robots performing complex grasping and manipulation tasks in kitchen and industrial environments, learning actions from human observation and experience as well as reasoning about object-action relations.

I will present present a hierarchical reinforcement learning agent that is intrinsically motivated to learn how to control its observation space in the fastest possible manner by optimizing learning progress. The agent mimicks to some extend the development process how a baby learns. Our agent learns what can be controlled, how to allocate time and attention, and the relations between objects in the environment. In a nutshell, our work combines several task-level planning ideas (e.g., backtracking search on task graph, probabilistic road-maps, allocation of search efforts, etc.), in the form of structured prior, with deep RL control and relational reasoning to learn from scratch.

I am studying the question how robots can autonomously develop skills. Considering children, it seems natural that they have their own agenda. They explore their environment in a playful way, without the necessity for somebody to tell them what to do next. With robots the situation is different. There are many methods to let robots learn to do something, but it is always about learning to do a specific task from a supervision signal. Unfortunately, these methods do not scale well to systems with many degrees of freedom, except a good prestructuring is available. The hypothesis is that if the robots first learn to use their bodies and interact with the environment in a playful way they can acquire many small skills with which they can later solve complicated tasks much quicker. In the talk I will present my steps into this direction. Starting from some general information theoretic consideration we provide robots with an own drive to do something and explore their behavioral capabilities. Technically, this is achieved by considering the sensorimotor loop as a dynamical system, whose parameters are adapted online according to a gradient ascent on an approximated information quantity. I'll show examples of simulated and real robots behaving in a self-determined way and present future directions.

Current robot learning, and machine learning in general, requires carefully-engineered setups (environments, objective functions, training data, etc.) for learning to succeed. Perception and action spaces are specially crafted to meet the requirements of the learning objective, which is specified in advance. How can we construct robot learning systems that can learn in an open-ended fashion, acquire skills not foreseen by its designers, and scale up to virtually unlimited levels of complexity? I argue that a key to achieving this lies in the robot's ability to learn abstract concepts that can be reused as a basis for future learning, both in autonomous exploration and for teaching by humans.

Human-made objects are constructed with functional parts. These are closely related to affordances; e.g., a handle is a part whose function is to be grasped and that affords grasping. It is these affordances that robots need to detect to perform tasks flexibly using unknown objects. Since affordances are often associated with parts, it makes sense for the segmentation of unknown objects into parts for affordance detection to be driven by known affordances. Then, the suitability of an object for a given task is often characterized by the combination of the actions its various parts afford. For example, to drive a nail into the wall you would like an object that possesses a part that affords pounding and another that affords grasping. I will discuss methods that allow robots to learn to segment objects in terms of their affordances, and to learn the suitability of objects for specific tasks as a function of their combination of affordances.

The title of this talk is the motto of my research group. In this talk, I describe our high-level agenda and present some representative research projects it motivated. These include a robot vision system for object detection, recognition and pose estimation using probabilistic models, and methods allowing a robot to bootstrap higher-level knowledge and skills by exploratory interaction with objects. All our work heavily relies on machine learning.