Organic Machine Learning (OML)
Goal of the project »Organic Machine Learning« (OML) is to break the conventional, rigid approach of training an deploying of machine learning systems and to develop methods for machine learning which resemble organic learning, especially human learning, where systems learn throughout their entire lifetime - especially during application.
The way of learning shall become more organic. Instead of learning on very large, clean and well-structured training data, which have been prepared in a time-consuming manner, the systems developed in OML shall learn from heterogeneous data with little or no preparation, like they occur in real world scenarios, and shall require less training data, like humans do. Different sources, such as interaction with humans and own experience, shall be combined in a multimodal way and learning shall be focused on cases of uncertainty. To achieve this, systems must be able to detect in which cases they are unsure and where further learning is necessary. Furthermore, learning systems should not be a »black box« whose functioning outsiders cannot see into. Instead, they should be able to explain their decision-making and behavior. With the ability to justify their decisions, those system will become more accepted by humans and their application in real world environments possible.
Finally, the developed system shall be integrated in a robotic system. In an interactive robot programming scenario, the robot shall learn new skills from scratch by learning from physical, visual and verbal interaction with a human as well as own experience - just how an apprentice is taught by his master.
Robdekon: Robot systems for the decontamination in misanthropic environments
ROBDEKON stands for »Robot systems for decontamination in misanthropic environments« and is a competence center that is dedicated to research of autonomous and semi-autonomous robotic systems. In the future, these robots are supposed to carry out decontamination work independently of human beings so that people can stay away from danger zones.
The aim of ROBDEKON is the research and development of novel robotic systems for decontamination tasks. Research topics include mobile robots for rough terrain, autonomous machines for construction, robotic manipulators as well as decontamination concepts, planning algorithms, multisensory 3D-mappings of the environment and teleoperational techniques using virtual reality. Artificial intelligence methods enable robots to perform assigned tasks either fully autonomously or semi-autonomously.
The research focus of H2T within the competence center ROBDEKON lies in the development of methods for single and multi arm grasping, mobile manipulation and planning of manipulation actions for handling contaminated objects, whereby the visual perception and autonomous execution of decontamination tasks represent the central questions of research.
ARCHES: Autonomous Robotic Networks to Help Modern Societies
The aim of the Helmholtz Future Project ARCHES is the development of heterogeneous, autonomous and interconnected robotic systems in a consortium of the Helmholtz Centers DLR, AWI, GEOMAR and KIT from 2018 to 2020. The aspired fields of application are as heterogeneous as the robotic systems and span from environmental monitoring of the oceans over technical crisis intervention to the exploration of the solar system.
KIT leads the efforts to develop methods for grasping and manipulation in the domains of planetary and ocean exploration. In particular, KIT will develop a manipulation toolkit providing methods for grasp, motion and manipulation planning as well as approaches for coordinated execution of manipulation tasks in robot teams. The focus lays on providing grasping and manipulation skills for known, familiar and unknown rigid objects. The skills will be tailored to several levels of autonomy ranging from teleoperation to semi-autonomous and autonomous execution.
TERRINet: The European Robotics Research Infrastructure Network
EU H2020 Projekt
TERRINet represents the effort made by Europe to maintain its leadership in Robotics, by focusing on research. The project is committed to grow and educate a new generation of researchers in Robotics able to design, develop and manage future robots.
Therefore, partners in the TERRINet project offer top quality infrastructure, platforms, excellent research services and training to researchers, entrepreneurs, students and industry partners to fully exploit their potential and foster their impact on research and innovation.
Next to the shared access to top quality infrastructure and services, TERRINet aims at serving as a multi-disciplinary, trans-national environment to facilitate cross-fertilization of ideas and sharing of excellent scientific research. All partners appreciate the reinforced partnership with research and industry organisations. The integrated and harmonised access to infrastructures, platforms, service resources and ideas allows TERRINet to establish testing and evaluation facilities for innovation and will trigger more efficient technology transfer in Europe.
KIT provides access to its infrastructure and humanoid robot platforms to conduct research on humanoid robotics and wearable robotics including advanced control, vision- and haptic-based grasping, mobile manipulation, learning from human observation and experience, motion analysis and syntheses as well as context-sensitive control of prostheses and orthoses.
IMAGINE: Robots Understanding Their Actions by Imagining Their Effects
The objective of IMAGINE is to enable robots to understand how their environment behaves and how to interact with it, resulting in substantial advances in autonomy, flexibility and robustness of interaction in the presence of unforeseen changes in the environment. Specifically, in the context of IMAGINE understanding means the ability of the robot 1) to infer which actions possibly apply in a given state of the environment, and how to parametrize the actions to achieve a desired effect; and 2) to discern to what extent its actions succeed, and if the action effect differs from the expected outcome, to infer possible reasons and recover from failures. The developed methods should be evaluated in the context of robot-driven disassembly and recycling of electromechanical devices, such as hard discs. To this end, the project will augment the level of structural understanding of devices and functional understanding of actions available to robots, enabling them to infer how to disassemble entire categories of electromechanical devices and appliances, to monitor the success of their own disassembly actions, and to synthesize effective recovery strategies for problems they encounter during disassembly.
KIT is responsible for the development of a multi-functional reconfigurable gripper, designed for the disassembly of electromechanical devices and appliances. KIT also leads the work on learning disassembly tasks from human observation and the integration of different components for object and affordance perception, planning, simulation and action execution into a holistic system architecture, resulting in multiple functional demonstrators over the course of the project.
INOPRO: Intelligent orthotics and prosthetics
The INOPRO project aims at developing personalized intelligent prostheses and orthoses with an increased level of autonomy achieved by embodied machine intelligence, which facilitates symbiotic interaction with the human and more intuitive support, and better adaptation to the human’s needs, leading to substantial enhancement of the quality of life of prostheses and orthoses users.
KIT is leading the development of a new personalized intelligent hand prostheses, which are adaptable to the user’s needs in size and functionality and are equipped with machine intelligence to facilitate intuitive human-machine interaction and reduce the cognitive load of the user. To achieve this goal, we rely on a tight integration of intelligent mechanics, novel 3D printing technologies, embedded systems, and robust control with few, simple control signals. To enhance the autonomous abilities of hand prostheses in the sense of context-sensitive grasping of daily objects, we are working on the development of low-dimensional grasping representations and data-driven learning of a library of grasping skills from human observation as well as on resource-aware algorithms for image processing using artificial neural networks.
Furthermore, KIT contributes to the design and control of novel leg prostheses and orthoses by developing novel joint mechanisms and mechanisms for the realisation of synergetic coupling in the lower limb as well as methods for action classification, data-driven user state estimation based on multi-sensory information, and context sensitive-control.
SECONDHANDS: A robot assistant for industrial maintenance
The goal of SecondHands is to design a robot that can offer help to a maintenance technician in a pro-active manner. The robot will be a second pair of hands that can assist the technician when he/she is in need of help. Thus, the robot should recognize human activity, anticipate humans needs and pro-actively offer assistance when appropriate, in real-time, and in a dynamically changing real world. The robot assistant will increase the efficiency and productivity of maintenance technicians in order to ensure a smooth running of production machinery thereby maximizing output and return on investment.
KIT leads the tasks of the design of the new robot, grasping and mobile manipulation as well as natural language and multimodal interfaces. Based on the ARMAR humanoid robot technologies, a new humanoid robot for maintenance tasks will be developed and validated in several maintenance tasks in a warehouse environment. A special focus is the development of methods for task-specific grasping of familiar objects, object handover, vison- and haptic-based reactive grasping strategies as well as mobile manipulation tasks while taking into account uncertainties in perception and action. To facilitate natural human-robot interaction, KIT will implement a speech recognition, and dialogue management system to enable maintenance technicians to conduct spoken dialog with the robot.
In the Transregional Collaborative Research Center Invasive Computing (abbr. InvasIC), we are investigating a completely novel paradigm for designing and programming future parallel computing systems called invasive computing.
Subproject D1 will explore the specific benefits and restrictions of invasive architectures in challenging real-time embedded systems and in particular in humanoid robotics. We will investigate the implementation of a cognitive robot control architecture with its different processing hierarchies, both on invasive TCPA and RISC-based MPSoC. The goal is to explore techniques of self-organisation to efficiently allocate available resources for the timely varying requirements of robotic applications. We expect that less computing resources are needed to fulfill the application requirements compared to traditional resource assignment at compile-time.
TIMESTORM: Mind and Time - Investigation of the temporal attributes of human-machine synergetic interaction
TimeStorm aims at equipping artificial systems with humanlike cognitive skills that benefit from the flow of time by shifting the focus of human-machine confluence to the temporal, short- and long-term aspects of symbiotic interaction. The integrative pursuit of research and technological developments in time perception will contribute significantly to ongoing efforts in deciphering the relevant brain circuitry and will also give rise to innovative implementations of artifacts with profoundly enhanced cognitive capacities. TimeStorm promotes time perception as a fundamental capacity of autonomous living biological and computational systems that plays a key role in the development of intelligence. In particular, time is important for encoding, revisiting and exploiting experiences (knowing), for making plans to accomplish timely goals at certain moments (doing), for maintaining the identity of self over time despite changing contexts (being).
The main role of KIT in TimeStorm is to investigate the temporal information in the perception and execution of manipulation actions and to integrate time processing mechanisms in humanoid robots. In particular, we investigate how semantic representation (top-down) and hierarchical segmentation (bottom-up) of human demonstrations based on spatio-temporal object interactions can be combined to facilitate generalization of action durations. This would allow a robot to scale perceived and learnt temporal information of an action in order to perform the same and other actions with various temporal lengths.
Robots Exploring Tools as Extensions to their Bodies Autonomously (REBA+)
While neuroscientific research is unravelling a remarkable complexity and flexibility of body representations of biological agents during actions and the use of tools (Cardinali et al., 2012; Umilta et al., 2008; Maravita and Iriki, 2004), available approaches for representing body schemas for robots (Hoffmann et al., 2010a) still largely lack equally sophisticated, adaptive and dynamically extensible representations. Associated and largely open challenges are rich representations that marry body morphology, control, and the exploitation of redundant degrees of freedom in representations that offer strong priors for rapid learning that in turn support a flexible adaptation and extension of these representations to realize capabilities such as tool use or graceful degradation in case of malfunction of parts of the body tree. This motivates the present project: to develop, implement and evaluate for a robot rich extensions of its body schema, along with learning algorithms that use these representations as strong priors in order to enable rapid and autonomous usage of tools and a flexible coping with novel mechanical linkages between the body, the grasped tool and target objects. As a major scientific contribution to Autonomous Learning we aim to address these key aspects of interaction learning:
- enhancing the scope from the body morphology to a representation of body-tool-environment linkages
- enhancing the scope from a representation of morphology to a representation that includes control
- enhancing the scope from minimal DOF systems to systems that offer and exploit redundant degrees of freedom
We will develop these representations in the context of a systematically chosen ”matrix” of real-world interaction situations, arranged to pose learning challenges in increasing order of complexity along the above three dimensions. Thereby, we will build on the previous project, where we have developed a basic framework for adaptive body schemes emphasising the kinematics level. The project will thus directly contribute to enhance the autonomy of robots for adjusting their physical interaction with the world to the variety of situations that is characteristic of many natural environments. It will also advance the state of the art of representations that can support such capabilities, including representations that can autonomously extend themselves as a result of autonomous exploration.
The I-SUPPORT project envisions the development and integration of an innovative, modular, ICT-supported service robotics system that supports and enhances frail older adults’ motion and force abilities and assists them in successfully, safely and independently completing the entire sequence of bathing tasks, such as properly washing their back, their upper parts, their lower limbs, their buttocks and groin, and to effectively use the towel for drying purposes. Advanced modules of cognition, sensing, context awareness and actuation will be developed and seamlessly integrated into the service robotics system to enable the robotic bathing system.
KIT leads the tasks concerning the learning motion primitives from human observation and kinesthetic teaching for a soft robot arm which should provide help in washing and drying tasks. The learned motion primitives should be represented in a way, which allow the adaptation to different context (soaping, washing, drying), body parts (back, upper lower limbs, neck) and users. To achieve this, adaptive representations will be developed of the learned motion primitives which are able to account for the different requirements such as encoding different motion styles (circular and linear), adaptation to different softness of different body parts, etc. To enable the handling of different washing and trying tools with varying softness by the robot arm, we will investigate how motion primitives can be enriched with models which encode correlations between objects properties and action parameters. Furthermore, KIT is addressing the task of personalization and adaptation of the robotic bathing system to the user by taking into the users‘ preference and previous sensorimotor experience. Based on a reference model of the human body, the Master Motor Map, which defines the kinematics and dynamics of the human body with regard to global body parameters such as height, weight, we will derive individual models of the different users. These models will be used to generate initial washing and/or drying behavior which will be refined based on sensorimotor experience obtained from the robot arm.
WALK-MAN: Whole-body Adaptive Locomotion and Manipulation
WALK-MAN targets at enhancing the capabilities of existing humanoid robots, permitting them to assist or replace humans in emergency situations including rescue operations in damaged or dangerous sites like destroyed buildings or power plants. The WALK-MAN robot will demonstrate human-level capabilities in locomotion, balance and manipulation. The scenario challenges the robot in several ways: Walking on unstructured terrain, in cluttered environments, among a crowd of people as well as crawling over a debris pile. The project's results will be evaluated using realistic scenarios, also consulting civil defence bodies.
KIT leads the tasks concerning multimodal perception for loco-manipulation and the representation of whole-body affordances. The partly unknown environments, in which the robot has to operate, motivate an exploration-based approach to perception. This approach will integrate whole-body actions and multimodal perceptual modalities such as visual, haptic, inertial and proprioceptive sensory information. For the representation of whole-body affordances, i.e. co-joint perception-action representations of whole-body actions associated with objects and/or environmental elements, we will rely on our previous work on Object-Action Complexes (OAC), a grounded representation of sensorimotor experience, which binds objects, actions, and attributes in a causal model and links sensorimotor information to symbolic information. We will investigate the transferability of grasping OACs to balancing OACs, inspired by the analogy between a stable whole-body configuration and a stable grasp of an object.
KoroiBot: Improving humanoid walking capabilities by human-inspired mathematical models, optimization and learning
KoroiBot a three years project funded by the European Commission under FP7-ICT-2013-10. The goal of the project is to investigate the way humans walk, e.g., on stairs and slopes, on soft and slippery ground, over beams and seesaws and create mathematical models and learning methods for humanoid walking. The project will study human walking, develop techniques for increased humanoid walking performance and evaluate them on existing state of the art humanoid robots.
KIT leads the tasks concerning human walking experiments, the establishment of large scale human walking database and the development of human and humanoid models as basis for the development of general motion and control laws transfer rules between different embodiments and for the generation of different walking types. The developed models and transfer rules, we will study how to implement balancing and push recovery strategies to deal with different types of perturbation in free and constrained situations. Furthermore, we will investigate the role of prediction in walking as well as the role different sensory feedback such as vision, vestibular and foot haptics in balancing. Therefore, we will implement a control and action selection schema emphasizing predictive control mechanisms, which rely on the estimation of expected perturbation based on multimodal sensory feedback and past sensorimotor experience. The control schema will be validated in the context of prediction and selection of push recovery actions.
XPERIENCE: Robots Bootstrapped through Learning from Experience
Current research in enactive, embodied cognition is built on two central ideas: 1) Physical interaction with and exploration of the world allows an agent to acquire and extend intrinsically grounded, cognitive representations and, 2) representations built from such interactions are much better adapted to guiding behaviour than human crafted rules or control logic. Exploration and discriminative learning, however are relatively slow processes. Humans, on the other hand, are able to rapidly create new concepts and react to unanticipated situations using their experience. “Imagining” and “internal simulation”, hence generative mechanisms which rely on prior knowledge are employed to predict the immediate future and are key in increasing bandwidth and speed of cognitive development. Current artificial cognitive systems are limited in this respect as they do not yet make efficient use of such generative mechanisms for the extension of their cognitive properties.
The Xperience project will address this problem by structural bootstrapping, an idea taken from child language acquisition research. Structural bootstrapping is a method of building generative models, leveraging existing experience to predict unexplored action effects and to focus the hypothesis space for learning novel concepts. This developmental approach enables rapid generalization and acquisition of new knowledge and skills from little additional training data. Moreover, thanks to shared concepts, structural bootstrapping enables enactive agents to communicate effectively with each other and with humans. Structural bootstrapping can be employed at all levels of cognitive development (e.g. sensorimotor, planning, communication).
- Xperience will demonstrate that state‐of‐the‐art enactive systems can be significantly extended by using structural bootstrapping to generate new knowledge. This process is founded on explorative knowledge acquisition, and subsequently validated through experience‐based generalization.
- Xperience will implement, adapt, and extend a complete robot system for automating introspective, predictive, and interactive understanding of actions and dynamic situations. Xperience will evaluate, and benchmark this approach on existing state‐of‐the‐art humanoid robots, integrating the different components into a complete system that can interact with humans.
By equipping embodied artificial agents with the means to exploit prior experience via generative inner models, the methods to be developed here are expected to impact a wide range of autonomous robotics applications that benefit from efficient learning through exploration, predictive reasoning and external guidance.
HEiKA-EXO: Optimization-based development and control of an exoskeleton for medical applications
HEiKA-EXO is a project, which is financed by the HEiKA joint research initiative of the University of Heidelberg and the Karlsruhe Institute of Technology. The project is a prototype study for the design and control of a lower leg exoskeleton for medical applications based on expertise available at KIT in humanoid robotics and the University of Heidelberg in mathematical modeling. The first exoskeleton prototype feature two active degrees of freedom in the knee and ankle joints.
GRASP is an Integrated Project funded by the European Commission through its Cognition Unit under the Information Society Technologies of the seventh Framework Programme (FP7). The project was launched on 1st of March 2008 and will run for a total of 48 months.
The aim of GRASP is the design of a cognitive system capable of performing grasping and manipulation tasks in open-ended environments, dealing with novelty, uncertainty and unforeseen situations. To meet the aim of the project, studying the problem of object manipulation and grasping will provide a theoretical and measurable basis for system design that is valid in both human and artificial systems. This is of utmost importance for the design of artificial cognitive systems that are to be deployed in real environments and interact with humans and other agents. Such systems need the ability to exploit the innate knowledge and self-understanding to gradually develop cognitive capabilities. To demonstrate the feasibility of our approach, we will instantiate, implement and evaluate our theories and hypotheses on robot systems with different embodiments and complexity.
GRASP goes beyond the classical perceive-act or act-perceive approach and implements a predict-act-perceive paradigm that originates from findings of human brain research and results of mental training in humans where the self-knowledge is retrieved through different emulation principles. The knowledge of grasping in humans can be used to provide the initial model of the grasping process that then has to be grounded through introspection to the specific embodiment. To achieve open-ended cognitive behaviour, we use surprise to steer the generation of grasping knowledge and modelling
PACO-PLUS: Perception, Action and Cognition through Learning of Object-Action Complexes
PACO-PLUS is an integrated Project funded by the European Commission through its Cognition Unit under the Information Society Technologies of the sixth Framework Programme (FP6). The project was launched on 1st of February 2006 and will run for a total of 48 months.
PACO-PLUS brings together an interdisciplinary research team to design and build cognitive robots capable of developing perceptual, behavioural and cognitive categories that can be used, communicated and shared with other humans and artificial agents. To demonstrate our approach we are building robot systems that will display increasingly advanced cognitive capabilities over the course of the programme. They will learn to operate in the real world and to interact and communicate with humans. To do this they must model and reflectively reason about their perceptions and actions in order to learn, act and react appropriately. We hypothesize that such understanding can only be attained by embodied agents and requires the simultaneous consideration of perception and action.
Our approach rests on three foundational assumptions:
- Objects and Actions are inseparably intertwined in cognitive processing; that is “Object-Action Complexes” (OACs) are the building blocks of cognition.
- Cognition is based on reflective learning, contextualizing and then reinterpreting OACs to learn more abstract OACs, through a grounded sensing and action cycle.
- The core measure of effectiveness for all learned cognitive structures is: Do they increase situation reproducibility and/or reduce situational uncertainty in ways that allow the agent to achieve its goals?
Collaborative Research Center (SFB) 588: Humanoid Robots - Learning and Cooperating Multimodal Robots
The Collaborative Research Center 588 "Humanoid Robots - Learning and Cooperating Multimodal Robots" was established on the 1st of July 2001 by the Deutsche Forschungs-gemeinschaft (DFG) and will run until June 30th, 2012.
The goal of this project is to generate concepts, methods and concrete mechatronical components for a humanoid robot, which will be able to share his activity space with a human partner. With the aid of this partially anthromorphic robot system, it will be possible to step out of the "robot cage" to realise a direct contact to humans.
The term multimodality includes the communication modalities intuitive for the user such as speech, gesture and haptics (physical contact between the human and the robot), which will be used to command or instruct the robot system directly.
Concerning the cooperation between the user and the robot - for example in the joint manipulation of objects - it is important for the robot to recognise the human's intention, to remember the acts that have already been carried out together and to apply this knowledge correctly in the individual case. Great effort is spent on safety, as this is a very important aspect of the man-machine-cooperation.
An outstanding property of the system is its ability to learn. The reason for this is the possibility to lead the system to new, formerly unknown problems, for example to new terms and new objects. Even new motions will be learned with the aid of the human and they can be corrected in an interactive way by the user.
The Collaborative Research Center 588 is assigned to the Department of Informatics. More than 40 scientists and 13 institutes are involved in this project. They belong to the department of Informatics, the Faculty of Electrical and Information Engineering, the Faculty of Mechanical Engineering and Faculty of Humanities and Social Sciences as well as Fraunhofer Institute of Optronics, System Technologies and Image Exploitation (IOSB) and the Forschungszentrum Informatik Karlsruhe (FZI).
EU FET Flagship Initiative "Robot Companions for Citizens"
The coordination action CA-RoboCom will design and describe the FET Flagship initiative “Robot Companions for Citizens” (RCC) including its: Scientific and Technological framework, governance, financial and legal structure, funding scheme, competitiveness strategy and risk analysis.
The FET Flagship initiative RCC will realize a unique and unforeseen multidisciplinary science and engineering program supporting a radically new approach towards machines and how we deploy them in our society.
Robot Companions for Citizens is an ecology of sentient machines that will help and assist humans in the broadest possible sense to support and sustain our welfare. RCC will have soft bodies based on the novel integration of solid articulated structures with flexible properties and display soft behavior based on new levels of perceptual, cognitive and emotive capabilities. RCC will be cognizant and aware of their physical and social world and respond accordingly. RCC will attain these properties because of their grounding in the most advanced sentient machines we know: animals.
Robot Companions for Citizens will validate our understanding of the general design principles underlying biological bodies and brains, establishing a positive feedback between science and engineering.
Driven by the vision and ambition of RCC, CA-RoboCom will, by means of an appropriate outreach strategy, involve all pertinent stakeholders: science and technology, society, finance, politics and industry. Other than the commitment of its Consortium, CA-RoboCom will involve a wide range of external experts in its working groups, its Advisory Board, and in its European and International Cooperation board. The CA-RoboCom consortium believes that given the potential transformative and disruptive effects of RCC in our society their development and deployment has to be based on a the broadest possible support platform.