Contributions published at Artificial Intelligence Lab

In recent years, breakthroughs have been made in both biology and robotics by the close cooperation between biologists and roboticists. Researchers in the fields of octopus biological study and soft robotics can also benefit from working together and inspiring each other. However, this collaboration is not easy because of the different research motivations and terminologies used. This paper starts from challenges for controlling soft robots, through biological inspirations from the octopus, to their engineering interpretation for creating a behavior control architecture for soft robots. Hypotheses related to the octopus biology from the engineering perspective are also proposed. This work serves as a case study in biologist-roboticist collaboration. It wishes to promote mutual understanding and cooperation between these two disciplines.

The cognitivistic paradigm, which states that cognition is a result of computation with symbols that represent the world, has been challenged by many. The opponents have primarily criticized the detachment from direct interaction with the world and pointed to some fundamental problems (for instance the symbol grounding problem). Instead, they emphasized the constitutive role of embodied interaction with the environment. This has motivated the advancement of synthetic methodologies: the phenomenon of interest (cognition) can be studied by building and investigating whole brain-body-environment systems. Our work is centered around a compliant quadruped robot equipped with a multimodal sensory set. In a series of case studies, we investigate the structure of the sensorimotor space that the application of different actions in different environments by the robot brings about. Then, we study how the agent can autonomously abstract the regularities that are induced by the different conditions and use them to improve its behavior. The agent is engaged in path integration, terrain discrimination and gait adaptation, and moving target following tasks. The nature of the tasks forces the robot to leave the ``here-and-now'' time scale of simple reactive stimulus-response behaviors and to learn from its experience, thus creating a ``minimally cognitive'' setting. Solutions to these problems are developed by the agent in a bottom-up fashion. The complete scenarios are then used to illuminate the concepts that are believed to lie at the basis of cognition: sensorimotor contingencies, body schema, and forward internal models. Finally, we discuss how the presented solutions are relevant for applications in robotics, in particular in the area of autonomous model acquisition and adaptation, and, in mobile robots, in dead reckoning and traversability detection.

This paper focuses on learning and adaptation of sensorimotor contingencies. As a specific case, we investigate the application of prism glasses, which change visual-motor contingencies. After an initial disruption of sensorimotor coordination, humans quickly adapt. However, scope and generalization of that adaptation is highly dependent on the type of feedback and exhibits markedly different degrees of generalization. We apply a model with a specific interaction of forward and inverse models to a robotic setup and subject it to the identical experiments that have been used on previous human psychophysical studies. Our model demonstrates both locally specific adaptation and global generalization in accordance with the psychophysical experiments. These results emphasize the role of the motor system for sensory processes and open an avenue to improve on sensorimotor processing.

In conventional “sense-think-act” control architectures, perception is reduced to a passive collection of sensory information, followed by a mapping onto a prestructured internal world model. For biological agents, Sensorimotor Contingency Theory (SMCT) posits that perception is not an isolated processing step, but is constituted by knowing and exercising the law-like relations between actions and resulting changes in sensory stimulation. We present a computational model of SMCT for controlling the behavior of a quadruped robot running on different terrains. Our experimental study demonstrates that: (i) Sensory-Motor Contingencies (SMC) provide better discrimination capabilities of environmental properties than conventional recognition from the sensory signals alone; (ii) discrimination is further improved by considering the action context on a longer time scale; (iii) the robot can utilize this knowledge to adapt its behavior for maximizing its stability.

Background Increasing the communication speed of brain-computer interfaces (BCIs) is a major aim of current BCI-research. The idea to automatically detect error-related potentials (ErrPs) in order to veto erroneous decisions of a BCI has been existing for more than one decade, but this approach was so far little investigated in online mode. Methods In our study with eleven participants, an ErrP detection mechanism was implemented in an electroencephalography (EEG) based gaze-independent visual speller. Results Single-trial ErrPs were detected with a mean accuracy of 89.1% (AUC 0.90). The spelling speed was increased on average by 49.0% using ErrP detection. The improvement in spelling speed due to error detection was largest for participants with low spelling accuracy. Conclusion The performance of BCIs can be increased by using an automatic error detection mechanism. The benefit for patients with motor disorders is potentially high since they often have rather low spelling accuracies compared to healthy people.

In this paper, we present the concept of adapting to changes in ground conditions like stiffness, damping and friction, using a novel two degree of freedom reconfigurable leg length hopping robot with a fixed passive compliance. In such a robot, the change in the dynamics of the single legged hopper can be induced by the change in coupled stiffness and damping of the system, i.e., stiffness and damping of the ground coupled with the stiffness and damping of the robotic leg. It is experimentally shown by in-place hopping of a robotic leg on various grounds (stiff, less stiff and soft) that the leg can effectively adapt to changes in coupled stiffness and damping by the rate and the amplitude at which the leg length changes. This is true, while the leg hops in-place as the role of ground friction is negligible. However, in forward motion where the ground friction dominates, a change in initial effective leg length, i.e., shortening or lengthening can provide an additional support to the hip motor in overcoming even large variations in ground friction. This is demonstrated through a planar locomotion experiment on different ground surfaces. The overall results provide strong support for this concept.

his dissertation discusses the importance of swarming concepts such as self-organization in human crowds in relation to the use of collective intelligence obtained through (voluntary) participation via the Internet. It offers practical advice on how to manage such a (potentially large) group in order to maximize the resulting collective intelligence for application in (both) scientific inquiry and in business environments, including the development of innovations.Through recent socio-technical developments (i.e., Web 2.0), humans are augmenting the immense computing power of computers to solve difficult problems. They are doing this by contributing human heuristics and collective intelligence. Potentially large groups not necessarily comprised of experts are working on and communicating about problems that could not be solved before by individuals, smaller groups of experts, or machines. This “collective intelligence” is complex as it includes a large number of nonlinear dynamics resulting from interaction of (irrational) individuals with each other. This dissertation addresses the phenomena of crowdsourcing/open innovation currently involving hundreds of thousands of Internet users including public and private organizations. It comprises of an introductory part, including an extended literature review crossing discipline boundaries, and nine separate research papers in the appendix. The dissertation contributes to the theory of collective intelligence by addressing a research terrain previously unexplored that is concerned with the characteristics of distributed problem solving in large groups in both academic and business settings. It also provides an interdisciplinary contribution as it crosses academic discipline boundaries by applying knowledge and techniques from artificial intelligence research to investigating phenomena that, up until now, were mainly described by sociologists and business scholars. Selected results of this dissertation include: > Human groups collaborating through the Internet show many similarities to natural (and artificial) swarms > A few management actions that build on human specifics (like maintaining the diversity of groups against the human tendency to build teams with similar backgrounds) can increase the output from the group beyond pure “swarm intelligence” > Agent design principles compiled by embodied artificial intelligence scholars can be useful for conventional organizations to make the interaction with and within such groups more flexible, scalable, and robust in particular process steps of collaborative problem solving > The process of “crowdsourcing” results in useful inputs for scientists in an academic setting (e.g., a research university) for almost all tasks during a scientific inquiry (e.g., identifying qualified partners or analyzing experimental data) The knowledge gathered in the process of developing this dissertation is synthesized in a multi-agent system that simulates a pivotal set of the interactions mentioned above. First tests in academia and start-up companies indicate that the simulator is a useful tool for both scholars and practitioners.Suggestions for future theory building and research are outlined at the end of the dissertation.

Motor control theories from human neuroscience postulate that the body seems to reduce dimensions in its task of controlling the redundant, nonlinear, high dimensional musculo skeletal system. This thesis proposes an architecture to learn such a reduced order model in an unsupervised fashion for the pendulum robot developed at the AI Lab. The goal of the first phase is to track pendulum using a vision algorithm during a type of system identification process known as "motor babbling". This is followed by two learning steps to learn correlations and scaling factors. The result is a working control model which allows the robot to perform targeting tasks.

This thesis develops a gaming engine platform, a video game based system to allow experimental comparisons of many different approaches and technologies for the control of prosthetic devices, permitting the investigation of alternative controller architectures not yet experimented with. The controller interface implemented in this project uses surface EMG signals as input; given several examples of EMG controllers in the literature, the path implemented in the thesis tries to be different, providing a simple but complete control over the game. Further, some experimental setups are proposed, in preparation for future research based on the platform.

null

null

null

null

We present a global education project in Artificial Intelligence (AI) called the “ShanghAI Lectures”: A lecture series held annually via videoconference among 15 to 20 universities around the globe. The lectures are complemented by a novel three-dimensional collaborative virtual environment for international student teamwork, and a web-based resource designed as a knowledge base and for community building. This paper summarizes the lessons learned from the first edition of the ShanghAI Lectures, which may guide future global teaching and learning projects of this kind.

null

Any person and almost any organization often face problems which cannot be solved by the given know-how at hand. The traditional way of outsourcing these problems to a certain service provider bears the risk of receiving a biased or just no solution, and comes with a dependence on that particular provider. Hence it is desirable to have a designated expert tackling the problem accurately and with ease. We review the theoretical background of know-how trading by means of an online community and present the implementation of such a platform based on the theoretical insights and show the impact of applying those. This realization implements a virtual company blurring the border between online and offline world by direct user participation in the management of the corresponding offline company in the physical world.

In previous work, we have argued that a sophisticated cognitive system with a complex body must possess configurable models of itself (or at least its body) and the world, along with the necessary infrastructure to use the modelled interactions between these two components to select relatively advantageous actions. These models may be used to generate representations of the future (imagination) and the past (episodic memory). In this paper we will explore some problems surrounding the representation of the present arising from the use of such models in the artificial cognitive system under development within the ECCEROBOT project. There are two aspects to consider: the representation of the state of the robot’s body within the self model, and the representation of the state of the external world within the world model. In both natural and robotic systems, the processing of the sensory data carrying state information takes a considerable time, and so any estimates of the present states of both the agent and the world would have to be obtained by using predictive models. However, it appears that there is no need for any such representations to be generated in the course of selecting a course of action using self and world models, since representations are only of the future or the past. This may call into question the utility and timing of the apparent perception of the present in humans.

The biological hypothesis of spinal engine states that the locomotion is mainly achieved by the spine, while legs only serve as assistance. Inspired by this spinal engine hypothesis, a compliant, multi-DOF, biologically inspired spine has been developed and embedded into a quadruped robot without actuation on legs. The experimental results support this spinal engine hypothesis and reveal that this kind of robot can achieve rapid, stable, and even dynamical locomotion by appropriately tuning the spine’s morphological parameters, e.g., rearranging the silicone blocks.

In dynamic locomotion, robots’ morphology and the ability to adapt it online play an important role in energy efficiency and coping with the highly unpredictable perturbations from the environment. In this paper, we present the design and implementation of a quadruped robot whose morphology is particularly targeted towards energy-efficient dynamic locomotion. We propose a combination of mechanisms which allow for energy-efficient actuation, ground clearance, and gait versatility through adaptation of morphology (morphosis). We report on a series of experiments to validate the robot’s performance in different locomotion conditions.