There was a nice workshop on “Advances in Biologically Inspired Brain-Like Cognition and Control for Learning Robots” at IROS. I presented the work on recruiting the internal model for planning ahead and cognition as an extended abstract and poster there. The abstract can be found here.
We applied the MMC approach to a population-code based representation – this allows to decompose complex robotic structures into simple, local geometric relations that can be easily solved in parallel. The population-code representation shall be utilized in the future to allow for integration of additional sensory signals. The paper has been presented at the IJCNN in Ireland, 2015:
Population based encodings allow to represent probabilistic and fuzzy state estimates. Such a representation will be introduced and applied for the case of a redundant manipulator. Following the Mean of Multiple Computations principle, a neural network model (PbMMC) is presented in which the overall complexity is divided into multiple local relationships. This allows to solve inverse, forward and mixed kinematic problems. The local transformations in between the kinematic variables can be sufficiently well learned by small single MLP layers. … the model as such is quite flexible as it can keep track of multiple possible solutions at the same time.
Citation: Baum, M., Meier, M., and Schilling, M. (2015), “Population based Mean of Multiple Computations Networks: A Building Block for Kinematic Models”. International Joint Conference on Neural Networks 2015, Killarney (Ireland).
I presented our paper at ICRA 2015 in Seattle on how sensory influences and central oscillation are both important in the generation of emergent walking behavior:
Locomotion control deals with the generation of (quasi-)rhythmic behaviors. There are two general approaches for the generation of such behavioral patterns. On the one hand, a central approach in which a pattern is generated open-loop, driving the motor output without relying on sensory feedback. On the other hand, a sensory driven approach relies on sensory feedback that dominates motor control. Both show different advantages and seem to serve different functions depending on the context. … we want to support a middle ground which tries to bring those two approaches together to provide robots with the adaptive and versatile motor control of animals.
In an extension to the internal model based approach, our paper got accepted at ICRA 2015 to present how the stance movement is controlled and (in addition) how an evading reflex can be incorporated into the overall control architecture:
a concept that allows the inherently compliant hexapod robot HECTOR to walk on uneven terrain and to overcome moderate obstacles by means of a decentralized walking controller. An important prerequisite is the availability of inherently compliant joint drives with an increased power/weight-ratio.
Citation: Paskarbeit, J., Schilling, M., Schmitz, J., and Schneider, A. (2015), “Obstacle crossing of a real, compliant robot based on local evasion movements and averaging of stance heights using singular value decomposition,” in IEEE International Conference on Robotics and Automation (ICRA), pp.3140-3145.
Our minimal cognitive systems approach provides a good starting point to analyze emergent high level properties. We just published in the MIND collection an article:
we propose a bottom-up approach to higher-level mental states, such as emotions, attention, intention, volition, or consciousness. The idea behind this bottom-up approach is that higher-level properties may arise as emergent properties, i.e., occur without requiring explicit implementation of the phenomenon under examination. Using a neural architecture that shows the abilities of autonomous agents, we want to come up with quantitative hypotheses concerning cognitive mechanisms, i.e., to come up with testable predictions concerning the underlying structure and functioning of an autonomous system that can be tested in a robot-control system. … our network does not only show emergent properties on the reactive level; it also shows that mental states, such as emotions, attention, intention, volition, or consciousness can be observed, too. [For example] … the property of global availability, which means that elements of the procedural memory can be addressed even if they do not belong to the current context.
Citation: Cruse, H. and Schilling, M. (2015), “Mental states as emergent properties. From walking to consciousness”. In: Open Mind. Metzinger, T., Windt, J. (eds.); Frankfurt/M.: MIND Group Frankfurt/M.
Finally, the hexapod robot has done his first steps. It is controlled by the decentralized Walknet architecture (see Schilling et al., 2013a and 2013b below) and produces adaptive walking behavior.
A video can be found here.
For more details on the robotic structure see Schneider, A., Paskarbeit, J., Schilling, M. and Schmitz, J. (2014), “HECTOR, A Bio-Inspired and Compliant Hexapod Robot”. In: A. Duff, T. Prescott, P. Verschure, N. Lepora (eds.): Living Machines 2014, LNAI 8608, pp. 427–429.
Insect-inspired control approaches often provide adaptive, robust and cheap solutions to common robotic problems. … We present practical observations focusing on an insectoid robot equipped with active antennae, and then expose verbal communication concepts for embodied cognition. Finally, we discuss how to combine insect-like processes intelligible to humans (e.g. visuotactile attention) with capabilities exceeding those of the insects (e.g. intuitive gestures, verbal communication).
Citation: Hoinville, T., Krause, A., Schilling, M., and Cruse, H. (2014). “Bridging an Interspecies Gap? Toward Human-Insectoid Robot Interaction.” In Proceedings of 9th ACM/IEEE International Conference on Human-Robot Interaction (HRI 2014), Bielefeld, Germany.
The Motivation Network deals with the selection of action. It is realized as a recurrent neural network which forms local winner-take-all structures. The local competition allows for concurrent activation and decentralized control structures as required, e.g., for six legged walking. In the example case of six legged walking, it produces emergent stable and adaptive gaits.
Citation: Schilling, M., Paskarbeit, J., Hüffmeier, A., Schneider, A., Schmitz, J., and Cruse, H. (2013), “A hexapod walker using a heterarchical architecture for action selection”. Frontiers in Computational Neuroscience 7:126.
Our review on the reactive control system Walknet and a review on the biological findings on walking in stick insects appeared in Biological Cybernetics (as Open Access):
This review summarizes the most important biological find- ings described by Walknet and how they can be simulated. Walknet shows how a number of properties observed in insects may emerge from a decentralized architecture.
Citation: Schilling, M., Hoinville, T., Schmitz, J. and Cruse, H. (2013), “Walknet, a bio-inspired controller for hexapod walking”. Biological Cybernetics, 107(4), pages 397-419.
Towards the connection to language, the concept has been published at the AAAI Conference in Stanford together with Srini Narayanan:
… we present a system of connected knowledge representations that is used to control a robot through instructions. As actions are a key component of in- structions and the robot’s behavior the representation of ac- tion is central in our approach. First, the system consists of a conceptual schema representation which provides a param- eter interface for action. Second, we present an intermediate representation of the temporal structure of action and show how this generic action structure can be mapped to detailed action controllers as well as language.
Citation: Schilling, M. and Narayanan, S. (2013), “Communicating with Executable Action Representations”. In Proceedings of AAAI Spring Symposium Series 2013, Stanford.