Periodic Reporting for period 1 - COROBOREES (COllective ROBOtics through Real-time Entrainment of Evolved dynamical Systems)
Periodo di rendicontazione: 2021-09-01 al 2023-08-31
As intelligent animals, we make extensive use of social information at all cognitive levels, including the most basic. One striking example of this is that we unconsciously synchronize our walking, not only with musical rhythms, but also with each other. Such unconscious imitation means that at a basic level our sensorimotor systems are open to feedback not only from the physical environment, but also from the social environment. This has profoundly shaped how humans acquire behaviours, and hence how we learn and communicate in general.
The robotics field has made great strides in sensorimotor control and learning using environmental feedback. This has included biologically-inspired use of central pattern generators (CPGs) modeled after spinal circuits, with many impressive examples of robust and flexible locomotion that emerge from interactions with the environment. Such “situated” robotics is often seen as a necessary condition for truly intelligent behaviour. However, low-level social feedback and learning have not received nearly the same level of focus, and have not been investigated in artificial CPGs.
The novel research goal of the COROBOREES project was to develop robotic agents with socially adaptive CPGs that can overcome morphological differences and synchronize complex movement patterns. Achieving this goal can be seen as a significant step towards socially situated - and socially intelligent - robots.
The project had two main objectives, each with its own work package devoted to it:
1. To develop and compare mechanisms for entrainment of CPG-based controllers to rhythmic acoustic signals.
2. To understand how sensorimotor control mechanisms affect the emergent outcome of multiple agents interacting.
The robotics field has made great strides in sensorimotor control and learning using environmental feedback. This has included biologically-inspired use of central pattern generators (CPGs) modeled after spinal circuits, with many impressive examples of robust and flexible locomotion that emerge from interactions with the environment. Such “situated” robotics is often seen as a necessary condition for truly intelligent behaviour. However, low-level social feedback and learning have not received nearly the same level of focus, and have not been investigated in artificial CPGs.
The novel research goal of the COROBOREES project was to develop robotic agents with socially adaptive CPGs that can overcome morphological differences and synchronize complex movement patterns. Achieving this goal can be seen as a significant step towards socially situated - and socially intelligent - robots.
The project had two main objectives, each with its own work package devoted to it:
1. To develop and compare mechanisms for entrainment of CPG-based controllers to rhythmic acoustic signals.
2. To understand how sensorimotor control mechanisms affect the emergent outcome of multiple agents interacting.
The project consisted of five work packages, two of which were devoted to pursuing the project milestones and deliverables (WP1&WP2).
In work package WP1, theoretical approaches to entrainment, oscillator models and CPG architectures were compared. A new oscillator model was developed, based on the commonly used Matsuoka oscillator, but with added nonlinearity to promote flexibility. Multi-objective evolution was used to generate a diverse population of controllers that were interfaced with virtual quadrupeds. The phenomenon of spontaneous entrainment was studied by introducing rhythmic input to the network. Emergent properties of the CPG networks were analysed, with the result that sensitivity of gait frequency to input was predictive of entrainment ability. Complex inputs were also analysed using real musical excerpts, with the finding that pulse clarity partially predicted entrainment performance.
In work package WP2, a virtual hexapod was created from an open source design, with the controller using the same CPG modules as the quadruped in WP1. The controller evolution was repeated for these virtual hexapods. A morphology-agnostic algorithm was designed to learn new movement patterns during robot-to-robot interaction, and tested on a heterogeneous population of quadrupeds and hexapods. It was found that entrainment ability was mainly learner dependent, and that the learning scheme improved stability of movement compared to pure entrainment. A prototype of the hexapod design was also built and interfaced with an evolved controller, and was used to demonstrate entrainment to simple rhythms.
These results were detailed in two open-access peer-reviewed articles, two conference papers and two Erasmus project reports. The results were disseminated at three workshops in the host country and four international conferences and workshops.
In work package WP1, theoretical approaches to entrainment, oscillator models and CPG architectures were compared. A new oscillator model was developed, based on the commonly used Matsuoka oscillator, but with added nonlinearity to promote flexibility. Multi-objective evolution was used to generate a diverse population of controllers that were interfaced with virtual quadrupeds. The phenomenon of spontaneous entrainment was studied by introducing rhythmic input to the network. Emergent properties of the CPG networks were analysed, with the result that sensitivity of gait frequency to input was predictive of entrainment ability. Complex inputs were also analysed using real musical excerpts, with the finding that pulse clarity partially predicted entrainment performance.
In work package WP2, a virtual hexapod was created from an open source design, with the controller using the same CPG modules as the quadruped in WP1. The controller evolution was repeated for these virtual hexapods. A morphology-agnostic algorithm was designed to learn new movement patterns during robot-to-robot interaction, and tested on a heterogeneous population of quadrupeds and hexapods. It was found that entrainment ability was mainly learner dependent, and that the learning scheme improved stability of movement compared to pure entrainment. A prototype of the hexapod design was also built and interfaced with an evolved controller, and was used to demonstrate entrainment to simple rhythms.
These results were detailed in two open-access peer-reviewed articles, two conference papers and two Erasmus project reports. The results were disseminated at three workshops in the host country and four international conferences and workshops.
The COROBOREES project investigated the potential for evolving dynamical systems in social and evolutionary robotics. For the first time, it demonstrated spontaneous entrainment of virtual and real robots to rhythmic inputs, and the transfer of movement patterns between virtual robots using this entrainment. In the process of achieving these goals, the action demonstrated several novel methods and results, which provide a fundamental scientific basis to future developments in social, collective, developmental and evolutionary robotics.
Mechanistic agents with configurable parameters, as developed in the project may also provide important insights into the neuroscience of the sensorimotor system, and are expected to also be useful in controlled interaction studies. The project outputs, such as the robot prototype, will be used by the host institution to further investigate rhythm and motion in humans.
Mechanistic agents with configurable parameters, as developed in the project may also provide important insights into the neuroscience of the sensorimotor system, and are expected to also be useful in controlled interaction studies. The project outputs, such as the robot prototype, will be used by the host institution to further investigate rhythm and motion in humans.