From Models of Cognition to Robot Control and Back Using SNNs
177
this decision; for example whether the prime purpose of the agent is to be an
explanatory model of biological cognition or whether it takes biological inspiration merely to be successful at some task [25], or to what degree physiological
details are essential to cognition, as opposed to mere constraints by biology [24].
In this context, spiking neural networks (SNNs) have recently attracted interest in biologically inspired robotics. In part, this is driven by the recent advent
of neuromorphic hardware platforms that promise to be able to run relatively
large networks in real time (a fundamental requirement for any model operating on a real robotic platform) and, for example, the potential energy efficiency
such models may provide [6]. From a more cognitive perspective, however, it
is possible that SNNs are a suitable level of abstraction to adequately model
natural cognition. This is, for example, the assumption in the so-called neuroengineering framework [13] underlying the semantic pointer architecture [12], a
biologically inspired cognitive architecture instantiated entirely in – and constrained by properties of – spiking neurons.
From the perspective of artificial cognitive systems, these two aspects are
very compelling: on the one hand, it is increasingly possible to run even large
SNNs on state-of-the-art hardware and on the other, there is a large selection
of models of cognitive mechanisms in the cognitive sciences literature that are
instantiated in SNNs, from simple motor control [10] to higher cognition [14].
While there therefore is a large potential for bringing both strands together
in near-future artificial cognitive systems, some caveats remain. Here, we are
primarily interested in those arising from the fact that models of interesting
cognitive phenomena, by present-day necessity, are rarely instantiated on a physical agent; rather they are demonstrated in simulations that contain the necessary but also sufficient detail to highlight the model’s functionality. We therefore explore how the porting of such a model onto a physical platform would
work. We start from the aforementioned model of motor control [10], which was
demonstrated using a simulated robot arm with two degrees of freedom (DOF)
and operating in two dimensions and implement it on a physical robot arm (an
UR5e), adapting it for operation in three dimensions. We report on the process
as well as on performance comparisons. We chose this model because it captures
an essential aspect of sensorimotor cognition that may, at least according to
embodied theories of cognition, serve as a fundamental aspect of all cognition
[28]; it is therefore likely that most if not all cognitively interesting models on
real robots will need to implement similar aspects.
Since spiking neural networks in artificial cognitive systems is a nascent field,
this kind of exploration helps adjust expectations and gauge the possibilities and
limitations. In addition to illustrating the details of this particular implementation of a (in cognitive terms) fundamental system on a real platform, the paper
thus highlights the need for more explorations in this direction so that the potential of taking existing cognitive models onto a real machine operating in the real
world can be fully exploited in the future.
In the remainder of the paper, we first present a brief background on some of
the particular challenges that SNNs entail for robots and existing SNN models of
reaching task. We then present our implementation and reflect on the possibilities
and limitations this demonstrates.
177
this decision; for example whether the prime purpose of the agent is to be an
explanatory model of biological cognition or whether it takes biological inspiration merely to be successful at some task [25], or to what degree physiological
details are essential to cognition, as opposed to mere constraints by biology [24].
In this context, spiking neural networks (SNNs) have recently attracted interest in biologically inspired robotics. In part, this is driven by the recent advent
of neuromorphic hardware platforms that promise to be able to run relatively
large networks in real time (a fundamental requirement for any model operating on a real robotic platform) and, for example, the potential energy efficiency
such models may provide [6]. From a more cognitive perspective, however, it
is possible that SNNs are a suitable level of abstraction to adequately model
natural cognition. This is, for example, the assumption in the so-called neuroengineering framework [13] underlying the semantic pointer architecture [12], a
biologically inspired cognitive architecture instantiated entirely in – and constrained by properties of – spiking neurons.
From the perspective of artificial cognitive systems, these two aspects are
very compelling: on the one hand, it is increasingly possible to run even large
SNNs on state-of-the-art hardware and on the other, there is a large selection
of models of cognitive mechanisms in the cognitive sciences literature that are
instantiated in SNNs, from simple motor control [10] to higher cognition [14].
While there therefore is a large potential for bringing both strands together
in near-future artificial cognitive systems, some caveats remain. Here, we are
primarily interested in those arising from the fact that models of interesting
cognitive phenomena, by present-day necessity, are rarely instantiated on a physical agent; rather they are demonstrated in simulations that contain the necessary but also sufficient detail to highlight the model’s functionality. We therefore explore how the porting of such a model onto a physical platform would
work. We start from the aforementioned model of motor control [10], which was
demonstrated using a simulated robot arm with two degrees of freedom (DOF)
and operating in two dimensions and implement it on a physical robot arm (an
UR5e), adapting it for operation in three dimensions. We report on the process
as well as on performance comparisons. We chose this model because it captures
an essential aspect of sensorimotor cognition that may, at least according to
embodied theories of cognition, serve as a fundamental aspect of all cognition
[28]; it is therefore likely that most if not all cognitively interesting models on
real robots will need to implement similar aspects.
Since spiking neural networks in artificial cognitive systems is a nascent field,
this kind of exploration helps adjust expectations and gauge the possibilities and
limitations. In addition to illustrating the details of this particular implementation of a (in cognitive terms) fundamental system on a real platform, the paper
thus highlights the need for more explorations in this direction so that the potential of taking existing cognitive models onto a real machine operating in the real
world can be fully exploited in the future.
In the remainder of the paper, we first present a brief background on some of
the particular challenges that SNNs entail for robots and existing SNN models of
reaching task. We then present our implementation and reflect on the possibilities
and limitations this demonstrates.