From Models of Cognition to Robot Control and Back Using SNNs
189
5 Discussion and Conclusions
In this paper, we have adapted a biologically plausible adaptive spiking neural
network model of reaching [10] from a 2D, 2DOF simulated environment onto
a physical UR5e operating in 3D, albeit limited to 3DOF for this initial implementation. A major limitation is that we did not evaluate this implementation
extensively, in part because we did not implement the model on neuromorphic
hardware and used a standard computer instead, which incurs a time penalty. It
is therefore likely that future work could find stronger performance than we did
in this initial implementation. Nonetheless, we found that the model was able to
adapt to force field perturbations on the physical platform as one would expect
from the simulated results from the model itself, even when adapted to 3D operation. In that sense, the initial results are encouraging for future implementations
of models of cognitive mechanisms on physical platforms.
While the present model is arguably rather limited in any cognitively interesting sense since it merely reaches towards target positions, it does capture a
necessary part of sensorimotor cognition and as such forms a basis on which
to build more cognitively interesting agents. Grounding of cognitive processes
in sensorimotor interaction underlies both general models of cognition [2,12]
and specific robotic implementations of, for example, concept grounding [23].
In a larger context, endowing machines such as robots with artificial cognition
instantiated in SNNs will require the entire model to be captured by these spiking neurons, including the sensorimotor interactions. This is an aspect that is
often abstracted away or simplified in a model that only focusses on a specific
mechanism (such as concept grounding, see e.g.. [23]). If the aim is however
to achieve a complete robotic cognitive agent, then all these aspects need to
be captured in a manner that likely incorporates models such as the REACH
model [10]. While framweorks such as the NEF/SPA [12] provide the theory of
how this might be achieved, many challenges remain to be resolved in practice,
as demonstrated, for instance, by the present implementation of merely porting
REACH onto a physical platform operating in three dimensions.
Overall, we have therefore presented an initial implementation of a model that
can underlie much cognitive behaviour on a commonly used robotic platform,
noting in particular the necessary changes and adapations that were required
for this. While our current results are modest at best in terms of performance,
we demonstrate the potential that exists as well as details of using a UR5e
specifically in this context.
References
1. Andersen, T.T., Rasmussen, S., Exner, F., Steffen, L., Schnell, T.: Universal robots ros driver (2020). https://github.com/UniversalRobots/Universal
Robots ROS Driver
2. Barsalou, L.W.: Perceptual symbol systems. Behav. Brain Sci. 22(4), 577–660
(1999)
189
5 Discussion and Conclusions
In this paper, we have adapted a biologically plausible adaptive spiking neural
network model of reaching [10] from a 2D, 2DOF simulated environment onto
a physical UR5e operating in 3D, albeit limited to 3DOF for this initial implementation. A major limitation is that we did not evaluate this implementation
extensively, in part because we did not implement the model on neuromorphic
hardware and used a standard computer instead, which incurs a time penalty. It
is therefore likely that future work could find stronger performance than we did
in this initial implementation. Nonetheless, we found that the model was able to
adapt to force field perturbations on the physical platform as one would expect
from the simulated results from the model itself, even when adapted to 3D operation. In that sense, the initial results are encouraging for future implementations
of models of cognitive mechanisms on physical platforms.
While the present model is arguably rather limited in any cognitively interesting sense since it merely reaches towards target positions, it does capture a
necessary part of sensorimotor cognition and as such forms a basis on which
to build more cognitively interesting agents. Grounding of cognitive processes
in sensorimotor interaction underlies both general models of cognition [2,12]
and specific robotic implementations of, for example, concept grounding [23].
In a larger context, endowing machines such as robots with artificial cognition
instantiated in SNNs will require the entire model to be captured by these spiking neurons, including the sensorimotor interactions. This is an aspect that is
often abstracted away or simplified in a model that only focusses on a specific
mechanism (such as concept grounding, see e.g.. [23]). If the aim is however
to achieve a complete robotic cognitive agent, then all these aspects need to
be captured in a manner that likely incorporates models such as the REACH
model [10]. While framweorks such as the NEF/SPA [12] provide the theory of
how this might be achieved, many challenges remain to be resolved in practice,
as demonstrated, for instance, by the present implementation of merely porting
REACH onto a physical platform operating in three dimensions.
Overall, we have therefore presented an initial implementation of a model that
can underlie much cognitive behaviour on a commonly used robotic platform,
noting in particular the necessary changes and adapations that were required
for this. While our current results are modest at best in terms of performance,
we demonstrate the potential that exists as well as details of using a UR5e
specifically in this context.
References
1. Andersen, T.T., Rasmussen, S., Exner, F., Steffen, L., Schnell, T.: Universal robots ros driver (2020). https://github.com/UniversalRobots/Universal
Robots ROS Driver
2. Barsalou, L.W.: Perceptual symbol systems. Behav. Brain Sci. 22(4), 577–660
(1999)