158
O. G. Rosado and P. F. M. J. Verschure
signal for safety). This sensory information forms progressively sensory representations
by associative learning at the Adaptive Layer and these associations will then be stored in
Contextual Layer modules if the associated behaviours allow reaching goal states. Thus,
robotic grippers are not just able to perform reflexive actions such as unscrewing. When
endowed with adaptive and contextual capabilities, they could, for example, correctly
locate the bolt, apply the appropriate velocity and pressure, and predict when the bolt is
going to be unscrewed.
At single-agent level, we consider three essential areas for the successful implementation of a recycling plant where robots work alongside humans for e-waste disassembly:
Navigation, HRI and Motor Control. DAC has been tested in these areas supporting its
implementation in both mobile and humanoid robots.
4.1 Previous Implementations of DAC: Navigation
Robotic navigation has already been achieved without requiring a cognitive architecture.
However, the characteristic of the Industry 4.0 context demands goal-oriented navigation,
which adapts to changing environments, needs to be aware of the material transported
and the state of the aCell, and ensures the safety of other robots and humans along the
trajectory. Due to the complexity of navigation in this context, an architectural approach is more suitable, and the DAC architecture has largely demonstrated its strengths
performing foraging tasks with mobile robots.
Aiming to prove that the different computational models proposed by DAC account
for functional mapping of specific brain areas and work complementing each other
when the system operates as a whole, Maffei et al. [15] embed the version DAC-X in
a mobile robot performing a foraging task. In this study, the Somatic Layer computed
input signals from the robot’s sensors, while the output was calculated as the total motor
signal provided by the architecture. Finally, the actions were constrained by the robot’s
body morphology. The Reactive Layer reflexively mapped sensory states into actions by
using feedback controllers that approximated the role of the Brainstem nuclei. Reflexive
object avoidance, visual target orientation and computation of bodily states such as needs
and drives were obtained by computational models, mimicking the Trigeminal Nucleus,
Superior Colliculus and Hypothalamus functions respectively. In the Adaptive Layer, a
model of the cerebellar microcircuit allowed associative learning by coupling neutral
sensory cues with adaptive responses. The motivation for action arose by modelling
the Ventral Tegmental Area for the computation of low-level internal states, and actionselection for behavioural plans was achieved by modelling the Basal Ganglia. Finally, the
memory systems of the Contextual Layer comprised a biologically constrained model
of the Hippocampus by which the agent acquired an internal representation of the environment; and a model of the Prefrontal Cortex that included mechanisms for storing
decision-making and goal-dependent information.
Analysis that discretised the behavioural performance into three phases (early, middle
and late trials) showed that at the beginning, the naive agent relies on its reflexes to
explore the arena and seek for resources. This initial navigation emerges primarily from
the work of the Reactive Layer, resulting in a stochastic trajectory pattern that covered
a large part of the arena and had a low item collection rate. After a few trials, the
Adaptive Layer took advantage of the local visual landmarks deployed in the floor,
O. G. Rosado and P. F. M. J. Verschure
signal for safety). This sensory information forms progressively sensory representations
by associative learning at the Adaptive Layer and these associations will then be stored in
Contextual Layer modules if the associated behaviours allow reaching goal states. Thus,
robotic grippers are not just able to perform reflexive actions such as unscrewing. When
endowed with adaptive and contextual capabilities, they could, for example, correctly
locate the bolt, apply the appropriate velocity and pressure, and predict when the bolt is
going to be unscrewed.
At single-agent level, we consider three essential areas for the successful implementation of a recycling plant where robots work alongside humans for e-waste disassembly:
Navigation, HRI and Motor Control. DAC has been tested in these areas supporting its
implementation in both mobile and humanoid robots.
4.1 Previous Implementations of DAC: Navigation
Robotic navigation has already been achieved without requiring a cognitive architecture.
However, the characteristic of the Industry 4.0 context demands goal-oriented navigation,
which adapts to changing environments, needs to be aware of the material transported
and the state of the aCell, and ensures the safety of other robots and humans along the
trajectory. Due to the complexity of navigation in this context, an architectural approach is more suitable, and the DAC architecture has largely demonstrated its strengths
performing foraging tasks with mobile robots.
Aiming to prove that the different computational models proposed by DAC account
for functional mapping of specific brain areas and work complementing each other
when the system operates as a whole, Maffei et al. [15] embed the version DAC-X in
a mobile robot performing a foraging task. In this study, the Somatic Layer computed
input signals from the robot’s sensors, while the output was calculated as the total motor
signal provided by the architecture. Finally, the actions were constrained by the robot’s
body morphology. The Reactive Layer reflexively mapped sensory states into actions by
using feedback controllers that approximated the role of the Brainstem nuclei. Reflexive
object avoidance, visual target orientation and computation of bodily states such as needs
and drives were obtained by computational models, mimicking the Trigeminal Nucleus,
Superior Colliculus and Hypothalamus functions respectively. In the Adaptive Layer, a
model of the cerebellar microcircuit allowed associative learning by coupling neutral
sensory cues with adaptive responses. The motivation for action arose by modelling
the Ventral Tegmental Area for the computation of low-level internal states, and actionselection for behavioural plans was achieved by modelling the Basal Ganglia. Finally, the
memory systems of the Contextual Layer comprised a biologically constrained model
of the Hippocampus by which the agent acquired an internal representation of the environment; and a model of the Prefrontal Cortex that included mechanisms for storing
decision-making and goal-dependent information.
Analysis that discretised the behavioural performance into three phases (early, middle
and late trials) showed that at the beginning, the naive agent relies on its reflexes to
explore the arena and seek for resources. This initial navigation emerges primarily from
the work of the Reactive Layer, resulting in a stochastic trajectory pattern that covered
a large part of the arena and had a low item collection rate. After a few trials, the
Adaptive Layer took advantage of the local visual landmarks deployed in the floor,