Distributed Adaptive Control: An Ideal Cognitive Architecture Candidate
157
plan recalling by sensory matching and internal chaining. Thus, the Contextual Layer
shows action-dependent learning as observed in operant conditioning, allowing abilities
such as allocentric-based trajectory planning, crucial for the mobile robot’s navigation.
In addition to this layered horizontal organisation, the architecture is also vertically distributed across three columns: states of the world obtained by exosensing, states of the
self obtained by endosensing and their interaction through action.
The DAC architecture fulfils both the theoretical and practical criteria that must
be considered when designing a proper cognitive architecture. On the one hand, at the
theoretical level, DAC applies: a) biologically-inspired learning rules such as Hebbian
learning, Oja learning rule or associative competition for different purposes, b) provides
a solution to the fundamental Symbol Grounding Problem by acquiring the state space
of the agent, based on its interaction with the environment, c) escapes from the now by
generating behavioural plans or policies based on sensory matching with representations
of environment and action states (stored in memory systems), and d) reverse the Referential Indeterminacy Problem, in which the agent has to extract the external concept
that was referred, by endowing the system with proactivity to acquire knowledge. On
the other hand, DAC has been validated in both single-agent (i.e. robots) and large-scale
levels [14–17]. These implementations have supported the adequacy of DAC, from a
pragmatic point of view, to perform a diverse set of tasks including foraging, object
manipulation or Human-Robot Interaction (HRI).
4 DAC at the Single-Agent Level
Within the recycling context, we propose two types of robots working as single agents
that allow for the transportation, disassembling and classification of the different WEEE
components. The two robot categories are robotic grippers and mobile robots. We propose that the implementation of DAC does not differ between robots. However, it needs
to be adapted for the different tasks these robots perform, based on the data provided by
the different sensors, the needs and goals, and the robots’ actuators.
In mobile robots, the needs range in different dimensions depending on their related
aCell. Based on the project, we define as aCell the workbench related to a specific worker
equipped with a robotic arm where different WEEE is disassembled. Mobile robots aim to
transport WEEE in an adaptative way; for instance, taking into consideration the disposal
of materials or the specific disassembled and classified components. Here, the Reactive
Layer is responsible for driving the needs of the robots towards different navigation
patterns as well as pick up and place behaviours, that are carried out by the actuators
(i.e. motors of the wheels and the lifting platforms). Sensors such as wheels’ encoders,
proximity sensors and RGB cameras provide the Reactive Layer with the information
needed to trigger reflexive behaviours, and the Adaptive Layer with the information
required to learn associations, which in turn, assist in the behavioural policies formation
by the Contextual Layer.
In the robotic grippers, the needs change from navigation-oriented to motor controloriented goals, since its porpoise is assist in the disassembling procedure itself. Here, the
Reactive Layer processes information regarding pressure, proximity and torque sensors,
along with data provided by a camera and triggers reflexive behaviours (like a stop
157
plan recalling by sensory matching and internal chaining. Thus, the Contextual Layer
shows action-dependent learning as observed in operant conditioning, allowing abilities
such as allocentric-based trajectory planning, crucial for the mobile robot’s navigation.
In addition to this layered horizontal organisation, the architecture is also vertically distributed across three columns: states of the world obtained by exosensing, states of the
self obtained by endosensing and their interaction through action.
The DAC architecture fulfils both the theoretical and practical criteria that must
be considered when designing a proper cognitive architecture. On the one hand, at the
theoretical level, DAC applies: a) biologically-inspired learning rules such as Hebbian
learning, Oja learning rule or associative competition for different purposes, b) provides
a solution to the fundamental Symbol Grounding Problem by acquiring the state space
of the agent, based on its interaction with the environment, c) escapes from the now by
generating behavioural plans or policies based on sensory matching with representations
of environment and action states (stored in memory systems), and d) reverse the Referential Indeterminacy Problem, in which the agent has to extract the external concept
that was referred, by endowing the system with proactivity to acquire knowledge. On
the other hand, DAC has been validated in both single-agent (i.e. robots) and large-scale
levels [14–17]. These implementations have supported the adequacy of DAC, from a
pragmatic point of view, to perform a diverse set of tasks including foraging, object
manipulation or Human-Robot Interaction (HRI).
4 DAC at the Single-Agent Level
Within the recycling context, we propose two types of robots working as single agents
that allow for the transportation, disassembling and classification of the different WEEE
components. The two robot categories are robotic grippers and mobile robots. We propose that the implementation of DAC does not differ between robots. However, it needs
to be adapted for the different tasks these robots perform, based on the data provided by
the different sensors, the needs and goals, and the robots’ actuators.
In mobile robots, the needs range in different dimensions depending on their related
aCell. Based on the project, we define as aCell the workbench related to a specific worker
equipped with a robotic arm where different WEEE is disassembled. Mobile robots aim to
transport WEEE in an adaptative way; for instance, taking into consideration the disposal
of materials or the specific disassembled and classified components. Here, the Reactive
Layer is responsible for driving the needs of the robots towards different navigation
patterns as well as pick up and place behaviours, that are carried out by the actuators
(i.e. motors of the wheels and the lifting platforms). Sensors such as wheels’ encoders,
proximity sensors and RGB cameras provide the Reactive Layer with the information
needed to trigger reflexive behaviours, and the Adaptive Layer with the information
required to learn associations, which in turn, assist in the behavioural policies formation
by the Contextual Layer.
In the robotic grippers, the needs change from navigation-oriented to motor controloriented goals, since its porpoise is assist in the disassembling procedure itself. Here, the
Reactive Layer processes information regarding pressure, proximity and torque sensors,
along with data provided by a camera and triggers reflexive behaviours (like a stop