Hybrid model frameworks combining synaptic plasticity-dependent neural firing with simple biomechanics at speeds faster than real time illustrate how invertebrate learning directly inspires “intelligent” robotics [114,115]. Such frameworks exploit a multifunctional model of Aplysia feeding rhythms, which are capable of repeatedly reproducing three types of behavior: biting, swallowing, and rejecting. These simulate behavioral switching in response to external sensory cues. Model approaches incorporate synaptic learning and neural connectivity in a simple mechanical model of the feeding apparatus [116], with testable hypotheses in the context of robot movement control. 

Learning-induced synaptic plasticity in such modular circuitry controls behavioral switching (Figure 7), as recently simulated in biologically inspired model approaches directly exploitable for multifunctional robot control. For the model equations, the reader is referred to [116,117,118,119,120]. This modeling framework can be extended to a variety of scenarios for multifunctional robot movement and rhythm control, and it has several advantages. It allows rapid simulation of multifunctional behavior and it includes the known functional circuitry and simplified biomechanics of peripheral anatomy. The direct relationship with the underlying neural circuitry makes it possible to both generate and test specific neurobiological hypotheses. The relative simplicity of the network (Figure 7) makes it attractive as a basis for robot control. Unlike other artificial neural network architectures, synthetic nervous systems are explainable in terms of structures directly informing the functional system output [121,122,123,124,125]. Although the connections and trained weights of other artificial neural networks may provide similar control capabilities, these, unlike synthetic nervous systems, have to be trained on large datasets. The very strength of synthetic nervous systems is that they use a restricted, functionally identified set of biological neuron dynamics, thereby generating robust control without the need for additional training [125]. Neural network learning-inspired robotics include reactive systems emulating reflexes, neural oscillators to generate movement patterns, and neural networks for filtering high-dimensional sensory information [126]. To such effect, biologically motivated neural-network-based robot controllers, inspired by control structures in the sensory brain, where information is routed through the network using facilitating dynamic synapses with short-term plasticity, have been proposed [123,124,125,126,127,128]. Learning occurs through long-term synaptic plasticity using temporal difference learning rules to enable the robot to learn to associate a given movement with the correct, i.e., appropriate as defined, input conditions. Self-organizing network dynamics [127,129] provide memory representations of the environments that the robot encounters.

Recent progress in neuromorphic sensory systems which mimic the biological receptor functions and sensorial processing [129,130,131,132] trends toward sensors and sensory systems that communicate through asynchronous digital signals analogous to neural spikes [127], improving the performance of sensors and suggesting novel sensory processing principles that exploit spike timing [128], leading to experiments in robotics and human–robot interaction that can impact how we think the brain processes sensory information. Sensory memory is formed at the earliest stages of neural processing (Figure 8), underlying perception and interaction of an agent with the environment. Sensory memory is based on the same plasticity principles as all true learning, and it is, therefore, an important source of intelligence in a more general sense. Sensory memory is consolidated while perceiving and interacting with the environment, and a primary source of intelligence in all living species. Transferring such biological concepts into electronic implementation aims at achieving perceptual intelligence, which would profoundly advance a broad spectrum of applications, such as prosthetics, robotics, and cyborg systems [129]. Moreover, transferring biologically intelligent sensory processing into electronic implementations [130,131,132] achieves new forms of perceptual intelligence (Figure 8). These have the potential to profoundly advance a broader spectrum of applications in robotics, artificial intelligence, and control systems.

These new, bioinspired systems offer unprecedented opportunities for hardware architecture solutions coupled with artificial intelligence, with the potential of extending the capabilities of current digital systems to psychological attributes such as sensations and emotions. Challenges to be met in this field concern integration levels, energy efficiency, and functionality to shed light on the translational potential of such implementations. Neuronal activity and the development of functionally specific neural networks in the brain are continuously modulated by sensory signals. The somatosensory cortical network [133] in the primate brain refers to a neocortical area that responds primarily to tactile stimulation of skin or hair. This cortical area is conceptualized in the current state of the art [133,134,135,136] as containing a single map of the receptor periphery, connected to a cortical neural network with modular functional architecture and connectivity binding functionally distinct neuronal subpopulations from other cortical areas into motor circuit modules at several hierarchical levels [133,134,135,136]. These functional modules display a hierarchy of interleaved circuits connecting, via interneurons in the spinal cord, to visual and auditory sensory areas, and to the motor cortex, with feedback loops and bilateral communication with the supraspinal centers [135,136,137]. This enables ’from-local-to-global’ functional organization [134], a ground condition for self-organization [106,107], with plastic connectivity patterns that are correlated with specific behavioral variations such as variations in motor output or grip force, which fulfills an important adaptive role and ensures that humans are able to reliably grasp and manipulate objects in the physical world under constantly changing conditions in their immediate sensory environment. Neuroscience-inspired biosensor technology for the development of robot-assisted minimally invasive surgical training [138,139,140,141,142,143] is a currently relevant field of application here as it has direct clinical, ergonomic, and functional implications, with clearly identified advantages over traditional surgical procedures [144,145]. Individual grip force profiling using wireless wearable (gloves or glove-like assemblies) sensor systems for the monitoring of task skill parameters [138,139,140,141] and their evolution in real time on robotic surgery platforms [141,142,143,146,147,148] permits studying the learning curves [140,141,142] of experienced robotic surgeons, surgeons with experience as robotic platform tableside assistants, surgeons with laparoscopic experience, surgeons without laparoscopic experience, and complete novices. Grip force monitoring in robotic surgery [146,147,148] is a highly useful means of tracking the evolution of the surgeon’s individual force profile during task execution. Multimodal feedback systems may represent a slight advantage over the not very effective traditional feedback solutions, and the monitoring of individual grip forces of a surgeon or a trainee in robotic task execution through wearable multisensory systems is by far the superior solution, as real-time grip force profiling by such wearable systems can directly help prevent incidents [146,147] because it includes the possibility of sending a signal (sound or light) to the surgeon before their grip force exceeds a critical limit, and damage occurs. Proficiency, or expertise, in the control of a robotic system for minimally invasive surgery is reflected by a lesser grip force during task execution, as well as by a shorter task execution times [146,147,148]. Grip forces self-organize progressively in a way that is similar to the self-organization of neural oscillations during task learning, and, in surgical human–robot interaction, a self-organizing neural network model was found to reliably account for grip force expertise [149].

To move neural processing models for robotics beyond reactive behavior, the capacity to selectively filter relevant sensory input and to autonomously generate sequences of processing steps is critical, as in cases where a robot has to search for specific visual objects in the environment, and then reach for these objects in a specific, instructed serial order [150,151]. In robotic tasks where the simultaneous control of object dynamics and internal forces exerted by the robot limb(s) to follow a trajectory with the object attached to it is required, plasticity and adaptation permit to deal with external perturbations acting on the robot–object system. On the basis of mere feedback through the internal dynamics of an object, a robot is, like a human, able to relate to specific objects with a very specific sensorimotor pattern. When the object-specific dynamical patterns are combined with hand coordinates obtained from a camera, dedicated hand-eye coordination self-organizes spontaneously [152,153,154] without any higher-order cognitive control. Robots are currently not capable of any form of genuine cognition. Cognition controls behavior in living brains, where sensing and acting are no longer linked directly to ensure control, as is the case for any robot currently, including humanoids. When an action is based on sensory information that is no longer directly available in the processing loop at the time where action is to ensue, the relevant information must be represented in a memory structure, as it is in any living brain. Information for the control of action then becomes abstracted from sensor data through the neural memory representations and mechanisms of memory-based decision making [150]. Plastic mechanisms in neural network-based control architectures (Figure 9) effectively contribute to the learning of dynamics of robot–object systems, enabling adaptive corrections and/or offset detection.

This allows for progressive error reduction by incorporating distributed synaptic plasticity according to feedback from actual movements in the given environment. It has been shown previously that such feedback processes are omnipresent in voluntary motor actions of human agents [154], where rapid corrective responses occur even for very small disturbances that approach the natural variability of limb movements. Robot control toward autonomy [155] ultimately implies that the robot generalizes across time and space, is capable of stopping when an element is missing, and updates a planned action sequence autonomously in real time when a scenario suddenly changes. Using biologically plausible neural learning, the flow of behavior generated can emerge new neural system dynamics through self-organization without any further control or supervision algorithm(s). In robotic control based on biologically inspired neural network learning, the universal training method is based on Hebbian synaptic learning. Several variants of the latter are discussed and compared, with detailed equations, in [150]. The neural network dynamics described therein can, in principle, be combined with other network structures that receive reward information as part of their input in an extended model approach based on known functional dynamics of the mammalian brain. As discussed in previous sections, the orbitofrontal cortex receives converging sensory and reward inputs, which makes the network as a whole capable of acquiring task structure and support reinforcement learning by encoding combinations of sensory and reward events [81,82].

Such networks possess self-organizing state-encoding dynamics of the type shown here above, based on the principle of multiple-stage decision tasks, where a human agent or robot has to choose between decisions (options) and their consequences. More knowledge from and interaction between the fields of cognitive neuroscience and robotics are needed here to further explore existing possibilities.