It is since 1971, when Leon Chua predicted the existence of the memristor (a nonlinear resistor with memory) as the “missing circuit element”, followed 37 years later by the first physical implementation of a monolithic memristor device achieved in the HP labs, that scientists look at the resemblance of these device characteristics to those of biological neurons and synapses, enablers of neuromorphic computing architectures. 

Building upon these studies, an attractive line of research is devoted to the use of these two-terminal devices in mimicking functionalities that are known to be characteristic features of the human brain, and often regarded as the foundation of intelligence, such as the ability of learning from and adapting to external stimuli by adapting the human behaviour by leveraging experience, managing risks and taking decisions in real-time while facing unknown events or environments.

Despite intense research carried out by scientists and engineers to try and replicate this behavior in artificial neural networks and brain-like computing architectures, it proves extremely difficult for a robot or an autonomous guided vehicle to operate out from the lab in unstructured environments, when facing events and processing stimuli not experienced before, in order to perform simple tasks accurately and successfully since the first attempt.

In our paper entitled “Memristor-based adaptive neuromorphic perception in unstructured environments”, we introduce a differential neuromorphic computing architecture which leverages the bifurcation of memristor states associated with sensory stimuli to achieve fine-grained adaptive sensing capabilities and use memristor device characteristics to reproduce functions that in nature lead to perception abilities and support intelligent behaviour.

We demonstrated this capability, in two practical use case scenarios.

In a first scenario, we combined a multilayered non-volatile memristor device with a tactile force sensor architecture to provide a robotic hand with the ability of self-adapting to the surface features of an unknown object in order to secure a safe and stable grasp of both sharp and slippery objects - a functionality that mimic the action of tactile receptors and sensory neurons involved in tactile stimuli perception, such as nociceptors, fast-adapting, and  slow-adapting  receptors, along  with  their  respective neural pathways.

In a second use-case scenario, we demonstrate the effectiveness of a memristors array-based architecture to control navigation in an unknown environment by recognising and reacting to rapidly moving obstacles under different light changing conditions into both planned and free driving conditions – a functionality that mimic the ability of expert drivers to rapidly recognise and locate dangerous situations with sudden moving obstacles and react accordingly. 

We believe that this work paves the way towards further developments of human-like information processing pipelines, opening the possibility for intelligent machines to operate in unstructured environments efficiently.