Robots are widely used in both everyday life and professional fields, ranging from assisting with household chores and conducting surgeries in healthcare to optimizing manufacturing processes in industries. Their applications consistently evolve, branching into emergent fields^[1−5]. However, there is still a lack of truly intelligent and agile robots^{[6, 7]}. The actions and behavioral capabilities achieved by current robots are primarily controlled through predefined rules. They lack multi-modal perception akin to the human brain and the ability for brain-like autonomous decision-making based on spatiotemporal information^{[8, 9]}. Consequently, they struggle to perform tasks and operations with the same precision, stability, and adaptability as humans in complex and variable environments. Making robots into more human-like and capable assistants is challenging but promising.

The advent of cutting-edge achievements in neuroscience provides an effective pathway to address this challenge by deeply integrating neuroscience with robotics^{[10, 11]}. Researchers are imitating essential neural mechanisms of humans from the inside out, focusing on mechanisms and structures. So far, significant progress has been made in exploring neural mechanisms, including visual perception, motor control, and cooperative coding^[12−14]. Brain-inspired software and hardware systems are established through computational modeling, device simulation, etc. At the software level, algorithms to merger associative memory mechanisms and attentional control have been proposed, enhancing the robustness of object recognition for robots in complex natural environments, real-world spatial orientations, and multi-obstacle scenarios^[15−17]. These days, the software algorithm models for brain-like simulations have attained a mature stage. However, the computational power demand still encounters significant challenges since information technology progresses into the era of artificial intelligence, big data, and cloud computing. The reason behind this challenge is the existence of storage and power consumption issues due to the bottleneck of the Von Neumann architecture, in which memory and processing units are separate^[18−20]. Therefore, at the hardware level, the advantages of integrating storage and computation are more pronounced using electronic circuits and brain-like devices that simulate neurons and synapse characteristics. The neuromorphic computing circuits structures are promising in replacing the traditional Von Neumann computing architecture, significantly improves computational performance and demonstrating advantages in low-power processing control^[21−26].

In the neuromorphic computing circuits, primary requirement for simulating the plasticity of biological neural synapses is the ability to continuously adjust the conductance (or resistance) of artificial synapse devices, similar to biological synapses^{[27, 28]}. Unlike the existing CMOS-based memory technologies that rely on volatile capacitor states, the successful development of nanoscale memristive devices (also known as memristors) has sparked an analog revolution^[29−31]. These devices can alter their resistance values based on the amount of charge flowing through them and maintain these changes even after power-off, offering the potential to construct hardware that mimics neural structures. Based on this feature, it has been demonstrated that neuromorphic circuits utilizing memristors can unlock circuit functionalities that traditional electronic products can hardly achieve^{[26, 32−39]}. Especially, combining the natural biomimetic characteristics of memristors, the analog circuit platform based on memristors, and the mixed-mode control platform using memristors as analog-to-digital conversion interfaces are effective approaches to break the Von Neumann bottleneck in this field.

Considering the current research on memristors in the field of biomimicry, this review focuses primarily on the operation of memristor-based neuromorphic circuits in emulating the brain's information processing and controlling drives. A detailed comparison in parallel between the human nervous system and neuromorphic circuit systems based on memristors is illustrated by the schematic in Fig. 1^[40−42]. The correspondence between the two in terms of perception, information processing, and decision-making is as follows: The physical environmental information perceived by sensory neurons in the human nervous system can be collected through analog sensors. The neuromorphic circuits based on memristors build hardware neural networks through four approaches: analog-digital hybrid platforms, innovative device designs, multiple control mechanisms, and array structures. These circuits mimic synaptic connections in the human brain and process signal integration. Similarly, motor neurons in the human nervous system can be simulated by bio-inspired drive systems based on memristors or traditional electromechanical actuators. The generated outcomes can be sensed and feedback to the hardware neural networks, thereby achieving complete biomimetic system control over robot behavior.

In this current review, we initially provided a brief introduction to mainstream research on the principles of memristors, along with the device structures designed based on their functionalities and their biomimetic characteristics. Following this, the review will delve into the biomimetic applications achieved by memristor-based neuromorphic systems employing novel devices and architectures. These applications mainly involve cognitive learning based on synaptic plasticity, brain-like information processing for multimodal perception and spatiotemporal information processing, as well as biomimetic control systems based on both local and global paradigms. Finally, a conclusion and future prospects are then discussed.

The memristor is the basic unit used in the neuromorphic circuit system as shown in Fig. 1. There are different types of memristors, including ionic migration memristors, spin memristors, ferroelectric memristors, and phase-change memristors^[47−56]. The specific switching mechanisms of memristors determine the electrical characteristics of neuromorphic devices^[57−63]. Within this section, an overview of various memristors switching mechanisms will be summarized.

Ionic migration memristors consist of a dielectric layer structure sandwiched between two metal electrodes. The resistance variation is caused by the migration of cations (silver and copper cations) and anions (oxygen, halide, nitride, and sulfurs ions)^{[29, 64]}. According to the specific mechanism of carrier migration within the dielectric layer, it can be categorized into conductive filamentary type, interface switching type, and ionic gating type based on electrochemical doping. The schematic of an electrochemical metallization (ECM) memristor based on cation migration, comprising an active metal electrode/electrolyte thin film/inertial metal electrode structure, is illustrated in Fig. 2(a). When applying an external electric field, the separation of active metal cations, it will induce an electrochemical reaction that forms a conductive bridge between the active metal electrode and the inert electrode. Consequently, this adjustment influences the conductivity of the device. Displayed in Fig. 2(b) is the valence change mechanism (VCM) memristor, which operates via anion migration and exhibits a structural similarity to ECM memristor. But in this case, both metal electrodes are inert. Under the influence of an applied electric field, anionic vacancies migrate within the insulator of VCM memristors, altering the valence state of metal cations and forming defects that act as conductive filaments, consequently relizing adjustments in the device's conductivity. Fig. 2(c) depicts a schematic of interface transitions within non-filamentary devices, which are also based on the mechanism of ion migration. When subjected to an external electric field, variations in its conductivity occur as a result of ion migration, altering the configuration of defect structures at the interface between the metal electrode and the conversion layer, subsequently modifying the Schottky barrier at the interface^{[49, 51, 52]}. At presented in Fig. 2(d), the last type of memristor, which operates on the mechanism of ion migration, known as the ionic gating memristor employing electrochemical doping. Its mechanism for altering device conductivity is realized by applying a voltage to the gate electrode, which drives ions to migriate into the channel materials, causing electrochemical redox reactions. This process leads to carrier doping at the interface between the ionic liquid and various materials^{[50, 65]}.

Differing from the ion migration mechanism, the memristive switches of spin memristors rely on magnetic tunnel junction structures, ferroelectric memristors is based on ferroelectric tunnel junction structures, and phase-change memristors display physical behavior associated with tunneling magnetoresistance effect, quantum tunneling effect and resistance inversion effect of electrons^[54]. The operational process of a spin memristor with a magnetic tunnel junction (MTJ) structure is shown in Fig. 2(e), where there is a spacer layer located between two ferromagnetic layers. The thicker layer serves as the fixed layer, while the thinner one works as the free layer. The conductivity of the spain memristor channel is modulated by the polarized current passing through the MTJ, which utilizes the tunneling magnetoresistance phenomenon to adjust the arrangement of the free layer's spins. Moreover, the tunneling probability of electrons is associated with the magnetization direction of the two ferromagnetic layers. Likewise configured with a tunneling junction as its fundamental unit, a typical ferroelectric memristor featuring a ferroelectric tunnel junction (FTJ) structure is depicted in Fig. 2(f). This device consists of metal/ferroelectric thin film/metal layers. The tunneling barrier and electron tunneling rate are altered by the external electric field-induced ferroelectric polarization, achieving the memristive effect by adjusting the tunnel resistance value. Phase-change memristors, on the other hand, usually have a structure comprising a layer of phase-change material sandwiched between two electrodes. This phase-change material can exist in at least two stable^{[47, 48, 54]}. Fig. 2(g) demonstrates the transition process between the crystalline and amorphous states in a phase-change memristor, which is controlled by the applied heating time.

Memristors possess reconfigurable properties and inherently exhibits synaptic plasticity and nonlinear response characteristics at the device level, enabling them to express biomimetic functions. However, traditional memristors are hard to achieve more biomimetic processing functions due to their limited degrees of freedom. In response to this challenge, proposed advancements started with making multiple modulation devices and heterogeneous synapse devices, aiming to fabricate more realistic basic units characteristized by intricate spatiotemporal dynamics. From the illustration in Fig. 3, there are various types of memristors, including traditional two-terminal memristors^[66−69], three-terminal memristors^{[79, 81−83]}, and multi-terminal memristor devices^{[55, 56, 69−73]}, as well as memristor arrays constructed using these devices^{[66, 74, 75]}. As the degrees of freedom increases with device structure, memristor-based neuromorphic circuits can mimic more complex biomimetic features. In addition to the traditional adjustability of memristor conductance values, these devices exhibit characteristics ranging from short-term to long-term synaptic memory transition^{[68, 76]}. Moreover, multi-modulation devices^{[76, 77]} or multi-terminal structured devices can potentially enable neuromorphic circuits based on memristors to achieve multimodal, spatiotemporal integration, and parallel computing, similar to certain aspects of brain-like functionalities. Different configurations of memristors are employed in designing neuromorphic hardware circuits, and it becomes feasible to partially realize the human-like neural system functions depicted in Fig. 1.

Contrary to the traditional Von Neumann architecture of computers, neuromorphic computing based on memristors integrates computation and storage in the same device or circuit. This fusion of information storage and processing enhances the speed and efficiency of information processing, which resembles the information processing mechanism of human brain. It seems that imitating the information processing patterns of the nervous system lays the foundation for developing energy-efficient and multifunctional neuromorphic applications^[85]. Thus, this section focuses on discussing the advantages brought to the field of robotics by memristor-based neuromorphic systems employing various devices and peripheral structures.

In particular, a summary of the current efforts in memristor-based neuromorphic circuits related to robot-environment interaction, associative learning, autonomous motion control with collaboration between the brain and cerebellum, multi-input information processing, and brain-like parallel computation is provided in this section. These efforts are inspired by biological behaviors (associative learning) or brain characteristics (neurons, synapses, brain-cerebellum coordination). The discussion further underscores the significant difference between memristor-based neuromorphic circuits and conventional methods in controlling robots. For example, memristors’ synaptic plasticity can enable robots to adapt to environments and learn new tasks independently, making them more flexible without strictly programmed rules. By adopting the collaborative autonomous motion control logic of the brain and cerebellum, memristor-based analog-digital hybrid control platforms show potential in addressing the aforementioned features of low energy consumption and low latency. Through the utilization of biomimetic brain neurons for multiple signal integration and parallel computing, the development of the field of robot control with more complex and efficient information processing capabilities is achieved by employing various memristor device structures and peripheral circuit configurations.

The use of neuromorphic circuits based on memristor for brain-like information processing proves to be highly effective. Therefore, this section supplements and summarizes applications that endow robots with learning capabilities of handling complex signals processing involving multiple and spatiotemporal information, and exploring applications based on memristor arrays via peripheral circuits or novel device designs for parallel computing and control.

The synapse stands as a manifestation of connectivity between preceding and succeeding neurons, thereby serving as a representation of how robots engage in rule-based learning. Within the realm of cognitive learning platform based on memristor, the focus is on establishing connectionism applications between stimuli and expected actions. Moreover, the utilization of the memristor’s capability to transition from short-term memory to long-term memory allows robots to adjust their actions in diverse environmental contexts, which is similar to experiential learning for biology.

The connectionism based on supervised learning and reinforcement learning: Given the memristor’s capacity to modify its conductivity following external stimuli and maintain these alterations post-power-off, contemporary investigations are centered on hardware neural networks using memristors to train robots. Supervised learning and reinforcement learning stand as two pivotal methods in robot control, and connectionism-based robots achieve cognitive learning in diverse environments via these methods.

Supervised learning trains models using known inputs and expected outputs, enabling robots to predict and execute tasks. Wang et al.'s work on the Braitenberg vehicle based on memristors explains learning as establishing connections between stimuli and biological responses^[67]. The Braitenberg vehicle is a conceptual model used to describe the responsive behavior of animals or insects to stimuli, and it is widely used in developing autonomous vehicles adaptable to diverse environments (Fig. 4(a) (i)). Neuromorphic processors were usually used for the vehicle's software, but no demonstration of a Braitenberg vehicle based on neuromorphic hardware has been shown until now. In the work by Wang et al., adaptive changes in the vehicle's behavior, derived from connectionism were achieved through the non-volatile and reconfigurable weights of memristors using supervised learning. The neuromorphic circuit centers around a memristor array serving as its processing core. It utilizes two analog grayscale sensor signals for input processing and the vehicle's servo motor for motion behavior. Processing the analog sensor signals by neuromorphic circuit eliminates the need for analog-to-digital conversion. By adjusting the conductance values of the memristor array through supervised learning update rules, effective learning guided by supervised signals occurs (Fig. 4(a) (ii)). This generates control signals that drive the servo motor for accurate steering and learning the correct tracing rules. Fig. 4(a) (iii) shows the flowchart of the supervised learning process for the Braitenberg vehicle, along with the circuit schematic for steering the servo motor which resembles the components of a biological system: sensors, neural system, and effectors. This research demonstrates the feasibility of using neuromorphic circuits based on memristor for implementing Braitenberg vehicles and suggests a potential direction for creating smarter and more adaptable robots.

Unlike supervised learning, the uniqueness of reinforcement learning allows robots to learn by interacting with the environment without the need for labeled data. Robots can adjust their strategies based on the rewards or penalties, gradually improving their performance. This learning approach allows robots to learn through trial and error, forming optimized behavioral patterns over time, beneficial for controlling robot in complex tasks and uncertain environments. It enables autonomous learning without prior knowledge and facilitates adaptive responses to new challenges. Similarly inspired by Braitenberg, path-planning robot with an organic neuromorphic circuit for sensorimotor integration^[83] employs a perception-action system feedback signal representing environmental stimuli in reinforcement learning to train an organic neuromorphic circuit processor. This achieves in-situ navigation learning within a two-dimensional maze, forming the correct associative relationship between sensory signals and motor responses. A trainable voltage divider circuit with neuromorphic devices processes sensory signals and punishment signals in reinforcement learning. The output of this circuit guides the car's steering (Fig. 4(b) (i)) based on probability. Before training, the car was unable to respond effectively to white labels. However, during training, instances of the car deviating from the planned path or reaching the maze boundaries incorrectly increased. This resulted in external mechanical stimuli from colliding with walls providing a coherent feed back to the organic transistor gate of the voltage divider circuit. This feedback reinforced the visuo-motor associative connection, (Fig. 4(b) (iii)) specifically the identification of white labels and correct movement behavior.

Both works aim to establish associations between perception and motion. However, the Braitenberg vehicle neuromorphic circuit work relies more on predefined neural network structures and supervised signal training, while Krauhausen et al.'s path-planning robot emphasizes the robot's interaction within the environment and the trial-and-error learning process (Fig. 4(b) (ii)). The demonstrated autonomous learning and adaptability capabilities based on memristor neuromorphic circuits in these two works enable robots to be more flexible and have greater potential for applications when faced with unknown environments and tasks.

Experience-based learning with long-term plasticity: The conductance (or resistance) of memristor devices can be adjusted continuously. Depending on the working mechanism, some memristors exhibit volatile behavior, with some devices spontaneously returning to a high-resistance state over time. This characteristic of short-term memory retention is comparable to short-term memory in the human brain. By repeatedly stimulating the devices, long-term memory can be achieved^{[86, 87]}. The synaptic characteristics of the memristor are shown in Fig. 2(b) (ii). These characteristics can be continuously adjusted through external control and can transition from short-term to long-term memory. This simulation mimics the cognitive function of human brain’s memory and forgetting, giving robots a brain-like behavioral learning approach. This section will focus on repeated training based on the memristor characteristics, which bring changes in response time to the robot and enable a brain-like application with varying sensitivities.

Similar to how athletes accelerate muscle response times through repetitive training, the long-term synaptic plasticity based on memristors is commonly employed to endow robots with the ability to adapt to changes in response time scales. The neural pathway not only facilitates neuromorphic learning, forgetting, and memory but also plays an important role in processing analog information detected by biological sensory systems. This information traverses intricate neural pathways to the brain, spinal cord, or individual synapses for further processing, and subsequently returns through motor pathways to control movement. Therefore, an artificial neural pathway (ANP)^[68] is proposed. In this work, calcium titanate photoreceptors detect light stimuli and convert them into electrical signals, which are further processed within memristor devices based on zeolitic imidazolate frameworks-8 to handle light signals. Electrical stimulation can precisely modulate memristor conductance values by regulating the growth of nanoparticles within the memristor. The memristor exhibits spontaneous relaxation phenomena in the high-resistance state (HRS) at the initial stage, which can be perceived as the device's learning and forgetting process in response to input stimuli. After repeated electrical stimulation, the device demonstrates excellent retention properties in terms of conductance values, resembling the memory process accumulated from past input stimuli learning experiences. In this way, the ANP utilizes memristor devices to capture the distribution of short-term to long-term memory changes in response to input stimuli, achieving the fundamental neuromorphic functions of learning and memory during autonomous learning processes. By receiving and processing similar light stimuli, the ANP improves the reaction of the robotic arm based on past learning experiences, allowing the arm to gradually approach and grasp a small cube located in front of it. Additionally, ongoing research is focused on the long-term plasticity of memristors in control aspects (Fig. 5(a) (i)). The multifunctional optoelectronic hybrid-integrated neuron based on silver nanoparticle-modified MXene integrates spatiotemporal optoelectronic information. Integrating optical sensing signals and electrical training signals^{[76, 79]} (Fig. 5(a) (ii)) significantly reduces the response time of robotic claw system. A novel multi-level micro-patterned triboelectric nanogenerator (M-TENG), achieves contact/separation delay through the unique spacing between different friction layers, generating multiple pulses in a single touch. Utilizing multiple charge pulses generated by M-TENG to drive organic electrochemical transistor achieves amplification and transmission of multi-synaptic signals, successfully demonstrating a low-energy artificial synaptic device. Long-term plasticity effects are achieved through continuous memory training of the robotic hand^[81].

The neuromorphic circuit based on the memristor, enables robot control to exhibit humanlike response time through experiential learning. It also can simulate a mechanism like sensory adaptability found in biology. The robot can adjust its perceptual sensitivity and behavioral response according to environmental changes, gradually adapting to its surroundings or exhibiting evasive behavior in response to strong harmful stimuli. The resistance of memristors fabricated using single-crystal LiNbO₃^[88] films changes gradually with an increase in the number of voltage pulses. This change in memristor conductivity represents the accumulated sensitivity changes to external stimuli. The magnitude of external voltage pulses determines the length of the conductive filament formed in the LiNbO₃ memristor. Temperature and pressure sensing is achieved through multi-modal environmental sensors and fixed resistors connected in series for voltage division. Threshold judgments on temperature and pressure are made, and when the sensed information surpasses the set threshold, a voltage pulse sufficient to affect the conductivity of the LiNbO₃ memristor filament is generated. This memristor current signal guides the lifting of the robotic arm. Experimental results indicate that as the number of touch stimuli increases, the response to the stimuli (the height of ascent after touching) decreases. In other words, the robotic hand gradually adapts to such stimuli (Fig. 5(b) (i)). For another example, a synaptic transistor using ZnO nanowires^[89], on a flexible substrate is applied in a robotic hand for perceiving external stimuli. This electronic skin enables associative learning based on teacher signals and self-training. The synaptic connections between neurons are influenced by action potentials. Through this learning process, the robotic hand acquires reflex behavior similar to that of humans when encountering harmful stimuli, avoiding further damage (Fig. 5(b) (ii)). A high thermoelectric figure of merit memristor was fabricated using Bi₂Se₃ in an Ag/poly (methyl methacrylate) (PMMA)/Bi₂Se₃/ITO structure (APBI)^[69]. This work utilized the memristor to demonstrate changes in short-term to long-term memory relaxation times. Using the memristor, he robotic claw simulated protective behaviors in response to different temperature-induced injuries. Fig. 5(b) (ii) depicts the artificial thermal nociceptor device exhibiting a human-like conditioned reflex process when exposed to various temperature stimulations. Under normal temperature conditions, the memristor remained in a high-resistance state (HRS), which is deemed as an undamaged state. As the temperature increased, it took a longer time for the device to relax into the HRS state. By exploiting this characteristic, the degree of bending of the robotic claw's index finger mirrored states similar to those observed in living organisms. This incurred a period required for recovery after a low-temperature burn, leaving burn marks, as well as a state requiring substantial time and quality for recovery from severe burns caused by high temperatures. The experiments demonstrated the rapid bending of the robotic claw's finger at high temperatures, mimicking the involuntary reflex protective mechanism observed in human fingers when exposed to high-temperature stimuli.

The Braitenberg vehicle and path-planning robot learned tracking rules and the maze-exit path through environmental feedback using supervised and reinforcement learning. Works like ANP completed bio-inspired training and learning processes by repeatedly learning input signals, accomplishing learning, forgetting, and memory processes for predefined actions. The plasticity of memristors, similar to biological neural synapses, enables robots to autonomously learn and adjust their perceptual abilities. Therefore, researchers have used memristor-based neural circuitry to demonstrate how robots can have adaptive responses to external stimuli or reflexive protective mechanisms against severe external stimuli. These brain-like features enable robots to intelligently explore environmental rules and adapt more effectively to various work environments and task requirements.

The intricate and elaborate neural network distribution within the brain enables it to process an extraordinarily diverse range of information, serving as a paradigm that the field of control presently seeks to emulate. The multi-input information processing platform based on memristor holds the promise of achieving brain-like processing capabilities, which can been demonstrated through the use of device control mechanisms and the structure facets of the devices. This facilitates the processing of memristors to accomplish the information processing of multiple types of sensory and characterized by spatiotemporal features. Additionally, it involves the implementation of parallel signal processing and control based on memristor arrays and device configurations.

Multimodal perception and spatiotemporal information processing: Multimodal information integration in bionics mimics how biological systems use different senses and sources of information simultaneously. Mammals optimally integrate visual and vestibular information to accomplish spatial cognitive tasks^[90], and stimuli under multisensory conditions lead to more accurate and faster behavioral responses compared with single-stimulus conditions^[91]. In robotics, achieving multimodal integration entails combining diverse information from different sensors and sources, aiding robots in comprehensively understanding their environment and performing more complex behaviors^[92−96].

At the same time, memristor with multi-modulation mechanisms and novel memristor structures also confer robots with similar multisensory integration capabilities (Fig. 3(b) (iii)). As shown in Fig. 6(a) (i), the Ag nanoparticle-modified MXene-based multifunctional optoelectronic hybrid integrated neuron (AOHN)^[76] can integrate spatial and temporal optoelectronic information. This solves the problem of single-mode computation and functionality in electrically driven neurons. The device concurrently integrates spatial information from dual-modal inputs. When the device receives two electrical stimuli simultaneously, or one electrical stimulus combined with a 400 nm light stimulus, or one electrical stimulus combined with a 500 nm light stimulus, the synaptic post-currents reach the set threshold at different times. This reflects the variations in the synaptic post-currents due to the integration of different spatial information by the device. Leveraging the ability to integrate spatiotemporal information under these dual-mode device characteristics, an integrated visual perception system mimicking the conditioned response in human has been developed. The artificial synapse device's gate receives presynaptic voltage (V_ps), generating postsynaptic current (I_ps). I_ps is converted into a voltage signal (V_NI) through a transimpedance amplifier circuit. AOHN selectively accepts V_NI stimulation and light stimulation to determine whether the AOHN's received stimuli surpass the firing threshold. Eventually, it converts into a voltage signal to drive the contraction of a mechanical claw.

Furthermore, alongside leveraging the multi-regulation mechanism of memristors, there is a dedicated pursuit to attain multi-sensory fusion functionality through device structure design. A multi-input neuromorphic device was fabricated using nano-particle-doped two-dimensional nanosheet films^[84]. The device's structural characteristics allow two separate gates to receive sensory stimuli independently, emulating multi-dendritic sensory neurons capable of spatiotemporal integration (Fig 6(a) (ii)). The postsynaptic current caused by the gate close to the drain is approximately twice that of the current near the source. These properties enable the device to emulate biological traits that facilitate multi-sensory fusion for different applications. An inertial measurement unit (IMU) sensor collects acceleration (Acc) and gyroscope (Gyro) signals to acquire information about human subjects' movement. These signals describe changes in human body posture and unmanned aerial vehicle flight modes. The two-dimensional signals from the sensor are wirelessly transmitted to a spike encoding circuit. Based on the relationship between amplitude information and instantaneous frequency, the circuit converts the sensor signals into a series of frequency-modulated dual-channel pulse signals allocated to gate1 and gate2 ports within the device, representing different weights. In this study, the acceleration signal is allocated to the gate close to the drain, while the angular velocity information is allocated to the gate near the source. By utilizing the postsynaptic current, this work achieves the recognition of human activity types and unmanned aerial vehicle flight modes, showcasing the device's ability to process multi-dimensional sensory information concurrently.

From previous studies, it is clear that integrating multimodal information often involves integrating spatiotemporal information as well. This characteristic is similar to that of the biological neural system, where neurons with multiple dendritic compartments exhibit impedance gradients along their dendritic positions^{[97, 98]}. This allows for location-related sequence detection. Multimodal information aims to connect different sensory systems or information sources to the same neuron through synaptic connections. This involves information entering the neuron at different dendritic locations and gate-controlled information modulation^[99]. Biological neurons could process multimodal signals and handle spatiotemporal information. Achieving these computations through circuits or software is challenging.

Both methods have achieved the localization of natural spatiotemporal information sources such as sound. In Fig. 6(b) (i), the (spike-timing-dependent plasticity) (STDP) synaptic weight adjustment circuit based on memristor is illustrated^[100]. By setting specific connection weights between left and right ear sensory neurons and their respective two postsynaptic neurons, a correspondence is established between the difference ΔV_int and the direction of sound source's (determined by the order of information transmitted from the left and right ear neurons and the intermediate time gap). This system utilizes the relationship between the spatiotemporal sequence of the system and the postsynaptic current to achieve binaural localization. The spatiotemporal encoding function is also proven through the dendritic integration of a single neuron, significantly enhancing the capabilities of dendritic information processing. In Fig. 6(b) (ii), it is demonstrated that a single multi-terminal oxide-based transistor, coupled with capacitance, can also perform this spatiotemporal information processing task^[70]. By designing distinct electrode distribution structures, a single neural transistor can simulate distinctions in spatiotemporal input sequences across individual dendritic branches. The amplitude of the postsynaptic current triggered by the same presynaptic pulse increases with the gate-to-channel distance. Utilizing this characteristic, a simple artificial neural network that mimics the brain's ability to distinguish sound direction is constructed within a single multi-terminal neural transistor.

Parallel computing: The human brain possesses the ability to process multiple information sources or tasks simultaneously. Parallel computation enables the brain to execute multiple tasks simultaneously, swiftly handling extensive information. This method enables the brain to efficiently accomplish various tasks, enhancing the speed and effectiveness of information processing^[101]. The bottleneck in traditional Von Neumann computer architecture mainly lies in the mismatch on data transfer speed between memories and processors^[78]. Leveraging continuous-time data representation and frequency multiplexing within nanoscale crossbar arrays not only enables simultaneous modulation and processing of analog signals but also extends the scope of large-scale parallel computing solutions.

The parallel processing system constructed by combining memristor arrays and frequency-division multiplexing demonstrates excellent performance in batch image processing. As shown in Fig. 7(a), a 25 × 16 crossbar array on the left is employed to store information from a batch of 16 images (converting a single image from a 5 × 5 grayscale matrix to a 25 × 1 vector matrix for storage). On the right, a 25 × 9 crossbar array stores pre-trained weight matrices for classification inference. The left crossbar matrix inputs 16 different frequency carrier signals to modulate the image information stored within the crossbar matrix. Each row in the matrix outputs different amplitude currents corresponding to the frequencies modulated after processing through a transimpedance amplifier and converting to voltage, serving as input signals for the right crossbar matrix. The input signals, processed through the weight matrix, complete the data processing and classification of the batch of images. In addition to achieving parallel computing using cross-array devices, parallel control functionalities can also be accomplished through device structures, as illustrated in Fig. 7(b). Eight synaptic transistors, based on ion-gel, connect to an ion-gel dielectric to create a synaptic array^[71]. Leveraging ion movement within synaptic transistors dependent on transmission distance, i.e., synaptic signal strength related to the gate position of the connecting synaptic transistor, multiple device signals merge into a unified signal feeding different gates of synaptic transistors, realizing parallel processing and single signal control of multiple drivers. Taking advantages of parallel multiplexing and direct analog control, multiple postsynaptic currents from synaptic transistors can directly drive voltage and current amplification circuits, leading to coordinated finger movements in a robotic hand. This outperforms traditional control systems in grasping objects. There is no doubt that neuromorphic circuits based on memristors facilitate fully analog domain parallel computing^[78].

By building diverse peripheral circuit structures (crossbar arrays) and designing various memristor device structures, data storage, and processing can occur simultaneously, reducing data transfer latency and power consumption. Simultaneously, it possesses the capability for large-scale parallel computing control. This approach could usher in more intricate and efficient information processing capabilities into the field of robotics control.

Most humanoid robot projects only imitate basic human body movements and do not undertake an in-depth analysis of the complexities of the human system^[102]. However, memristor-based neuromorphic circuits in the field of biomimetic control systems have made significant progress. These circuits emulate various biological mechanisms, including the coordination within the human muscular system and the neuro-muscular interactions governing eye movements. They have achieved complete analog-domain emulation of human musculoskeletal dynamics, ocular neuromuscular systems, and intricate feedback control mechanisms. Memristor-based neuromorphic circuits introduce a novel paradigm of integrating the coordination between the brain and cerebellum. They exhibit collaborative control characteristics, effectively addressing the memory wall issues associated with the traditional Von Neumann architectures. We believe that memristor-based neuromorphic biomimetic control systems can replicate the physiological complexity of human body functions. This replication can endow machines with similar adaptability, precision, and control efficiency, leading to transformative innovations in robotics, prosthetics, and autonomous systems.

The neuromuscular junction (NMJ) is a crucial synaptic connection between motor neurons and muscle fibers. Action potentials from the central nervous system are transmitted across the NMJ, triggering muscle movement. Synaptic transistors with CuInP₂S₆ (CIPS) ferroelectric films integrated with AlGaN/GaN high-electron-mobility transistors (HEMTs) have demonstrated artificial synaptic plasticity and the ability to drive significant energy for NMJ^[103]. Ultrasonic distance sensing information is converted into voltages that are applied to the gate of the synaptic transistor. The synaptic power of this transistor allows for the direct control of micro-electro-mechanical system (MEMS) actuators, mimicking the NMJ of the oculomotor nerve to trigger ocular muscle movement and enabling in-situ object tracking (Fig. 8(a) (i)). This system shows promise for biomimetic robotics and neuromorphic vision. Inspired by the voluntary motion mechanisms in the human body, muscle motion soft actuators were created using synaptic devices composed of indium tin oxide/indium gallium zinc oxide/organic−inorganic hybrid dielectric/aluminum structures^[45]. These actuators consist of excitatory and inhibitory synaptic circuits. The excitatory synaptic circuits (source-follower gate circuits) and inhibitory synaptic circuits (diode-loaded common-source circuits) resemble the coordinated actions of antagonistic muscles in muscle movement. The inhibitory circuit converts excitatory signals into inhibitory ones, allowing the antagonistic muscles to relax. The excitatory circuit directly stimulates the corresponding muscle's excitatory neurons, inducing muscle contraction. The V_pre pulse generated followed by the excitatory and inhibitory synaptic circuits is connected to a common electrode providing drain voltage, which drives the soft actuator through current amplification at the same voltage. Precise bidirectional motion control of the soft actuator is achieved by applying positive or negative V_pre pulses (Fig. 8(a) (ii)). In closed-loop control systems, feedback signals from the higher-level cortex of the human brain play a crucial role^[99]. The advanced cortex generates feedback signals through perceptual, cognitive, and decision-making processes, influencing behaviors and motor control by transmitting signals to the basic neural systems. This feedback mechanism allows humans to continuously adjust and adapt behaviors to changing environmental conditions. Similarly, closed-loop control systems use sensors to acquire, process and analyze information, and generate control signals to regulate the state or behavior of the system. The artificial synaptic multiplexer unit (Fig. 8(a) (iii)), which uses ion gel as a dielectric layer, enables multi-gate control of parallel inputs to be combined with feedback signals from actuators and bending angle sensors, eventually realizing closed-loop control systems^[82].

Regarding analog signal processing tasks, repeated analog-to-digital or digital-to-analog conversion processes are no longer necessary since memristors can act as both digital and analog domains. In that case, memristor-based neuromorphic circuits are faster and more efficient than traditional digital circuits. The analog circuit can directly receive and process continuous analog sensor input, while simultaneously allowing advanced algorithms from the digital domain to be integrated into the calculations performed by memristors within the analog circuit. This combination of analog and digital control is similar to the collaborative motion control between the brain’s cerebrum (which processes sensor data and controls motion like analog circuits) and cerebellum (which runs high-level algorithms like digital circuits). At the same time, these hybrid circuits are also capable of efficient parallel processing, can handle multiple inputs, complex computational tasks, and simulate large-scale neural networks.

Recent work has shown that memristor-based mixed-mode control platforms share traits such as fast response times and low power consumption. Fig. 8(b) shows the implementation of an adaptive balancing robot using a memristor-based hybrid analog-digital computing platform. This involved the development of memristor-based hybrid analog-digital sensor fusion algorithms and control algorithms. The goal is to achieve low-latency adaptive balancing for a mobile robot of single degree of freedom^[44] and robust flight control for unmanned aerial vehicles (UAVs)^[104]. The analog circuit section of these works constructed hardware sensor fusion circuits that combined accelerometer and gyroscope sensor data. It also implemented PID motion control hardware circuits^[105], which consist of proportional amplification, differential, and integral circuits. Both the sensor fusion and motion control circuit sections in the analog circuit utilized a memristor as an interface to receive control signals from advanced algorithms in the digital computing platform. This allowed for an analog-digital mixed-control platform without the need of circuits handling digital-to-analog conversion. Experimental results regarding the mobile robot showed that the mixed analog-digital platform, utilizing hardware Kalman filtering fusion and PD motion control circuits, had faster balance recovery and lower latency after external force impact by contrast with a purely digital control platform. In UAV control, Kalman filtering fusion and complementary filtering algorithms based on hardware circuit^[106] were used to accurately estimate the drone's rotational angles. The memristor-based PID analog-digital hybrid computing control platform also demonstrated excellent characteristics in the UAV domain. Compared with purely digital computing platforms, using sensor fusion based on memristors and PID hardware control algorithms consumes as less as 142.5 mJ power. This energy-efficient computing platform could reduce the burden on both batteries and computational capabilities in UAVs, leading to more robust flight control.

In this review, we have discussed significant memristive mechanisms used in constructing neuromorphic circuits, focusing on the storage and computing capabilities of the memristor. We have also summarized representative works in neuromorphic control, including cognitive learning, complex signal processing, and biomimetic driving achieved through combining peripheral circuits or designing different structures of memristors. The research in memristor-based neuromorphic circuits holds tremendous potential in biologically inspired intelligence and robot control. With a deeper understanding of the nervous system, memristor-based neuromorphic circuits will better simulate the characteristics of the human brain, promoting their applications in fields such as intelligent robots, artificial intelligence, and neural computing.

On one hand, leveraging the parallel computing capabilities and energy efficiency of memristors can realize the construction of more heavily loaded and efficient hardware neural networks, achieving deeper and more intelligent information processing. On the other hand, integrating memristor-based neuromorphic circuits with new architectures can emulate biological neurons and synapses. These circuits are incorporated features such as cognitive learning, multimodal information processing, spatiotemporal information processing, parallel computing, and biomimetic driving, thus enabling the construction of bio-inspired robots that mimic the entire human nervous system. This would offer more possibilities for autonomous perception, complex information processing, intelligent decision-making, and precise control in robots. However, integrating memristor-based neuromorphic computing into the sensing, information processing, and control aspects of robots may pose challenges. Addressing these challenges could be facilitated by constructing novel structured integrated memristors designed specifically for sensory-memory computing. Additionally, the issue of device uniformity needs to be addressed to ensure the consistency of performance across different units. At the same time, when applying memristors in the field of robot control, the detection of weak signals in devices is a topic worth special attention. The challenges mentioned above have been noticed by researchers and are being actively overcome. We anticipate that more advancements in memristor-based neuromorphic circuit research will be inspired in the near future, enabling robots to possess more human-like processing and control capabilities, thus opening up new possibilities for humankind.