Efficient and adaptive perception in unstructured environments is crucial for robotics and autonomous driving. Traditional methods struggle with the variability inherent in real-world scenarios. This research aims to address this limitation by developing a memristor-based neuromorphic computing system that mimics the differential processing of sensory information observed in biological systems. Humans efficiently understand unstructured environments through differential processing, where multiple receptors and neurons process sensory information differently and adapt their structure and synaptic weights based on stimulus features. Memristors, with their synapse-like characteristics, are ideally suited to emulate these biological functions. Existing methods often assign a single memristor to a fixed receptor, limiting the utilization of sensory data. This research proposes a differential neuromorphic computing method using memristors' multistate properties to extract features from unstructured data and modulate memristor states, enhancing the adaptability of robotic systems.

The paper reviews existing research in biomimetic sensory feedback, electronic skins for soft robots, learned robot manipulation, embodied neuromorphic intelligence, and neuromorphic computing using Loihi and other neuromorphic processors. It highlights the limitations of current memristor-based approaches, such as the restriction to fixed receptor processing, which omits useful information and fails to emulate the full spectrum of biological sensory processing. The literature also underscores the importance of multi-sensory integration and the role of various receptors and neurons in achieving human-like perception in unstructured environments. The use of memristors in neuromorphic computing is discussed, emphasizing their synapse-like characteristics and potential to replicate synaptic plasticity.

The proposed differential neuromorphic computing method uses memristors' intrinsic multistate properties to extract features from unstructured data and modulate memristor states. The method was tested in two application scenarios: object grasping and autonomous driving. **Tactile Perception:** A piezoresistive force sensing architecture and a self-directed channel (SDC) memristor were used to emulate nociceptors (amplifying hazardous stimuli) and adapting receptors (regulating mild stimuli). An FPGA platform was used for feature extraction and modulation scheme selection, allowing the memristor to operate at high, middle, and low resistance levels for nociception, adaptation, and recovery, respectively. Experiments involved grasping a 3D-printed object (cube and cone) and a soap to demonstrate safe grasping of sharp and slippery objects. The system achieved >720% amplification of hazardous stimuli and >50% attenuation of mild stimuli, mimicking biological tactile perception. A time window processing mechanism was implemented to distinguish sudden from persistent threats. **Visual Perception:** A driving recorder and a 40x25 memristor array were used to process differentially encoded visual motion information. The gray CMOS image was compressed, and filtering circuits extracted changes in visual information between frames. The system classified visual stimuli as fast or slow based on a preset threshold, encoding them into electrical pulses to modulate the memristor states. Slow information released the memristor from a low-resistance state. Experiments involved designed driving scenarios (pedestrian crossing) and free-driving scenarios (various traffic conditions). The system achieved 94% accuracy in extracting critical information for decision-making, comparable to human assessment. The system demonstrated successful detection of pedestrians, vehicles, road signs, and other crucial elements.

The key findings demonstrate the effectiveness of the proposed memristor-based differential neuromorphic computing method in enabling robots to perceive and adapt to unstructured environments. **Tactile Perception:** The system successfully mimicked biological nociception, adaptation, and recovery functions, achieving >720% amplification of hazardous stimuli and >50% attenuation of mild stimuli. It enabled a robot hand to safely grasp both sharp and slippery objects by dynamically adjusting grasping strategies based on tactile feedback. The adaptation speed could also be adjusted. **Visual Perception:** The system achieved 94% accuracy in extracting critical information for autonomous driving decision-making in diverse scenarios and various weather and lighting conditions, demonstrating its robustness and generalization capability. The system effectively distinguishes between fast and slow-moving objects, generating "afterimages" to indicate movement orientation and provide more actionable data for higher-level processing. The speed and accuracy of visual information processing were comparable to human drivers' perceptions.

The findings address the research question by demonstrating that memristor-based differential neuromorphic computing can enable robots to adapt and operate in unstructured environments with improved accuracy and efficiency compared to traditional methods. The successful emulation of biological perception mechanisms highlights the potential of this approach. The high accuracy in object grasping and autonomous driving scenarios showcases the system's real-world applicability. The ability to dynamically adjust the response to stimuli, based on both current and historical information, shows its advantage over static approaches like PID control. The fast response time of the system suggests potential for real-time applications.

This research presents a novel memristor-based differential neuromorphic computing method that allows robots to perceive and adapt to complex unstructured environments. This bio-inspired approach demonstrated high accuracy and efficiency in both tactile and visual perception tasks, showcasing its significant potential for advancing robotics and autonomous driving. Future research should focus on developing more automatic modulation schemes and control algorithms, designing scalable parallel circuits, and addressing the challenges of device variability and nonlinear dynamics to enhance the robustness and reliability of the system for real-world applications.

The current implementation relies on pre-set modulation schemes and hard threshold comparisons, lacking the flexibility for diverse real-world environments. Scalable parallel circuit design for controlling a large number of memristors is a significant challenge. Device variability and nonlinear memristive dynamics also need to be addressed for real-world applications. The visual system's performance might be affected under low-light conditions or with a narrower dynamic range compared to more advanced camera systems.