A photonic neuromorphic autonomous navigation method and system

Abstract

The application discloses a kind of photonic neural morphology autonomous navigation method and system, the method obtains environment point cloud data by laser radar, constructs environment state vector after sector division and feature extraction;Dynamic timing inference is carried out using pulse neural network, and its core calculation is realized through optical linear calculation module and optical nonlinear activation module: the former is completed by MZI interferometer linear operation of high speed, low power consumption, and the latter generates nonlinear pulse response using DFB-SA laser array;Finally, the output pulse signal is converted into robot control instruction.The application combines pulse neural network with photonic computing, fully gives the high parallelism and low power consumption advantage of photonic computing, and the high energy efficiency characteristics of pulse neural network, significantly improves the real-time performance, energy efficiency ratio and environmental adaptability of navigation system, and provides an efficient autonomous navigation solution for mobile robot.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence and autonomous systems technology, and in particular to a photonic neuromorphic autonomous navigation method and system. Background Technology

[0002] With the rapid development of autonomous driving, intelligent robots, drones, and edge intelligent devices, autonomous navigation systems have become one of the core technologies for realizing intelligent mobile robots. Currently, most autonomous navigation systems rely on electronic computing platforms to run traditional artificial neural networks (ANNs) or convolutional neural networks (CNNs) to process environmental sensor data, such as LiDAR, visual images, and IMUs, to complete tasks such as path planning, obstacle avoidance, and target tracking. However, these methods have significant limitations in terms of real-time performance, energy consumption, and system complexity.

[0003] High computational resource consumption: Traditional neural networks rely on high-performance GPUs or embedded CPUs for inference, which consumes a lot of power and has high latency, making it difficult to meet the needs of resource-constrained platforms such as micro-robots and drones.

[0004] Non-brain-like computing mode: Electronic neural networks use synchronous computing, which is fundamentally different from the asynchronous pulse processing mechanism of biological nervous systems, limiting their potential in low-power sensing and fast response.

[0005] Poor scalability and physical bottlenecks: Moore's Law is slowing down, and electronic devices face physical limitations such as bandwidth, energy consumption and interconnection latency, making it difficult to further improve computing density and energy efficiency.

[0006] In recent years, neuromorphic computing and photonic computing have emerged as cutting-edge directions for overcoming the aforementioned bottlenecks. Photonic neural networks, utilizing the physical characteristics of parallel propagation of optical signals and ultra-high bandwidth, can achieve energy-efficient execution of large-scale neural computations, exhibiting unique advantages, especially in navigation scenarios where deployment space is limited or real-time requirements are extremely high.

[0007] Currently, autonomous navigation systems are mostly based on deep reinforcement learning algorithms (such as DQN, DDPG, TD3, etc.) and perception fusion technology. They process LiDAR, visual images, or map information by constructing neural network models to achieve environmental modeling and path planning. These systems are mostly deployed on general-purpose electronic computing platforms, such as CPUs / GPUs, and rely on high-dimensional dense neural networks to complete end-to-end control strategy learning.

[0008] Some studies have attempted to use spiking neural networks (SNNs) for path planning and obstacle avoidance tasks [3, 4], leveraging the sparse pulses and asynchronous processing characteristics of SNNs to reduce power consumption. However, these implementations are still limited to running on digital processors or FPGAs and have not fundamentally broken the limitations of traditional electronic devices in terms of power consumption, bandwidth, and parallelism.

[0009] Meanwhile, photonic computing technology has made groundbreaking progress in recent years. Studies have shown that using Mach-Zehnder Interferometer (MZI) arrays to achieve tunable weighted matrix operations, combined with DFB-SA lasers to generate nonlinear modulation signals, provides the physical basis for realizing photonic neural networks. Yang Z's team constructed an optical neural network based on MZI and, combined with the deep Q-network framework from deep learning, proposed an optical deep Q-network algorithm. This algorithm achieved performance comparable to electronic neural networks in 2D path planning tasks, providing a new method for the efficient application of MZI photonic computing in reinforcement learning. However, existing research mostly focuses on static tasks such as image recognition and signal classification, lacking exploration of applications for decision control in dynamic scenarios, such as autonomous navigation, and has not yet formed an integrable system-level solution. Summary of the Invention

[0010] This invention provides a photonic neuromorphic autonomous navigation method and system, which solves the problem of lack of decision control in dynamic scenarios in the prior art, and realizes the real-time performance, energy efficiency and environmental adaptability of the navigation system.

[0011] In a first aspect, the present invention provides a photonic neuromorphic autonomous navigation method, the method comprising:

[0012] Point cloud data of the robot in the navigation environment is acquired by LiDAR, the point cloud data is divided into sectors and the minimum obstacle distance of each sector is extracted, and the environmental state vector is constructed by combining the robot pose information and target point information.

[0013] The environmental state vector is input into a pre-trained spiking neural network to obtain continuous action signals; wherein, the spiking neural network is based on the LIF neuron model and uses membrane potential accumulation and pulse firing mechanism to dynamically infer the state in the time dimension to obtain continuous action signals;

[0014] The weights of the spiking neural network are mapped to an explicit MZI interferometer, and linear calculations are performed through optical interference to obtain the optical interference result. The optical interference result is then converted into a photoelectric signal to obtain a linear electrical signal.

[0015] The linear electrical signal is injected into the DFB-SA laser array. By adjusting the current in the gain region and the reverse bias voltage in the saturation region, the DFB-SA array generates a nonlinear pulse response to the linear electrical signal, thus generating a pulse output signal.

[0016] Based on the pulse output signal, linear velocity and angular velocity control commands for the robot are generated, driving the robot to perform obstacle avoidance or navigation tasks.

[0017] In conjunction with the first aspect, in one possible implementation, the step of acquiring point cloud data of the robot in the navigation environment via LiDAR, dividing the point cloud data into sectors and extracting the minimum obstacle distance for each sector, and constructing an environmental state vector by combining the robot's pose information and target point information, includes:

[0018] The 360° point cloud data acquired by the lidar is divided into N sectors with equal angular ranges on the horizontal plane, and the minimum distance between all measurement points in each sector and the robot is extracted to form an obstacle distance vector.

[0019] Obtain the robot's current pose, including its global coordinate position and heading angle, and obtain the global coordinate position of the target point;

[0020] Based on the robot's current pose and the target point's position, calculate the Euclidean distance between the robot and the target point, as well as the orientation angle between the robot's current heading and the target point's direction;

[0021] The obstacle distance vector, the Euclidean distance, the orientation angle, and the robot's current linear velocity and angular velocity are combined to form the environmental state vector.

[0022] In conjunction with the first aspect, in one possible implementation, the spiking neural network includes: an input layer, at least one hidden layer, and an output layer; wherein each layer includes a linear transformation unit and a LIF neuron unit;

[0023] The linear transformation unit is used to perform weighted summation on the input signal of the current layer to obtain the input current of the neuron in the current layer;

[0024] The LIF neuron unit uses the input current as its driving signal, integrates the input in a timely manner through the accumulation and leakage of internal membrane potential, and triggers pulse output when the membrane potential exceeds the firing threshold; wherein, the update equation of the membrane potential is:

[0025] ;

[0026] in, Indicates the leakage coefficient; This is represented as the issuance threshold; This represents the pulse firing function, which fires a pulse and resets the membrane potential when the membrane potential exceeds a threshold.

[0027] In conjunction with the first aspect, in one possible implementation, the training method for the spiking neural network includes:

[0028] Initialize the Actor network, Critic network, and their respective target networks;

[0029] The robot is controlled to perform actions in a simulation environment and collect experience data; wherein, the experience data includes: current state, performed action, reward, next state, and round termination flag;

[0030] S3: Store the experience data in the experience replay pool;

[0031] S4: Randomly sample a batch of empirical data from the empirical replay pool, and update the parameters of the Critic network based on the temporal difference objective and loss function of the TD3 algorithm;

[0032] S5: Update the parameters of the Actor network at a predetermined update frequency by maximizing the Q value of one of the outputs of the Critic network;

[0033] S6: Using a soft update method, slowly update the parameters of the Actor network and the Critic network to their respective target networks;

[0034] S7: Repeat steps S2 to S6 until the Actor network converges;

[0035] The spiking neural network is built based on the LIF neuron model, and its spiking mechanism is backpropagated during training using a differentiable approximation function.

[0036] In conjunction with the first aspect, in one possible implementation, the loss function of the spiking neural network is expressed as:

[0037] The loss function of the Critic network is the mean squared error loss, expressed as:

[0038] ;

[0039] The loss function of the Actor network is a negative Q-value expectation, expressed as:

[0040] ;

[0041] in, Indicates batch size; The set of trainable parameters is The first Critic network pairs the states With action The estimated Q value was performed; The set of trainable parameters is The second Critic network for state With action The estimated Q value and the target Q value The error between; Indicates at time The state vector obtained from the environment; Indicates the state The following are the continuous actions output by the Critic network; This represents the target Q value in the TD3 algorithm; This represents the set of trainable parameters for the first Critic network; This represents the set of trainable parameters for the second Critic network.

[0042] In conjunction with the first aspect, in one possible implementation, the environmental state vector is input into a pre-trained spiking neural network to obtain continuous action signals; wherein the spiking neural network is based on a LIF neuron model, and through membrane potential accumulation and pulse firing mechanisms, it dynamically infers the state in the time dimension to obtain continuous action signals, including:

[0043] The environmental state vector corresponding to the current moment is used as the input of the spiking neural network;

[0044] In each layer of the spiking neural network, the following temporal dynamic inference process is performed:

[0045] A linear transformation is performed on the input signal or the pulse signal from the previous layer to obtain the input current of the neurons in the current layer;

[0046] Based on the LIF neuron model, the leakage integral mechanism is used to accumulate the result of the linear transformation with the membrane potential at the previous time step to obtain the membrane potential at the current time step.

[0047] The current membrane potential is compared with the firing threshold: if the membrane potential exceeds the threshold, the neuron fires a pulse and resets the membrane potential; if the threshold is not exceeded, the neuron does not fire a pulse and maintains the current membrane potential until the next time step.

[0048] The binary pulse signal indicating whether a pulse is being emitted at the current moment is used as the output of the neuron in this layer and transmitted to the next layer or the output layer.

[0049] The pulse sequence output from the last layer of the network is passed through a linear layer and mapped via the Tanh activation function to generate a continuous action signal located within a continuous interval.

[0050] In conjunction with the first aspect, in one possible implementation, mapping the weights of the spiking neural network to an explicit MZI interferometer, performing linear calculations through optical interference to obtain an optical interference result, and then performing photoelectric conversion on the optical interference result to obtain a linear electrical signal, includes:

[0051] Based on the weight matrix of the pulse neural network, the phase shifter phase of each MZI unit in the explicit MZI interferometer array is configured, and the weight matrix is ​​optically fixed in the MZI array.

[0052] The environmental state vector is modulated onto a multi-channel light source to generate an optical signal carrying the environmental state vector;

[0053] The optical signal is injected into an MZI array with a solidified weight matrix, and linear calculations are performed through optical interference to obtain the interference optical signal;

[0054] The interference optical signal is converted into a linear electrical signal by photoelectric conversion.

[0055] In conjunction with the first aspect, in one possible implementation, injecting the linear output into the DFB-SA laser array, and adjusting the gain region current and the saturation region reverse bias voltage to cause the DFB-SA array to generate a nonlinear pulse response to the input signal, thereby generating a pulse output signal, includes:

[0056] The linear electrical signal is injected into each corresponding DFB-SA laser in the DFB-SA laser array; wherein each DFB-SA laser corresponds to a neuron;

[0057] By adjusting the gain region current and saturation region reverse bias voltage of each DFB-SA laser, the corresponding neuron-like excitation threshold and dynamic response characteristics are set.

[0058] When the linear electrical signal exceeds a set threshold, the DFB-SA laser generates pulsed light output to achieve nonlinear activation, and the pulsed light output is used as the pulse output signal of the pulse neural network.

[0059] In conjunction with the first aspect, in one possible implementation, generating linear and angular velocity control commands for the robot based on the pulse output signal to drive the robot to perform obstacle avoidance or navigation tasks includes:

[0060] The pulse output signal is processed through a linear transformation layer to obtain a continuous motion vector;

[0061] The motion vector is scaled and mapped, with the first component mapped to linear velocity and the second component mapped to angular velocity; wherein, the linear velocity is mapped to the interval [0,1] and the angular velocity is mapped to the interval [-1,1].

[0062] Control commands are generated based on the mapped linear and angular velocities, and the robot is driven to move by the robot actuators.

[0063] Secondly, the present invention provides a photonic neuromorphic autonomous navigation system, comprising:

[0064] The sensor module is used to acquire point cloud data of the robot in the navigation environment through LiDAR, divide the point cloud data into sectors and extract the minimum obstacle distance of each sector, and construct an environmental state vector by combining the robot pose information and target point information.

[0065] The decision processing module is used to input the environmental state vector into a pre-trained spiking neural network to obtain continuous action signals; wherein, the spiking neural network is based on the LIF neuron model and uses membrane potential accumulation and pulse firing mechanism to dynamically reason about the state in the time dimension to obtain continuous action signals.

[0066] An optical linear calculation module is used to map the weights of the pulse neural network to an explicit MZI interferometer, perform linear calculations through optical interference, obtain optical interference results, and perform photoelectric conversion on the optical interference results to obtain a linear electrical signal;

[0067] An optical nonlinear activation module is used to inject the linear electrical signal into the DFB-SA laser array. By adjusting the current in the gain region and the reverse bias voltage in the saturation region, the DFB-SA array generates a nonlinear pulse response to the linear electrical signal, thus generating a pulse output signal.

[0068] The control module is used to generate linear and angular velocity control commands for the robot from the pulse output signal and drive the robot to perform obstacle avoidance or navigation tasks.

[0069] One or more technical solutions provided in this invention have at least the following technical effects or advantages:

[0070] This invention acquires point cloud data using LiDAR, divides it into sectors, extracts the minimum obstacle distance, and constructs an environmental state vector by combining robot pose and target point information. This step compresses high-dimensional, sparse point cloud data into a low-dimensional, structured environmental state vector, significantly reducing the amount of data required for subsequent processing and improving the system's real-time performance. Through sector division, the system can perceive the distribution of obstacles in different directions. Combined with pose and target information, the state vector contains local obstacle avoidance and global path information required for navigation, providing a comprehensive and efficient input foundation for intelligent decision-making. The environmental state vector is input into a pre-trained spiking neural network (SNN), and continuous action signals are obtained based on a LIF neuron model. Utilizing the event-driven characteristics of the SNN and the bio-simulation dynamics of the LIF neuron, this step achieves high energy efficiency and low-latency dynamic temporal information processing. The SNN performs inference in the time dimension, better handling continuous sensor data streams, and its sparse pulse computation characteristics significantly reduce power consumption. Compared to traditional artificial neural networks, it is more suitable for resource-constrained embedded navigation systems. The weights of the spiking neural network are mapped to an explicit MZI interferometer, and linear calculations and photoelectric conversions are performed through optical interference. This step leverages the high parallelism, ultra-high speed, and low power consumption advantages of photonic computing to physically accelerate the linear transformations in the spiking neural network. Optical interference calculations are completed instantaneously in the optical domain, breaking through the speed and power consumption bottlenecks of traditional electronic computing. It is particularly suitable for handling large-scale matrix operations, laying the core computational foundation for the rapid response of the entire system. A linear electrical signal is injected into a DFB-SA laser array, and parameters are adjusted to generate a nonlinear pulse response. This step cleverly utilizes the nonlinear characteristics of the DFB-SA laser to simulate the nonlinear activation and pulse firing functions in biological neurons. It converts high-speed photoelectric signals back into optical pulses, completing the key nonlinear operation in the computational loop while maintaining the advantages of signal processing in the optical domain. This method achieves high-speed, low-power "optical-electrical-optical" signal conversion and processing, a crucial step in neuromorphic optical computing. Linear and angular velocity control commands for the robot are generated based on the pulse output signals. This step translates the complex neuromorphic calculation results from the upper layers into low-level control commands that can directly drive the robot's actuators. It establishes a closed loop from intelligent perception and decision-making to specific physical actions, enabling robots to perform obstacle avoidance and navigation tasks in real time and autonomously. The entire system forms a complete, efficient, and low-power autonomous navigation solution from perception and computation to control. Attached Figure Description

[0071] Figure 1 A flowchart illustrating the steps of a photonic neuromorphic autonomous navigation method provided in this embodiment of the invention;

[0072] Figure 2 A schematic diagram of a spiking neural network structure provided in an embodiment of the present invention;

[0073] Figure 3 This is a schematic diagram of the simulation environment provided in the embodiments of the present invention;

[0074] Figure 4 This is a schematic diagram of the overall training framework for Spiking-TD3 provided in an embodiment of the present invention;

[0075] Figure 5 The graph showing the growth of reward and success rate with the number of training rounds is provided for embodiments of the present invention.

[0076] Figure 6 This is an explicit MZI network structure diagram provided in an embodiment of the present invention;

[0077] Figure 7 This is a flowchart illustrating an embodiment of the present invention. Detailed Implementation

[0078] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0079] In a first aspect, the present invention provides a photonic neuromorphic autonomous navigation method, see [link to previous document]. Figure 1 The method includes the following steps S101 to S105.

[0080] S101: The robot acquires point cloud data in the navigation environment through LiDAR, divides the point cloud data into sectors and extracts the minimum obstacle distance in each sector, and constructs an environmental state vector by combining the robot pose information and target point information.

[0081] Specifically, in step S101, point cloud data of the robot in the navigation environment is acquired by LiDAR, the point cloud data is divided into sectors and the minimum obstacle distance of each sector is extracted, and the environmental state vector is constructed by combining the robot pose information and target point information, including the following steps S1011 to S1014.

[0082] S1011 divides the 360° point cloud data acquired by the LiDAR into N sectors with equal angular ranges on the horizontal plane, and extracts the minimum distance between all measurement points in each sector and the robot to form an obstacle distance vector.

[0083] S1012, Obtain the robot's current pose, including its global coordinate position and heading angle, and obtain the global coordinate position of the target point;

[0084] S1013, based on the robot's current pose and the target point's position, calculate the Euclidean distance between the robot and the target point, as well as the orientation angle between the robot's current heading and the target point's direction;

[0085] S1014 combines the obstacle distance vector, Euclidean distance, orientation angle, and the robot's current linear and angular velocities to form the environment state vector.

[0086] For example, see Figure 3 For the simulation environment of this invention, radar point cloud data is obtained through the ROS topic / velodyne_points. The code performs dimensionality reduction processing on the point cloud:

[0087] The 360° laser scan results are divided into n sectors according to the angle, such as 20 sectors, and the minimum distance value of each sector is taken as the state input to describe the obstacle distribution.

[0088] The robot's pose is read in real time via the / p3dx / odom topic, and its position coordinates are extracted. and orientation angle It is used to calculate the Euclidean distance and orientation deviation from the target point.

[0089] Ultimately, the following is achieved through the env.step(action) interface: robot action execution → environment state update → reward calculation → new state return.

[0090] The state acquisition process is defined in the file velodyne_env.py. After each action is performed by the robot, the following information needs to be updated from the sensors and internal odometry:

[0091] Velodyne LiDAR data: The 360° point cloud is divided into a specified number of sectors (controlled by the environment_dim parameter, such as 20 sectors). Each sector takes the distance to the nearest obstacle within that angle range. If the minimum distance is too small, it indicates that the robot may have already approached an obstacle.

[0092] Robot's own state: including the current Euclidean distance from the target point, the angle between the robot's orientation and the target's orientation, and the linear and angular velocities of the last execution.

[0093] Therefore, the environment state vector can be represented as:

[0094] ;

[0095] in, For the first The distance values ​​for each sector are: dist2goal, which is the distance between the robot and the target point, and angle2goal, which is the angle between the robot's current heading and the target point's direction. and These represent linear velocity and angular velocity, respectively.

[0096] S102, The environmental state vector is input into a pre-trained spiking neural network to obtain continuous action signals; wherein, the spiking neural network is based on the LIF neuron model, and through the membrane potential accumulation and pulse firing mechanism, it performs dynamic reasoning on the state in the time dimension to obtain continuous action signals.

[0097] Here, the spiking neural network includes: an input layer, at least one hidden layer, and an output layer; wherein each layer includes a linear transformation unit and a LIF neuron unit;

[0098] The linear transformation unit is used to perform weighted summation on the input signals of the current layer to obtain the input current of the neurons in the current layer;

[0099] The LIF neuron unit uses the input current as the driving signal, integrates the input through the accumulation and leakage of internal membrane potential, and triggers pulse output when the membrane potential exceeds the firing threshold; the update equation for the membrane potential is:

[0100] ;

[0101] in, Indicates the leakage coefficient; This is represented as the issuance threshold; This represents the pulse firing function, which fires a pulse and resets the membrane potential when the membrane potential exceeds a threshold.

[0102] Here, the loss function of the spiking neural network is expressed as:

[0103] The loss function of the Critic network is the mean squared error loss, expressed as:

[0104] ;

[0105] The loss function of the Actor network is a negative Q-value expectation, expressed as:

[0106] ;

[0107] in, Indicates batch size; The set of trainable parameters is The first Critic network pairs the states With action The estimated Q value was performed; The set of trainable parameters is The second Critic network for state With action The estimated Q value and the target Q value The error between; Indicates at time The state vector obtained from the environment includes the distance of the LiDAR sector, the distance between the robot and the target point, the angle, and the speed. Indicates the state The continuous motion output by the Critic network includes linear velocity and angular velocity control values; This represents the target Q value in the TD3 algorithm; This represents the set of trainable parameters for the first Critic network; This represents the set of trainable parameters for the second Critic network.

[0108] Specifically, training methods for spiking neural networks include:

[0109] S1: Initialize the Actor network, Critic network, and their respective target networks;

[0110] S2: Control the robot to perform actions in a simulation environment and collect experience data; the experience data includes: current state, performed actions, reward, next state, and round termination flag;

[0111] S3: Store experience data in the experience replay pool;

[0112] S4: Randomly sample a batch of empirical data from the empirical replay pool, and update the parameters of the Critic network based on the temporal difference objective and loss function of the TD3 algorithm;

[0113] S5: Update the parameters of the Actor network at a predetermined update frequency by maximizing the Q value of one of the outputs of the Critic network;

[0114] S6: Use a soft update method to slowly update the parameters of the Actor network and Critic network to their respective target networks;

[0115] S7: Repeat steps S2 to S6 until the Actor network converges;

[0116] Spiking neural networks are built on the LIF neuron model, and their spiking mechanism is backpropagated during training using a differentiable approximation function.

[0117] Specifically, in step S102, the environmental state vector is input into a pre-trained spiking neural network to obtain continuous action signals; wherein, the spiking neural network is based on the LIF neuron model, and through membrane potential accumulation and pulse firing mechanisms, it dynamically infers the state in the time dimension to obtain continuous action signals, including:

[0118] (1) Use the environmental state vector corresponding to the current moment as the input of the spiking neural network;

[0119] (2) In each layer of the spiking neural network, the following time-series dynamic inference process is performed:

[0120] (2.1) Perform a linear transformation on the input signal or the pulse signal from the previous layer to obtain the input current of the neurons in the current layer;

[0121] (2.2) Based on the LIF neuron model, the leakage integral mechanism is used to accumulate the result after linear transformation with the membrane potential at the previous time step to obtain the membrane potential at the current time step.

[0122] (2.3) Compare the current membrane potential with the firing threshold: if the membrane potential exceeds the threshold, the neuron fires a pulse and resets the membrane potential; if the threshold is not exceeded, the neuron does not fire a pulse and maintains the current membrane potential until the next moment.

[0123] (2.4) The binary pulse signal indicating whether a pulse is being emitted at the current moment is used as the output of the neuron in this layer and transmitted to the next layer or the output layer;

[0124] (2.5) The pulse sequence output from the last layer of the network is passed through a linear layer and mapped by the Tanh activation function to generate a continuous action signal located in a continuous interval.

[0125] For example, this invention employs a multilayer spiking neural network as the Actor network structure. SNNs possess membrane potential integration and leakage characteristics, enabling them to capture the dynamic evolution of state sequences over time, thus exhibiting stronger representation capabilities for time-dependent tasks.

[0126] In its implementation, the core of SNN consists of alternating custom LIFSpike layers and linear layers (nn.Linear). Each neuron in each layer accumulates the input signal through membrane potential; if the membrane potential exceeds a threshold... This triggers a pulse output and resets the membrane potential. A differentiable approximation function of spike_fn is used to enable backpropagation training. Through a pulse firing rate and time averaging mechanism, sequential pulses are converted into continuous outputs, thus adapting to the continuous motion control requirements of the TD3 algorithm.

[0127] In the reinforcement learning phase, see Figure 4This invention employs the TD3 algorithm to train the SNN network. Based on DDPG (Deep Deterministic Policy Gradient), the TD3 algorithm introduces strategies such as dual Criticism and delayed Actor updates, which can effectively suppress overestimation and reduce the noise impact of policy updates while training the continuous controller, significantly improving the stability and convergence speed of the continuous control task.

[0128] 2.2.2 Construction and Formula Description of Spiking Neural Networks:

[0129] The spiking neural network (SNN) used in this invention is represented by the LIF (Leaky Integrate-and-Fire) model, which combines the leakage accumulation characteristics of membrane potential and the spike triggering mechanism. The specific implementation includes the following key steps.

[0130] For discrete time step Let the membrane potential of a certain neuron be denoted as . The input current (or input signal) is denoted as The membrane potential leakage coefficient is denoted as (0< <1), the threshold of the neuron is denoted as Then, at each time step, the membrane potential update of this neuron can be expressed as:

[0131] ;

[0132] in:

[0133] ;

[0134] When membrane potential When the threshold is exceeded, the neuron releases a pulse and resets the membrane potential to [value missing]. (or approximately 0), thus achieving the so-called "discharge" process. In the above formula, The smaller the leakage, the more obvious it is; the larger the leakage, the higher the degree of preservation of historical membrane potential.

[0135] Since the impulse function is inherently non-differentiable, this invention employs an approximate gradient technique called ZIF (Zero-Inflated Firing) or other alternatives to achieve a continuous approximation of the impulse firing function, enabling backpropagation to occur within the network. Specifically, an approximate activation function can be used. Alternative to hard jump:

[0136] ;

[0137] And by introducing the following in backpropagation: The gradient of the membrane potential can be calculated using methods such as or other smoothing functions.

[0138] To achieve temporal feature extraction and continuous action control, the overall structure of the spiking neural network (SNN Actor) designed in this invention adopts a temporally unfolded multilayer perceptron structure, and performs a three-stage operation of "linear weighting - membrane potential integration - threshold firing" at each time step.

[0139] The SNN Actor network structure is as follows: Figure 2 As shown, the designed SNN Actor network includes an input layer, two hidden layers (one and two layers), and an output layer.

[0140] Assume the input state is The computation process of each layer of the network can then be described as follows: See Figure 2 The input to the input layer is represented as: .

[0141] SNN Actor Network Layer 1:

[0142] Linear transformation unit calculation: ;

[0143] Linear transformation unit calculation: ; ;

[0144] SNN Actor Network Second Layer:

[0145] Linear transformation unit calculation: ;

[0146] Linear transformation unit calculation: ; ;

[0147] Output layer: Continuous action signals ;

[0148] in, and The layer weights and bias parameters, This indicates that the neurons in this layer are at time [time]. Whether to output a binary pulse, and the input of the input layer. The result is obtained by performing a linear transformation in the first layer. After nonlinear activation of LIF neurons, pulses are generated and sent to the second layer. As the input to the second hidden layer, it undergoes the same processing to produce the input to the output layer. After the final linear transformation and tanh constraint processing, the final output action is obtained. The dynamic and temporal superposition mechanism of membrane potential endows the network with temporal memory, enabling it to maintain stable decision-making performance in environments with noise or perception delays.

[0149] 2.2.3 TD3 Reinforcement Learning Algorithm Flow and Key Formulas:

[0150] While the spiking neural network (Actor) is used to output actions, this invention trains the Actor network and the dual Critic network using the TD3 algorithm. TD3 has the following key improvements over the traditional DDPG:

[0151] Dual Critic Networks: TD3 maintains two independent Critic networks, denoted as follows: and For the next state-action pair TD3 will calculate the two Critic outputs separately and take the minimum value:

[0152] ;

[0153] This approach can effectively reduce the problem of Q-value overestimation and improve training stability.

[0154] Target Q-value calculation and delayed update: In TD3, both the Actor and the Critic have their own corresponding target networks, denoted as: And Actor .

[0155] The parameters of the target network will be incremented by a small step. Perform a soft update (Polyak update), that is:

[0156] ;

[0157] Critic estimation from the target network output The target for the Q value at the next moment:

[0158] ;

[0159] in, Adding a certain amount of Gaussian or noise and then cropping it is called TargetPolicySmoothing.

[0160] Based on this, the Bellman approximation target can be obtained:

[0161] ;

[0162] in, It's an instant reward. It is a discount factor. This indicates whether the termination has been reached; when done=1, it means that updates will cease in the next moment.

[0163] Critic Loss Function: This invention uses mean squared error (MSE) to measure the loss during Critic training. With the goal The difference. For a dual-critic, the loss function can be expressed as:

[0164] ;

[0165] in, The batch size (batch_size). Based on stochastic gradient descent or the Adam optimizer... , Update.

[0166] Actor Delayed Update: To reduce frequent interference from Actor updates, the TD3 policy often sets a delay parameter `policy_freq`, updating the Actor only after a certain number of Critic updates. The goal of the Actor update is to maximize the Critic's Q-value for its output action, i.e., minimize:

[0167] ;

[0168] Only one of the Critics is used here (usually) That's it. Then use a soft update to synchronize the Actor parameters to the target Actor network.

[0169] 2.2.4 Training Process and Data Flow:

[0170] (1) Environment Interaction and Storage: This invention collects the current state by interacting with the simulation environment (GazeboEnv). ,action ,award and the next state This information, including the four-tuples or five-tuples and a sign indicating whether the input is complete, is stored in the replay buffer. The replay buffer's capacity is set to... A first-in-first-out (FIFO) queue mechanism is used to ensure that the latest data is retained first.

[0171] (2) Batch sampling and network update: After accumulating a certain amount of interactive data, a batch is randomly sampled from the Replay Buffer, with batch size=B, and the Critic and Actor are updated respectively according to the update process required by TD3.

[0172] Critic Update: Calculate loss based on batch samples ,right Perform gradient descent.

[0173] Actor Update (Delayed Trigger): Updates Actor parameters using Critic estimation. After each update, soft update to .

[0174] Random noise and exploration: In the early stages of training, in order to increase the exploration of the strategy, a certain amount of Gaussian noise is added to the actions output by the Actor, and the noise gradually decays as training progresses.

[0175] (3) Target point arrival and collision detection: In the actual interaction process, if the robot is close to the target point, such as less than the threshold of 0.3m, it is considered to have successfully arrived and can be given a large positive reward; if the lidar detects that the nearest obstacle is too close, below the safe distance threshold of 0.35m, it is considered a collision and a negative reward is given, and the current round ends. The sparse reward and punishment mechanism set up in this way can effectively guide the network to learn obstacle avoidance actions.

[0176] (4) Network Output Actions: The output of the SNN Actor network is typically mapped to the interval [-1, 1], and then transformed into linear velocity ∈ [0, 1] and angular velocity ∈ [0, 1]. [1,1], thereby driving the robot to move forward and turn.

[0177] Iterative training follows the above process; see [link / reference]. Figure 5 As the number of interactions with the environment increases, the parameters of the network are continuously improved, enabling the robot to learn more robust obstacle avoidance strategies in increasingly complex scenarios and ultimately reach its goal.

[0178] S103 maps the weights of the pulse neural network to an explicit MZI interferometer, performs linear calculations through optical interference, obtains the optical interference results, and performs photoelectric conversion on the optical interference results to obtain a linear electrical signal;

[0179] Specifically, in step S103, the weights of the spiking neural network are mapped to an explicit MZI interferometer, linear calculations are performed through optical interference to obtain optical interference results, and the optical interference results are converted into photoelectric signals to obtain linear electrical signals, including the following steps S1031 to S1034.

[0180] S1031, Based on the weight matrix of the pulse neural network, configure the phase shifter phase of each MZI unit in the explicit MZI interferometer array, and optically solidify the weight matrix onto the MZI array.

[0181] S1032 modulates the environmental state vector onto a multi-channel light source to generate an optical signal carrying the environmental state vector.

[0182] S1033, the optical signal is injected into the MZI array with the solidified weight matrix, and the linear calculation is completed through optical interference to obtain the interference optical signal;

[0183] S1034 performs photoelectric conversion on the interference optical signal to obtain a linear electrical signal.

[0184] For example, based on the aforementioned training process, the weight values ​​of each network layer are obtained. By adjusting the phase of each phase shifter in the explicit MZI network, the mapping between the weights and the hardware is achieved, completing the hardware deployment of the weights and realizing optical linear computation. See also... Figure 6 It is an explicit MZI network.

[0185] In the designed explicit MZI network, each independent MZI represents a weight, calculated as follows:

[0186] ;

[0187] in, It is the phase shift of the internal phase shifter of the MZI. It is the weight value represented by the current MZI, which can be adjusted. The size of the MZI is used to explicitly configure the weight matrix. .

[0188] A multi-channel light source with wavelengths spaced 0.5 nm apart was used as the input signal source. A pulse signal was applied to the light source using an optical modulator (MZM) to obtain an optical signal carrying the environmental state vector. .

[0189] See Figure 7 The light, loaded with a pulse signal, is injected into the input port of the explicit MZI network through an optical fiber. After entering the explicit MZI network, the optical signal undergoes interference within each MZI unit, photoelectric conversion by the BPD, current summation, and current-to-voltage transformation by the transimpedance amplifier, thus completing the process. The calculation yields a linear electrical signal at the output port. .

[0190] S104 injects a linear electrical signal into the DFB-SA laser array. By adjusting the current in the gain region and the reverse bias voltage in the saturation region, the DFB-SA array generates a nonlinear pulse response to the linear electrical signal, thus generating a pulse output signal.

[0191] Specifically, in step S104, the linear output is injected into the DFB-SA laser array. By adjusting the current in the gain region and the reverse bias voltage in the saturation region, the DFB-SA array generates a nonlinear pulse response to the input signal, thus generating a pulse output signal. This includes the following steps S1041 to S1044.

[0192] S1041, inject a linear electrical signal into each corresponding DFB-SA laser in the DFB-SA laser array; wherein, each DFB-SA laser corresponds to a neuron;

[0193] S1042, by adjusting the gain region current and saturation region reverse bias voltage of each DFB-SA laser, sets the corresponding neuron-like excitation threshold and dynamic response characteristics;

[0194] S1043 When the linear electrical signal exceeds the set threshold, the DFB-SA laser generates pulsed light output to achieve nonlinear activation, and uses the pulsed light output as the pulse output signal of the pulse neural network.

[0195] For example, the optical signal obtained after calculation by the explicit MZI network is converted into a current signal by a balanced photodetector (BPD). The current signals are connected in series by wires to achieve summation and then enter the TIA transimpedance amplifier to become a voltage signal, which includes the weighted summation result. This voltage signal is then used to drive the DFB-SA array, so that each DFB-SA processes the linear weighted summation result of one neuron.

[0196] By adjusting the current in the gain region and the reverse bias voltage in the saturation region of the DFB-SA laser, a neuron-like state is set up to make the DFB-SA laser produce a neuron-like response to the input signal, thus obtaining a pulsed output signal. .

[0197] S105 generates linear and angular velocity control commands for the robot based on the pulse output signal, driving the robot to perform obstacle avoidance or navigation tasks.

[0198] Specifically, in step S105, linear velocity and angular velocity control commands for the robot are generated based on the pulse output signal to drive the robot to perform obstacle avoidance or navigation tasks, including the following steps S1051 to S1053.

[0199] S1051 processes the pulse output signal through a linear transformation layer to obtain a continuous motion vector;

[0200] S1052, scale and map the motion vector, mapping the first component to linear velocity and the second component to angular velocity; wherein, the linear velocity is mapped to the interval [0,1] and the angular velocity is mapped to the interval [-1,1].

[0201] S1053 generates control commands based on the mapped linear velocity and angular velocity, and drives the robot to move through the robot actuator.

[0202] This invention introduces a spiking neural network into an autonomous navigation SNN Actor network. Its core feature is the use of mechanisms such as pulse firing and membrane potential accumulation to express and process temporal information more naturally. Compared to traditional deep networks based on static activation functions, it significantly improves the network's sensitivity to temporal features when processing dynamic sensing data (such as LiDAR sequence point clouds), thereby enhancing environmental perception and target navigation performance in dynamic scenes.

[0203] Low energy consumption and bio-inspired design. Due to the event-driven nature of spiking neurons, they only fire pulses when a threshold is exceeded, resulting in a more sparse overall computational overhead. In the future, running on neuromorphic hardware such as Loihi and TrueNorth will further demonstrate these advantages of low energy consumption and high real-time performance, providing a more bio-inspired and engineering-valuable solution for robot navigation in resource-constrained scenarios.

[0204] The weight configuration is simple and the response speed is fast. This patent adopts a more explicit and easily configurable MZI photonic network structure to replace the triangular or rectangular network structure used in traditional photonic computing, thereby simplifying the configuration process of the weight matrix in the photonic neural network.

[0205] This invention can be extended to actual robot platforms. It uses LiDAR data as the primary sensing source, combined with an event-triggered spiking neural network for control decisions, exhibiting strong portability. Whether in environments with more sensor fusion, such as vision and IMU, or in more complex outdoor or heterogeneous environments, the algorithm framework can still be extended to run on actual robots, and hardware acceleration can be used to meet real-time requirements, providing advanced intelligent mobility solutions for various service robots, logistics robots, inspection robots, and more.

[0206] Secondly, the present invention provides a photonic neuromorphic autonomous navigation system, comprising:

[0207] The sensor module is used to acquire point cloud data of the robot in the navigation environment through LiDAR, divide the point cloud data into sectors and extract the minimum obstacle distance of each sector, and construct an environmental state vector by combining the robot pose information and target point information.

[0208] The decision processing module is used to input the environmental state vector into a pre-trained spiking neural network to obtain continuous action signals. The spiking neural network is based on the LIF neuron model and uses membrane potential accumulation and pulse firing mechanism to dynamically reason about the state in the time dimension to obtain continuous action signals.

[0209] The optical linear calculation module is used to map the weights of the pulse neural network to an explicit MZI interferometer, perform linear calculations through optical interference, obtain optical interference results, and perform photoelectric conversion on the optical interference results to obtain a linear electrical signal;

[0210] The optical nonlinear activation module is used to inject a linear electrical signal into the DFB-SA laser array. By adjusting the current in the gain region and the reverse bias voltage in the saturation region, the DFB-SA array generates a nonlinear pulse response to the linear electrical signal, thus generating a pulse output signal.

[0211] The control module is used to generate linear and angular velocity control commands for the robot from the pulse output signals, and drive the robot to perform obstacle avoidance or navigation tasks.

[0212] For example, suppose a robot moves from starting point A to target point B in an office environment containing obstacles such as tables and chairs. After the system starts, the LiDAR in the sensor module begins scanning the surrounding environment to acquire real-time point cloud data. The LiDAR divides the 360-degree field of view into 20 sectors, each 18 degrees, and calculates the minimum obstacle distance in the point cloud data of each sector. For example, the minimum distance for sector 1 is 1.5 meters, and for sector 2 it is 0.8 meters, etc. Simultaneously, the sensor module combines the robot's current pose, such as position coordinates x=0, y=0, orientation angle θ=0, and target point pose (x=10, y=5), to construct an environmental state vector. This vector includes the minimum obstacle distance for each sector, the robot's direction and distance relative to the target point.

[0213] The decision processing module inputs the environmental state vector into a pre-trained spiking neural network. This network, based on a LIF neuron model, performs dynamic reasoning at time steps through membrane potential accumulation and pulse firing mechanisms. For example, it processes a state update at each time step, outputting continuous action signals such as forward speed and turning rate. The optical linear computation module maps the weights of the spiking neural network onto an explicit MZI interferometer array, performing linear operations such as matrix multiplication through optical interference to generate optical interference results. These results are then converted into linear electrical signals, such as voltage signals, by a photodetector. The optical nonlinear activation module injects this electrical signal into a DFB-SA laser array. By adjusting the gain region current to 50mA and the saturation region reverse bias voltage to 2V, the laser generates a nonlinear pulse response to the input signal, outputting a pulsed output signal, such as a series of optical pulses. The control module receives this pulsed output signal, decodes it into the robot's linear velocity (e.g., 0.5 m / s) and angular velocity (e.g., 0.3 radians / s), and provides control commands to drive the robot to perform obstacle avoidance actions, such as navigating around tables and chairs, and gradually navigating towards the target point B. Throughout the process, the system operates in real time to ensure that the robot reaches its destination safely and efficiently.

[0214] The various embodiments described in this specification are presented in a progressive manner. Similar or identical parts between embodiments can be referred to interchangeably. Each embodiment focuses on its differences from other embodiments. All or part of this invention can be used in numerous general-purpose or special-purpose computer system environments or configurations. Examples include: personal computers, server computers, handheld or portable devices, tablet devices, mobile communication terminals, multiprocessor systems, microprocessor-based systems, programmable electronic devices, network PCs, minicomputers, mainframe computers, and distributed computing environments including any of the above systems or devices, etc.

[0215] The above embodiments are only used to illustrate the technical solutions of the present invention, and are not intended to limit the present invention. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features therein. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the present invention.

Claims

1. A photonic neuromorphic autonomous navigation method, characterized in that, include: Point cloud data of the robot in the navigation environment is acquired by LiDAR, the point cloud data is divided into sectors and the minimum obstacle distance of each sector is extracted, and the environmental state vector is constructed by combining the robot pose information and target point information. The environmental state vector is input into a pre-trained spiking neural network to obtain continuous action signals. The spiking neural network is based on a LIF neuron model and uses membrane potential accumulation and pulse firing mechanisms to dynamically infer the state in the time dimension to obtain continuous action signals. The spiking neural network includes an input layer, at least one hidden layer, and an output layer. Each layer includes a linear transformation unit and a LIF neuron unit. The linear transformation unit is used to perform weighted summation on the input signal of the current layer to obtain the input current of the neuron in the current layer; The LIF neuron unit uses the input current as its driving signal, integrates the input in a timely manner through the accumulation and leakage of internal membrane potential, and triggers pulse output when the membrane potential exceeds the firing threshold; wherein, the update equation of the membrane potential is: ； in, Indicates the leakage coefficient; This is represented as the issuance threshold; This represents the pulse firing function, which fires a pulse and resets the membrane potential when the membrane potential exceeds a threshold. The weights of the spiking neural network are mapped to an explicit MZI interferometer, and linear calculations are performed through optical interference to obtain the optical interference result. The optical interference result is then converted into a photoelectric signal to obtain a linear electrical signal. The linear electrical signal is injected into the DFB-SA laser array. By adjusting the current in the gain region and the reverse bias voltage in the saturation region, the DFB-SA array generates a nonlinear pulse response to the linear electrical signal, thus generating a pulse output signal. Based on the pulse output signal, linear velocity and angular velocity control commands for the robot are generated, driving the robot to perform obstacle avoidance or navigation tasks.

2. The photonic neuromorphic autonomous navigation method according to claim 1, characterized in that, The process involves acquiring point cloud data of the robot in the navigation environment using LiDAR, dividing the point cloud data into sectors and extracting the minimum obstacle distance for each sector, and combining the robot's pose information with the target point information to construct an environmental state vector, including: The 360° point cloud data acquired by the lidar is divided into N sectors with equal angular ranges on the horizontal plane, and the minimum distance between all measurement points in each sector and the robot is extracted to form an obstacle distance vector. Obtain the robot's current pose, including its global coordinate position and heading angle, and obtain the global coordinate position of the target point; Based on the robot's current pose and the target point's position, calculate the Euclidean distance between the robot and the target point, as well as the orientation angle between the robot's current heading and the target point's direction; The obstacle distance vector, the Euclidean distance, the orientation angle, and the robot's current linear velocity and angular velocity are combined to form the environmental state vector.

3. The photonic neuromorphic autonomous navigation method according to claim 1, characterized in that, The training method for the spiking neural network includes: Initialize the Actor network, Critic network, and their respective target networks; The robot is controlled to perform actions in a simulation environment and collect experience data; wherein, the experience data includes: current state, performed action, reward, next state, and round termination flag; S3: Store the experience data in the experience replay pool; S4: Randomly sample a batch of empirical data from the empirical replay pool, and update the parameters of the Critic network based on the temporal difference objective and loss function of the TD3 algorithm; S5: Update the parameters of the Actor network at a predetermined update frequency by maximizing the Q value of one of the outputs of the Critic network; S6: Using a soft update method, slowly update the parameters of the Actor network and the Critic network to their respective target networks; S7: Repeat steps S2 to S6 until the Actor network converges; The spiking neural network is built based on the LIF neuron model, and its spiking mechanism is backpropagated during training using a differentiable approximation function.

4. The photonic neuromorphic autonomous navigation method according to claim 3, characterized in that, The loss function of the Critic network is the mean squared error loss, expressed as: ； The loss function of the Actor network is a negative Q-value expectation, expressed as: ； in, Indicates batch size; The set of trainable parameters is The first Critic network pairs the states With action The estimated Q value was performed; The set of trainable parameters is The second Critic network for state With action The estimated Q value and the target Q value The error between; Indicates at time The state vector obtained from the environment; Indicates the state The following are the continuous actions output by the Critic network; This represents the target Q value in the TD3 algorithm; This represents the set of trainable parameters for the first Critic network; This represents the set of trainable parameters for the second Critic network.

5. The photonic neuromorphic autonomous navigation method according to claim 1, characterized in that, The environmental state vector is input into a pre-trained spiking neural network to obtain continuous action signals; wherein, the spiking neural network is based on a LIF neuron model, and through membrane potential accumulation and pulse firing mechanisms, it dynamically infers the state in the time dimension to obtain continuous action signals, including: The environmental state vector corresponding to the current moment is used as the input of the spiking neural network; In each layer of the spiking neural network, the following temporal dynamic inference process is performed: A linear transformation is performed on the input signal or the pulse signal from the previous layer to obtain the input current of the neurons in the current layer; Based on the LIF neuron model, the leakage integral mechanism is used to accumulate the result of the linear transformation with the membrane potential at the previous time step to obtain the membrane potential at the current time step. The current membrane potential is compared with the firing threshold: if the membrane potential exceeds the threshold, the neuron fires a pulse and resets the membrane potential; if the threshold is not exceeded, the neuron does not fire a pulse and maintains the current membrane potential until the next time step. The binary pulse signal indicating whether a pulse is being emitted at the current moment is used as the output of the neuron in this layer and transmitted to the next layer or the output layer. The pulse sequence output from the last layer of the network is passed through a linear layer and mapped via the Tanh activation function to generate a continuous action signal located within a continuous interval.

6. The photonic neuromorphic autonomous navigation method according to claim 1, characterized in that, The process of mapping the weights of the pulse neural network to an explicit MZI interferometer, performing linear calculations through optical interference to obtain optical interference results, and then performing photoelectric conversion on the optical interference results to obtain a linear electrical signal includes: Based on the weight matrix of the pulse neural network, the phase shifter phase of each MZI unit in the explicit MZI interferometer array is configured, and the weight matrix is ​​optically fixed in the MZI array. The environmental state vector is modulated onto a multi-channel light source to generate an optical signal carrying the environmental state vector; The optical signal is injected into an MZI array with a solidified weight matrix, and linear calculations are performed through optical interference to obtain the interference optical signal; The interference optical signal is converted into a linear electrical signal by photoelectric conversion.

7. The photonic neuromorphic autonomous navigation method according to claim 1, characterized in that, The process of injecting the linear output into the DFB-SA laser array and generating a pulse output signal by adjusting the gain region current and the saturation region reverse bias voltage, includes: The linear electrical signal is injected into each corresponding DFB-SA laser in the DFB-SA laser array; wherein each DFB-SA laser corresponds to a neuron; By adjusting the gain region current and saturation region reverse bias voltage of each DFB-SA laser, the corresponding neuron-like excitation threshold and dynamic response characteristics are set. When the linear electrical signal exceeds a set threshold, the DFB-SA laser generates pulsed light output to achieve nonlinear activation, and the pulsed light output is used as the pulse output signal of the pulse neural network.

8. The photonic neuromorphic autonomous navigation method according to claim 1, characterized in that, The step of generating linear and angular velocity control commands for the robot based on the pulse output signal, and driving the robot to perform obstacle avoidance or navigation tasks, includes: The pulse output signal is processed through a linear transformation layer to obtain a continuous motion vector; The motion vector is scaled and mapped, with the first component mapped to linear velocity and the second component mapped to angular velocity; wherein, the linear velocity is mapped to the interval [0,1] and the angular velocity is mapped to the interval [-1,1]. Control commands are generated based on the mapped linear and angular velocities, and the robot is driven to move by the robot actuators.

9. A photonic neuromorphic autonomous navigation system, characterized in that, include: The sensor module is used to acquire point cloud data of the robot in the navigation environment through LiDAR, divide the point cloud data into sectors and extract the minimum obstacle distance of each sector, and construct an environmental state vector by combining the robot pose information and target point information. A decision processing module is used to input the environmental state vector into a pre-trained spiking neural network to obtain continuous action signals. The spiking neural network is based on a LIF neuron model and uses membrane potential accumulation and pulse firing mechanisms to dynamically infer the state over time to obtain continuous action signals. The spiking neural network includes an input layer, at least one hidden layer, and an output layer. Each layer includes a linear transformation unit and a LIF neuron unit. The linear transformation unit is used to perform weighted summation on the input signal of the current layer to obtain the input current of the neuron in the current layer; The LIF neuron unit uses the input current as its driving signal, integrates the input in a timely manner through the accumulation and leakage of internal membrane potential, and triggers pulse output when the membrane potential exceeds the firing threshold; wherein, the update equation of the membrane potential is: ； in, Indicates the leakage coefficient; This is represented as the issuance threshold; This represents the pulse firing function, which fires a pulse and resets the membrane potential when the membrane potential exceeds a threshold. An optical linear calculation module is used to map the weights of the pulse neural network to an explicit MZI interferometer, perform linear calculations through optical interference, obtain optical interference results, and perform photoelectric conversion on the optical interference results to obtain a linear electrical signal; An optical nonlinear activation module is used to inject the linear electrical signal into the DFB-SA laser array. By adjusting the current in the gain region and the reverse bias voltage in the saturation region, the DFB-SA array generates a nonlinear pulse response to the linear electrical signal, thus generating a pulse output signal. The control module is used to generate linear and angular velocity control commands for the robot from the pulse output signal and drive the robot to perform obstacle avoidance or navigation tasks.

Images