Robot path planning method, device, and electronic equipment based on pulse membrane potential readout

By using a path planning method based on pulse membrane potential readout, the problems of high computational load and high energy consumption in robot path planning are solved, achieving efficient path planning in dynamic obstacle environments and improving obstacle avoidance safety and path generation stability.

CN123113852APending Publication Date: 2026-11-17ZHONGSHAN INST OF CHANGCHUN UNIV OF SCI & TECH
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202611220774.4
Authority / Receiving Office
CN · China
Patent Type
Applications(China)
Current Assignee / Owner
Filing Date
2026-08-11
Publication Date
2026-11-17

AI Technical Summary

Technical Problem

Existing deep learning-based robot path planning schemes are computationally intensive and energy-intensive in dynamic obstacle environments, requiring high computational performance from the robot itself. Furthermore, the action value representation of spiking neural path planning schemes is unstable and difficult to use directly for continuous action evaluation.

Method used

By acquiring the robot's current path planning state data, path decision context data is generated and state units are divided. State correlation is calculated and pulse coding weights and triggering sequences are determined. Membrane potential accumulation and updating are performed using a pulse neural computation layer. Action value correlation is calculated in combination with non-pulse readout neurons. Candidate movement actions are evaluated, and the path decision context is updated through path feedback data.

Benefits of technology

It reduces the computational load and energy consumption of path planning, and improves the real-time performance of path planning, obstacle avoidance safety, and path generation stability in dynamic environments.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN123113852A_ABST
    Figure CN123113852A_ABST
Patent Text Reader

Abstract

This application discloses a robot path planning method, apparatus, and electronic device based on pulsed membrane potential readout. The method acquires the current path planning state data of the target robot in a target working environment, generates path decision context data, and divides it into multiple path state units. It calculates the state correlation between each path state unit and preset path decision factors, determines the pulse coding weight and pulse triggering sequence, and then generates a pulse state sequence. The pulse state sequence is input into a pulsed neural computation layer, and within a preset simulation time window, the corresponding pulsed neurons are updated using event-triggered membrane potential accumulation based on pulse events, and the pulse output state is determined. Based on the membrane potential state, pulse output state, and path decision context data, action readout input data is determined. Candidate movement actions are evaluated using membrane potential change data from non-pulse readout neurons, the target movement action is determined, and path nodes and path feedback evaluation data are updated, iteratively generating the target planned path. This reduces the computational load and energy consumption of path planning, and improves the real-time performance, obstacle avoidance safety, and stability of path planning in dynamic environments.
Need to check novelty before this filing date? Find Prior Art

Description

TECHNICAL FIELD

[0001] The present application relates to the technical field of robots, and in particular to a robot path planning method and device based on pulse membrane potential reading and electronic equipment. BACKGROUND

[0002] Robot path planning is one of the key technologies in mobile robot navigation control, and its main goal is to determine a feasible path that can guide the robot to safely reach the target position in a given work environment according to the current position of the robot, the target position and the distribution of environmental obstacles. In industrial inspection, warehouse logistics, medical services, agricultural operations and complex indoor and outdoor autonomous mobile scenarios, the path planning result directly affects the safety of robot obstacle avoidance, motion efficiency and task execution stability. With the increasing complexity of the work environment, robots often need to handle static obstacles, dynamic obstacles and multi-sensor detection information at the same time, which puts higher requirements on the real-time performance and resource consumption of the path planning method.

[0003] Existing robot path planning methods include graph search, artificial potential field, grid decomposition, swarm intelligence optimization and deep learning-based path decision methods. Traditional path planning methods have certain applicability in regular or known environments, but in dynamic obstacle-dense, frequently changing local information or high-dimensional sensor data scenarios, a large amount of state search, cost calculation or path recalculation is often required, resulting in increased computational overhead. Deep learning-based path planning methods can use high-dimensional environmental state data for path decision-making and have strong expression ability in complex scenarios, but they usually rely on continuous numerical activation and intensive calculation processes, which have high requirements on processor computing power, storage access bandwidth and energy supply.

[0004] For mobile robots, path planning calculations usually need to be completed in real time on the robot body side, and the computing resources and battery capacity on the robot body side are limited. Existing deep neural network-based path planning schemes usually need to perform continuous multiplication and accumulation operations on a large number of neurons or feature channels during inference, even if some environmental state changes are small, they may still participate in full-quantity calculation. The above calculation method will bring high energy consumption and large calculation delay, and in dynamic obstacle real-time avoidance, low-power cruising and edge autonomous navigation applications, it is easy to reduce the robot's endurance and path response speed.

[0005] Spiking neural computation utilizes discrete pulse events to transmit information, featuring event triggering, sparse activity, and low-power computation, making it suitable for resource-constrained robot path planning scenarios. However, some existing spiking neural path planning schemes still suffer from problems such as unstable action value representation, insufficient integration of path decision data with obstacle risk feedback, and difficulty in directly using pulse computation results for continuous action evaluation. Some schemes also require reliance on traditional continuous networks to assist in value estimation or policy conversion, leading to complex computational chains and hindering the full realization of the energy-saving advantages of pulse event-triggered computation.

[0006] Therefore, there is an urgent need for a low-energy path planning scheme for robot path planning scenarios, which can convert the robot's environmental state into sparse pulse events, evaluate candidate movement actions based on membrane potential accumulation and non-pulse membrane potential readout, and dynamically update the selection of subsequent path nodes by combining path feedback information. This reduces the computational load and energy consumption in the path planning process while ensuring obstacle avoidance safety and path continuity. Summary of the Invention

[0007] This application provides a robot path planning method, device, and electronic device based on pulsed membrane potential readout, which solves the problems of high computational load, high energy consumption, and high requirements for the computational performance of the robot body in existing deep learning-based robot path planning schemes in dynamic obstacle environments.

[0008] This application provides a robot path planning method based on pulsed membrane potential readout, comprising: acquiring current path planning state data of a target robot in a target working environment, wherein the current path planning state data is used to characterize the spatial and motion relationships between the target robot, target location objects, environmental obstacle objects, and sensor detection objects; generating path decision context data based on the current path planning state data, and dividing the current path planning state data into state units according to the path decision dimension to obtain multiple path state units, wherein the path decision context data is used to record the correspondence between current path nodes, historical path nodes, candidate movement actions, environmental obstacle change states, and path feedback evaluation data. The system calculates the state correlation between each path state unit and the preset path decision factors, and determines the pulse coding weight and pulse triggering sequence corresponding to each path state unit based on the state correlation. Based on the pulse coding weight and pulse triggering sequence, pulse timing coding is performed on each path state unit to obtain a pulse state sequence corresponding to the current path planning state data. The pulse state sequence is input into the pulse neural computation layer. Within a preset simulation time window, the membrane potential of the corresponding spiking neurons in the pulse neural computation layer is cumulatively updated according to the pulse events in the pulse state sequence. The pulse output state of the spiking neurons is determined based on the updated membrane potential state. Specifically, the membrane potential cumulative update is the update of the corresponding pulse output state when a pulse event is input. The process involves event-triggered state updates performed by neurons; determining action readout input data corresponding to multiple candidate movement actions based on the membrane potential state, pulse output state, and path decision context data of spiking neurons; inputting each action readout input data into the non-spiking readout neurons corresponding to the multiple candidate movement actions, and calculating the membrane potential change data of each non-spiking readout neuron within a preset simulation time window; calculating the action value correlation between each candidate movement action and preset path decision factors based on the membrane potential change data of each non-spiking readout neuron, and constraining and correcting the action value correlation according to the path decision context data to obtain the path action evaluation result corresponding to each candidate movement action; and further refining the path action evaluation result based on the path decision context data. The system selects the path action evaluation result corresponding to the selected movement action, determines the target movement action from multiple candidate movement actions, and updates the path node data of the target robot based on the target movement action. Based on the updated path node data and the environmental state change data in the target working environment, it generates path feedback evaluation data corresponding to the updated path nodes and updates the path decision context data based on the path feedback evaluation data. If the target robot has not reached the target position and has not met the preset termination condition, it iteratively determines the next path node based on the updated path decision context data and the updated current path planning state data, so as to generate the target planning path based on multiple path nodes determined from the current position to the target position.

[0009] According to an embodiment of this application, based on updated path decision context data and updated current path planning state data, the next path node is iteratively determined to generate a target planned path based on multiple path nodes determined from the current position to the target position. This includes: iteratively determining multiple path nodes based on updated path decision context data and updated current path planning state data, and generating an initial planned path according to the determined order of each path node; performing trajectory continuity processing on adjacent path nodes in the initial planned path to obtain continuous candidate trajectories; performing collision verification on the continuous candidate trajectories based on the spatial occupancy data and motion change data of environmental obstacles in the target work environment; and determining the continuous candidate trajectories as the target planned path if they meet the collision verification conditions.

[0010] According to the embodiments of this application, the current path planning status data includes the robot's current position data, target position data, static obstacle position data, dynamic obstacle position data, obstacle movement direction data, multi-directional sensor detection data, relative geometric data between the robot and the target position object, and relative risk data between the robot and environmental obstacle objects.

[0011] According to an embodiment of this application, the path state unit includes a target approach state unit, an obstacle proximity state unit, a dynamic obstacle motion state unit, a sensor direction detection state unit, and a safety angle state unit.

[0012] According to the embodiments of this application, the preset path decision factors include target approach factors, obstacle avoidance factors, motion direction continuity factors, and dynamic collision risk factors.

[0013] According to the embodiments of this application, the state correlation between each path state unit and the preset path decision factors is calculated, and the pulse coding weight and pulse triggering sequence corresponding to each path state unit are determined based on the state correlation. This includes: calculating the initial state correlation between each path state unit and the preset path decision factors based on the spatial position change data, motion trend change data, and sensor detection intensity data corresponding to each path state unit; performing context correction on the initial state correlation based on the historical path nodes and path feedback evaluation data in the path decision context data to obtain the corrected state correlation; and determining the pulse coding weight and pulse triggering sequence of each path state unit in the pulse timing coding process based on the corrected state correlation.

[0014] According to an embodiment of this application, pulse timing encoding is performed on each path state unit based on pulse coding weights and pulse triggering timing to obtain a pulse state sequence corresponding to the current path planning state data. This includes: determining the pulse triggering intensity of each path state unit within a preset simulation time window based on the pulse coding weights corresponding to each path state unit; determining the pulse triggering time step of each path state unit within a preset simulation time window based on the pulse triggering timing corresponding to each path state unit; and generating a pulse state sequence formed by arranging multiple pulse events in chronological order based on the pulse triggering intensity and the pulse triggering time step.

[0015] According to an embodiment of this application, within a preset simulation time window, the membrane potential of the corresponding spiking neuron in the spiking neural computation layer is cumulatively updated based on the pulse events in the pulse state sequence. This includes: at the current time step of the preset simulation time window, detecting whether there is a pulse event input to the target spiking neuron in the pulse state sequence; if there is a pulse event input to the target spiking neuron, updating the membrane potential of the target spiking neuron based on the pulse weight corresponding to the pulse event and the membrane potential state of the target spiking neuron in the previous time step; if there is no pulse event input to the target spiking neuron, maintaining the membrane potential state of the target spiking neuron in the previous time step, or updating the state of the target spiking neuron based on a preset membrane potential decay rule.

[0016] According to an embodiment of this application, determining the pulse output state of a spiking neuron based on the updated membrane potential state includes: comparing the updated membrane potential state with a preset membrane potential threshold; generating the pulse output state corresponding to the spiking neuron and performing a reset process on the membrane potential state of the spiking neuron if the updated membrane potential state meets the pulse triggering condition; and prohibiting the spiking neuron from outputting pulse events if the updated membrane potential state does not meet the pulse triggering condition.

[0017] According to an embodiment of this application, based on the membrane potential state, pulse output state, and path decision context data of a spiking neuron, action readout input data corresponding to multiple candidate movement actions is determined, including: determining multiple candidate movement actions and action constraints corresponding to each candidate movement action based on the path decision context data; calculating the action response correlation between the membrane potential state and pulse output state of the spiking neuron and each candidate movement action; and determining the action readout input data corresponding to each candidate movement action from the membrane potential state and pulse output state of the spiking neuron based on the action response correlation and action constraints.

[0018] According to an embodiment of this application, the correlation between action value and preset path decision factors is calculated based on the membrane potential change data of each non-pulse readout neuron, including: obtaining the membrane potential values ​​of each non-pulse readout neuron at multiple time steps within a preset simulation time window; determining the action value representation data corresponding to each candidate movement based on the maximum membrane potential of each non-pulse readout neuron within the preset simulation time window; and calculating the correlation between action value and preset path decision factors based on the action value representation data corresponding to each candidate movement.

[0019] According to an embodiment of this application, the path action evaluation result corresponding to each candidate movement action is obtained by constraining and correcting the action value correlation based on path decision context data. This includes: determining the target approach constraint, obstacle avoidance constraint, movement direction continuity constraint, and dynamic collision risk constraint corresponding to each candidate movement action based on path decision context data; weighting and correcting the action value correlation corresponding to each candidate movement action based on the target approach constraint, obstacle avoidance constraint, movement direction continuity constraint, and dynamic collision risk constraint; and obtaining the path action evaluation result corresponding to each candidate movement action based on the weighted and corrected action value correlation.

[0020] According to an embodiment of this application, based on the path action evaluation results corresponding to each candidate movement action, a target movement action is determined from multiple candidate movement actions, and the path node data of the target robot is updated according to the target movement action. This includes: sorting multiple candidate movement actions according to the path action evaluation results corresponding to each candidate movement action; determining the candidate movement action whose sorting result meets the preset action selection conditions as the target movement action; determining the next path node after the target robot moves from the current path node based on the movement direction and movement step size corresponding to the target movement action; and writing the next path node into the path node sequence in the path decision context data to update the path node data of the target robot.

[0021] According to an embodiment of this application, path feedback evaluation data corresponding to the updated path nodes is generated based on the updated path node data and environmental state change data in the target working environment. This includes: calculating the distance change data of the target robot relative to the target location object based on the path node data before and after the update; calculating the obstacle avoidance risk change data of the target robot relative to the environmental obstacle object based on the updated path node data and the position change data of the environmental obstacle object; calculating the motion direction continuity evaluation data of the target robot based on the updated path node data, historical path node data, and the target location object; and generating path feedback evaluation data based on the distance change data, obstacle avoidance risk change data, and motion direction continuity evaluation data.

[0022] According to an embodiment of this application, updating path decision context data based on path feedback evaluation data includes: determining the path benefit evaluation result and path risk evaluation result corresponding to the current target movement action based on the path feedback evaluation data; updating the historical evaluation relationship between each candidate movement action and preset path decision factors in the path decision context data based on the path benefit evaluation result and path risk evaluation result; and adjusting the state correlation calculation result and action value correlation correction result in the next path node determination process based on the updated historical evaluation relationship.

[0023] According to an embodiment of this application, trajectory continuity processing is performed on adjacent path nodes in the initial planned path to obtain a continuous candidate trajectory, including: determining multiple path segments to be smoothed based on the positional change relationship between adjacent path nodes in the initial planned path; performing curve smoothing processing on the initial planned path based on the connection direction and curvature continuity constraints between the multiple path segments to be smoothed to obtain a continuous candidate trajectory; wherein the continuous candidate trajectory passes through at least some of the multiple path nodes, or is located within the safe passage area corresponding to the multiple path nodes.

[0024] According to an embodiment of this application, collision verification is performed on continuous candidate trajectories based on the spatial occupancy data and motion change data of environmental obstacles in the target working environment. This includes: sampling the continuous candidate trajectories at a preset spatial sampling interval to obtain multiple trajectory sampling points; determining the predicted spatial occupancy area of ​​the environmental obstacle at the corresponding time based on the predicted arrival time of each trajectory sampling point; calculating the safety interval between each trajectory sampling point and the corresponding predicted spatial occupancy area; and determining that the continuous candidate trajectories meet the collision verification conditions if all safety intervals meet preset safety conditions.

[0025] This application provides a robot path planning device based on pulsed membrane potential readout, comprising: a state data acquisition module, used to acquire current path planning state data of a target robot in a target working environment, wherein the current path planning state data is used to characterize the spatial and motion relationships between the target robot, target location objects, environmental obstacle objects, and sensor detection objects; and a context and state unit determination module, used to generate path decision context data based on the current path planning state data, and to divide the current path planning state data into state units according to the path decision dimension to obtain multiple path state units, wherein the path decision context data is used to record current path nodes, historical path nodes, candidate movement actions, and environmental obstacles. The system establishes a correspondence between obstacle change states and path feedback evaluation data; a state correlation calculation module calculates the state correlation between each path state unit and preset path decision factors, and determines the pulse coding weight and pulse triggering sequence corresponding to each path state unit based on the state correlation; a pulse timing coding module performs pulse timing coding on each path state unit according to the pulse coding weight and pulse triggering sequence to obtain a pulse state sequence corresponding to the current path planning state data; and a pulse membrane potential calculation module inputs the pulse state sequence into the pulse neural computation layer, and performs cumulative update of the membrane potential of the corresponding pulse neurons in the pulse neural computation layer according to the pulse events in the pulse state sequence within a preset simulation time window, and based on... The pulse output state of the spiking neuron is determined based on the updated membrane potential state. The cumulative membrane potential update is an event-triggered state update performed on the corresponding spiking neuron when a pulse event input is present. The action readout input determination module determines the action readout input data corresponding to multiple candidate movement actions based on the spiking neuron's membrane potential state, pulse output state, and path decision context data. The non-pulse membrane potential readout module inputs each action readout input data into the non-pulse readout neurons corresponding to the multiple candidate movement actions and calculates the membrane potential change data of each non-pulse readout neuron within a preset simulation time window. The action evaluation result determination module determines the action evaluation result based on the membrane potential changes of each non-pulse readout neuron. The system analyzes data to calculate the correlation between the action value of each candidate movement action and preset path decision factors, and performs constraint correction on the action value correlation based on path decision context data to obtain the path action evaluation result corresponding to each candidate movement action; the path node update module is used to determine the target movement action from multiple candidate movement actions based on the path action evaluation results corresponding to each candidate movement action, and update the path node data of the target robot according to the target movement action; the path feedback update module is used to generate path feedback evaluation data corresponding to the updated path nodes based on the updated path node data and the environmental state change data in the target working environment, and update the path decision context data based on the path feedback evaluation data.The target path generation module is used to iteratively determine the next path node based on updated path decision context data and updated current path planning state data when the target robot has not reached the target position and the preset termination condition has not been met. This generates a target planned path based on multiple path nodes determined from the current position to the target position.

[0026] According to an embodiment of this application, the target path generation module includes: an initial path generation submodule, used to iteratively determine multiple path nodes based on updated path decision context data and updated current path planning state data, and generate an initial planned path according to the determination order of each path node; a trajectory continuity processing submodule, used to perform trajectory continuity processing on adjacent path nodes in the initial planned path to obtain continuous candidate trajectories; a collision verification submodule, used to perform collision verification on the continuous candidate trajectories based on the spatial occupancy data and motion change data of environmental obstacle objects in the target working environment; and a target path determination submodule, used to determine the continuous candidate trajectories as the target planned path if the continuous candidate trajectories meet the collision verification conditions.

[0027] According to the embodiments of this application, the current path planning status data includes the robot's current position data, target position data, static obstacle position data, dynamic obstacle position data, obstacle movement direction data, multi-directional sensor detection data, relative geometric data between the robot and the target position object, and relative risk data between the robot and environmental obstacle objects.

[0028] According to an embodiment of this application, the state correlation calculation module includes: an initial correlation calculation submodule, used to calculate the initial state correlation between each path state unit and a preset path decision factor based on the spatial position change data, motion trend change data, and sensor detection intensity data corresponding to each path state unit; a context correction submodule, used to perform context correction on the initial state correlation based on historical path nodes and path feedback evaluation data in the path decision context data to obtain the corrected state correlation; and a coding parameter determination submodule, used to determine the pulse coding weight and pulse triggering timing of each path state unit in the pulse timing coding process based on the corrected state correlation.

[0029] According to an embodiment of this application, the pulse timing coding module includes: a pulse trigger intensity determination submodule, used to determine the pulse trigger intensity of each path state unit within a preset simulation time window based on the pulse coding weight corresponding to each path state unit; a pulse trigger time step determination submodule, used to determine the pulse trigger time step of each path state unit within a preset simulation time window based on the pulse trigger timing corresponding to each path state unit; and a pulse state sequence generation submodule, used to generate a pulse state sequence formed by arranging multiple pulse events in chronological order based on the pulse trigger intensity and the pulse trigger time step.

[0030] According to an embodiment of this application, the pulse membrane potential calculation module includes: a pulse event detection submodule, used to detect whether there is a pulse event input to the target spiking neuron in the pulse state sequence at the current time step of a preset simulation time window; an event trigger update submodule, used to perform membrane potential accumulation update of the target spiking neuron based on the pulse weight corresponding to the pulse event and the membrane potential state of the target spiking neuron in the previous time step when there is a pulse event input to the target spiking neuron; and a non-event state processing submodule, used to maintain the membrane potential state of the target spiking neuron in the previous time step or to update the state of the target spiking neuron based on a preset membrane potential decay rule when there is no pulse event input to the target spiking neuron.

[0031] According to an embodiment of this application, the pulse membrane potential calculation module further includes: a membrane potential comparison submodule, used to compare the updated membrane potential state with a preset membrane potential threshold; a pulse output and reset submodule, used to generate a pulse output state corresponding to the spiking neuron and perform a reset process on the membrane potential state of the spiking neuron when the updated membrane potential state meets the pulse triggering condition; and a pulse inhibit output submodule, used to inhibit the spiking neuron from outputting pulse events when the updated membrane potential state does not meet the pulse triggering condition.

[0032] According to an embodiment of this application, the non-pulse membrane potential readout module includes: a membrane potential value acquisition submodule, used to acquire the membrane potential values ​​of each non-pulse readout neuron at multiple time steps within a preset simulation time window; an action value representation determination submodule, used to determine the action value representation data corresponding to each candidate movement action based on the maximum membrane potential of each non-pulse readout neuron within the preset simulation time window; and an action value correlation calculation submodule, used to calculate the action value correlation between each candidate movement action and preset path decision factors based on the action value representation data corresponding to each candidate movement action.

[0033] According to an embodiment of this application, the action evaluation result determination module includes: an action constraint determination submodule, used to determine the target approach constraint, obstacle avoidance constraint, motion direction continuity constraint, and dynamic collision risk constraint corresponding to each candidate movement action based on path decision context data; an association correction submodule, used to perform weighted correction on the action value association of each candidate movement action based on the target approach constraint, obstacle avoidance constraint, motion direction continuity constraint, and dynamic collision risk constraint; and an action evaluation result generation submodule, used to obtain the path action evaluation result corresponding to each candidate movement action based on the weighted correction of the action value association.

[0034] According to an embodiment of this application, the path feedback update module includes: a distance change calculation submodule, used to calculate the distance change data of the target robot relative to the target location object based on the path node data before and after the update; an obstacle avoidance risk change calculation submodule, used to calculate the obstacle avoidance risk change data of the target robot relative to the environmental obstacle object based on the updated path node data and the location change data of the environmental obstacle object; a direction continuity evaluation submodule, used to calculate the motion direction continuity evaluation data of the target robot based on the updated path node data, historical path node data, and the target location object; a feedback evaluation data generation submodule, used to generate path feedback evaluation data based on the distance change data, obstacle avoidance risk change data, and motion direction continuity evaluation data; and a context update submodule, used to update the path decision context data based on the path feedback evaluation data.

[0035] This application also provides an electronic device, including a processor and a memory. The memory stores a computer program, which, when executed by the processor, implements the robot path planning method based on pulse membrane potential readout in any of the above embodiments.

[0036] This application also provides a computer-readable storage medium storing a computer program. When the computer program is executed by a processor, it implements the robot path planning method based on pulse membrane potential readout in any of the above embodiments.

[0037] This application acquires current path planning state data to characterize the spatial and motion relationships between the robot, target location object, environmental obstacle object, and sensor detection object. It then generates path decision context data and divides the path state into units. Based on the state correlation between the path state units and preset path decision factors, it determines the pulse coding weight and pulse triggering sequence, allowing the path planning state to enter the pulse neural computation layer in the form of a pulse state sequence. Within a preset simulation time window, the pulse neural computation layer performs event-triggered membrane potential accumulation updates only on the corresponding pulse neurons based on pulse events. It also uses the membrane potential change data of non-pulse readout neurons to calculate the action value correlation between candidate movement actions and path decision factors. Finally, it combines the path decision context data to perform constraint correction, path node updates, and feedback evaluation updates. Therefore, compared to existing deep learning-based robot path planning schemes that perform high-frequency full-scale calculations on a large number of continuous state features, this application can transform the path state calculation process into a sparse pulse event-driven membrane potential accumulation process, reducing invalid calculations and the frequency of continuous multiplication and accumulation calculations in pulse-free input states, thereby reducing the computational load and energy consumption of path planning on the robot body. At the same time, by updating the path decision context data through path feedback evaluation data, subsequent path node selection can be dynamically corrected by combining feedback results such as target approach, obstacle changes, and historical motion continuity, thereby improving the real-time performance, obstacle avoidance safety, and path generation stability of path planning in dynamic obstacle environments.

[0038] The above description is only an overview of the technical solution of this application. In order to better understand the technical means of this application and to implement it in accordance with the contents of the specification, and to make the above and other objects, features and advantages of this application more obvious and understandable, specific embodiments of this application are given below. Attached Figure Description

[0039] Various other advantages and benefits will become apparent to those skilled in the art upon reading the following detailed description of preferred embodiments. The accompanying drawings are for illustrative purposes only and are not intended to limit the scope of this application. Furthermore, the same reference numerals denote the same parts throughout the drawings. In the drawings: Figure 1 A flowchart of an embodiment of the robot path planning method based on pulse membrane potential readout provided in this application; Figure 2 A schematic diagram of a structure of an embodiment of the robot path planning device based on pulse membrane potential readout provided in this application; Figure 3 A schematic diagram of the structure of an embodiment of the electronic device provided in this application. Detailed Implementation

[0040] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0041] In the description of this application, it should be understood that the terms "first," "second," etc., are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of technical features. Therefore, a feature specified with "first," "second," etc., may explicitly or implicitly include one or more of that feature. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0042] In this application, unless otherwise expressly specified and limited, the terms "connection" and "linked" should be interpreted broadly. For example, they can refer to direct connections or indirect connections; they can refer to electrical connections, communication connections, data transmission relationships, or function call relationships. Those skilled in the art can understand the specific meaning of the above terms in this application based on the specific circumstances.

[0043] First Embodiment Robot path planning is a crucial step in autonomous navigation for mobile robots. In a target work environment, the robot typically needs to determine its next movement in real time based on its own position, the target position, static obstacles, dynamic obstacles, and sensor detection results, and to form a planned path to the target location among multiple path nodes. In industrial inspection, warehouse handling, indoor service, park delivery, and complex ground movement scenarios, path planning needs to consider not only path length but also obstacle avoidance safety, the movement trends of dynamic obstacles, path continuity, and the computational resource consumption of the robot itself.

[0044] Existing path planning schemes typically include graph search-based path planning methods, sampling-based path planning methods, artificial potential field-based path planning methods, optimization algorithm-based path planning methods, and deep neural network-based path decision-making methods. Traditional path planning methods are applicable to known or regular environments, but they tend to incur high search or path recalculation costs in scenarios with frequently changing dynamic obstacles, high-dimensional multi-sensor inputs, and continuously updated local environmental states. While deep neural network-based path planning methods can represent the features of complex environmental states, they usually require continuous numerical activation and a large number of multiplication and accumulation operations, placing high demands on the computing power, storage access capabilities, and battery capacity of the robot's on-board processor.

[0045] When performing path planning on the robot itself, computational resources and energy supply are typically limited. If the path planning process performs full computation on a large number of state features and neural network nodes in each planning cycle, a significant amount of invalid computation will occur even if some environmental states do not change significantly. This type of computation easily leads to increased planning delays and power consumption, further affecting the obstacle avoidance response speed and endurance of the mobile robot in dynamic environments. These problems are particularly pronounced in scenarios such as low-power cruise, edge autonomous control, and real-time obstacle avoidance.

[0046] This application provides a robot path planning method based on pulsed membrane potential readout. This method divides the current path planning state data into multiple path state units according to the path decision dimension, and determines the pulse coding weight and pulse triggering sequence based on the state correlation between the path state units and preset path decision factors. Then, the path state information can be converted into a pulse state sequence, and the membrane potential is accumulated and updated in the pulse neural computation layer based on the pulse event triggering of the corresponding pulse neurons. Since the membrane potential accumulation and update can be performed when a pulse event input exists, the path planning state calculation process can be transformed from continuous full-scale calculation to sparse event-driven calculation.

[0047] Furthermore, this embodiment does not rely solely on the number of pulse outputs to directly determine the action. Instead, it combines the membrane potential state of the spiking neurons, the pulse output state, and path decision context data to determine the action readout input data corresponding to multiple candidate movement actions. It then calculates the action value correlation between candidate movement actions and preset path decision factors using the membrane potential change data of non-spiking readout neurons. This allows the temporal membrane potential information formed in the spiking neuron computation layer to be converted into readout results more suitable for action evaluation, improving the stability and interpretability of candidate movement action evaluation.

[0048] This application embodiment also records the correspondence between current path nodes, historical path nodes, candidate movement actions, environmental obstacle changes, and path feedback evaluation data through path decision context data. After each target movement action is determined and path node data is updated, path feedback evaluation data can be generated based on the updated path node data and environmental state change data, and the path decision context data is updated based on the path feedback evaluation data. In this way, the subsequent path node determination process can be dynamically corrected by combining historical path feedback, target approach status, obstacle avoidance risk changes, and movement direction continuity, reducing problems such as path oscillation, repeated detours, or delayed response to dynamic obstacles.

[0049] Through the above-described solution, the embodiments of this application can address the problems of high computational load, high energy consumption, insufficient dynamic obstacle avoidance response, and inadequate utilization of path feedback in existing robot path planning schemes based on deep learning or continuous state computation. By converting path state processing into a membrane potential accumulation process triggered by pulse events, invalid updates in the absence of pulse input are reduced, and candidate movement actions are evaluated through non-pulse membrane potential readout. Thus, while ensuring target approach, obstacle avoidance, and path continuity, the computational load and energy consumption in the robot path planning process are reduced, and the real-time performance, obstacle avoidance safety, and path generation stability of path planning in dynamic environments are improved.

[0050] Second Embodiment This embodiment combines Figure 1 This paper describes the robot path planning method based on pulse membrane potential readout provided in this application. Figure 1 A flowchart illustrating an embodiment of the method of this application is shown. The method according to an embodiment of this application includes: S101, acquire the current path planning state data of the target robot in the target working environment. The current path planning state data is used to characterize the spatial and motion relationships between the target robot, the target location object, the environmental obstacle object, and the sensing and detection object.

[0051] In this embodiment, the target robot can be a mobile robot, inspection robot, warehouse handling robot, delivery robot, service robot, or other robot that needs to move autonomously in the working environment. The target working environment can be an indoor environment, an outdoor environment, a semi-structured environment, or a dynamically changing environment. The current path planning status data can be acquired by the robot's positioning unit, inertial measurement unit, LiDAR, depth camera, millimeter-wave radar, ultrasonic sensor, vision sensor, or other environmental perception devices.

[0052] Current path planning state data describes the environmental and motion states required for path planning. For example, it may include the robot's current pose, target position, obstacle position, obstacle velocity, sensor detection distance, sensor detection direction, relative distance and angle between the robot and the target, safe distance between the robot and obstacles, and predicted motion direction of dynamic obstacles. Acquiring this data provides the foundational input for subsequent state unit partitioning, pulse coding, and motion evaluation.

[0053] In one implementation, the target robot can acquire current path planning state data according to a fixed planning cycle. For example, environmental state data can be acquired every preset time interval and combined with the path node data determined in the previous planning cycle to form the state input corresponding to the current planning cycle. In another implementation, the acquisition of current path planning state data can also be triggered when the sensor detects a change in the state of an obstacle, the robot deviates from the current path, or the target position is updated.

[0054] Furthermore, in this embodiment, the current path planning status data may include robot current position data, target position data, static obstacle position data, dynamic obstacle position data, obstacle movement direction data, multi-directional sensor detection data, relative geometric data between the robot and the target, and relative risk data between the robot and the obstacle.

[0055] Specifically, the robot's current position data can represent the robot's coordinates, orientation, or posture in the target working environment; the target position data can represent the location of the target point or target area; static obstacle position data can represent the space occupied by non-moving obstacles such as walls, shelves, equipment, and fixed facilities; dynamic obstacle position data and obstacle movement direction data can represent the position and movement trend of personnel, vehicles, mobile devices, or other moving targets; multi-directional sensor detection data can include distance detection values ​​in different directions, the probability of obstacle presence, or the degree of passability; relative geometric data can include the target relative distance, target azimuth angle, and target direction deviation; relative risk data can be determined based on the distance between the robot and the obstacle, the relative speed, and the predicted rendezvous time.

[0056] S102, generate path decision context data based on the current path planning state data, and divide the current path planning state data into state units according to the path decision dimension to obtain multiple path state units. The path decision context data is used to record the correspondence between the current path node, historical path node, candidate movement action, environmental obstacle change state, and path feedback evaluation data.

[0057] In this embodiment, path decision context data is used to store context information related to path decision-making during multiple path node determination processes. This context information may include the current path node, historical path nodes that have been traversed, a set of candidate movement actions, historical evaluation results of candidate movement actions, changes in the status of environmental obstacles, path feedback evaluation data, termination condition judgment results, and other data used for path planning and decision-making.

[0058] A path state unit is a state representation unit obtained by dividing the current path planning state data according to the path decision dimensions. By splitting complex environmental states into multiple path state units, the subsequent pulse coding process can be more closely aligned with the path planning task. For example, different state units can correspond to decision dimensions such as target approach, obstacle proximity, dynamic obstacle movement, sensor orientation detection, and safety angle. In this way, each path state unit can independently participate in state correlation calculations and further convert them into corresponding pulse events.

[0059] In one implementation, path decision context data can be stored in the form of a structured data table, a cache queue, a path node sequence, or a state dictionary. Current path nodes and historical path nodes can form a path node sequence; candidate movement actions can include forward, left turn, right turn, backward, stop, diagonal movement, or movement in multiple directions defined by the discrete action space; environmental obstacle changes can be determined based on the changes in the obstacle's position, speed, and risk level within a continuous planning cycle; path feedback evaluation data can be used to record whether the previous target movement action brought the robot closer to the target, away from obstacles, and maintained directional continuity.

[0060] Furthermore, in embodiments of this application, the path state unit may include a target approach state unit, an obstacle proximity state unit, a dynamic obstacle motion state unit, a sensor direction detection state unit, and a safety angle state unit.

[0061] Specifically, the target approach state unit can be used to indicate whether the robot's current movement trend is towards the target position; the obstacle proximity state unit can be used to indicate the distance and distribution density of obstacles around the robot; the dynamic obstacle motion state unit can be used to indicate the velocity direction, predicted position, and potential collision risk of dynamic obstacles; the sensor direction detection state unit can be used to indicate the passability detected by sensors in multiple directions; and the safety angle state unit can be used to indicate the safe angle between the candidate movement direction and the obstacle direction. Through this classification method, the path planning state data can be made more suitable for pulse timing coding and path action evaluation.

[0062] S103, calculate the state correlation between each path state unit and the preset path decision factors, and determine the pulse coding weight and pulse triggering timing corresponding to each path state unit based on the state correlation.

[0063] In this embodiment, the preset path decision factors can be understood as the main evaluation dimensions affecting path planning decisions. State correlation is used to characterize the degree of influence of a certain path state unit on a certain preset path decision factor. For example, there is usually a high correlation between the target approach state unit and the target approach factor; there is usually a high correlation between the obstacle proximity state unit and the obstacle avoidance factor; and there is usually a high correlation between the dynamic obstacle motion state unit and the dynamic collision risk factor.

[0064] Pulse coding weights can be used to represent the pulse contribution intensity of a path state unit during the pulse coding process. Higher state correlation allows the pulse events generated by the corresponding path state unit to have higher weights or a stronger impact on membrane potential accumulation and updating. Pulse triggering timing can be used to represent the occurrence time, order, or triggering frequency of pulse events corresponding to a path state unit within a preset simulation time window. Through pulse coding weights and pulse triggering timing, path planning state data can be converted into pulse event representations with both time and intensity information.

[0065] In one implementation, state correlation can be calculated based on data such as the numerical value of the path state unit, normalized risk level, target direction deviation, obstacle distance, and sensor detection intensity. Alternatively, correlation rules between state units and path decision factors can be pre-set. For example, when a sensor detects an obstacle distance less than a safety threshold in a certain direction, the obstacle avoidance correlation for that direction is increased; when the angle between the candidate movement direction and the target direction is small, the target approach correlation is increased; and when the predicted trajectory of a dynamic obstacle intersects with the robot's candidate movement direction, the dynamic collision risk correlation is increased.

[0066] Furthermore, in the embodiments of this application, the preset path decision factors may include target approach factors, obstacle avoidance factors, motion direction continuity factors, and dynamic collision risk factors.

[0067] Specifically, the target approach factor is used to evaluate whether the movement action helps to shorten the distance between the robot and the target position; the obstacle avoidance factor is used to evaluate whether the movement action can enable the robot to avoid static or dynamic obstacles; the motion direction continuity factor is used to evaluate whether the movement action maintains a smooth connection with the historical movement direction; and the dynamic collision risk factor is used to evaluate whether the robot is likely to collide with dynamic obstacles when executing candidate movement actions. By setting the above path decision factors, the path planning results can take into account efficiency, safety, and continuity.

[0068] Furthermore, embodiments of this application may also calculate the initial state correlation between each path state unit and preset path decision factors based on the spatial position change data, motion trend change data, and sensor detection intensity data corresponding to each path state unit; perform context correction on the initial state correlation based on historical path nodes and path feedback evaluation data in the path decision context data to obtain the corrected state correlation; and determine the pulse coding weight and pulse triggering timing of each path state unit in the pulse timing coding process based on the corrected state correlation.

[0069] Specifically, spatial position change data can represent changes in the relative spatial relationship between the robot, the target, and obstacles; motion trend change data can represent changes in the velocity direction of the robot or obstacles; sensor detection intensity data can represent the degree of obstacle approach or spatial passability detected by sensors in different directions. Initial state correlation can be calculated based on the above data. Subsequently, it can be corrected by combining historical path nodes and path feedback evaluation data. For example, if historical path feedback indicates that movement in a certain direction easily leads to the robot approaching an obstacle, the correlation between the relevant state unit in that direction and obstacle avoidance factors or dynamic collision risk factors can be increased; if a certain movement direction repeatedly causes the robot to approach the target without generating a high obstacle avoidance risk, the correlation between the corresponding state unit and target approach factors or motion direction continuity factors can be increased.

[0070] S104. Based on the pulse coding weight and pulse triggering timing, pulse timing coding is performed on each path state unit to obtain the pulse state sequence corresponding to the current path planning state data.

[0071] In this embodiment, pulse timing encoding is used to convert path state units into discrete pulse events. Each pulse event may include information such as event time, event source, event weight, and corresponding state unit identifier. A pulse state sequence can be formed by arranging multiple pulse events in chronological order for input into the pulse neural computation layer.

[0072] Pulse coding weights can influence the intensity of pulse events or their contribution to membrane potential accumulation, while pulse triggering timing can affect the triggering time of pulse events within a preset simulation time window. By combining pulse coding weights and pulse triggering timing, more important, urgent, or riskier path state units can be triggered earlier or generate stronger pulse inputs. For example, when an obstacle is close in front of the robot, the corresponding obstacle-adjacent state unit can generate a stronger pulse event at the beginning of the simulation time window, enabling the pulse neural computation layer to respond to obstacle avoidance requirements in a timely manner.

[0073] In one implementation, normalized state correlations can be mapped to pulse triggering intensity, and the urgency of state changes can be mapped to pulse triggering time steps. For target approach-related states, pulse events can be generated based on target direction deviation; for obstacle avoidance-related states, pulse events can be generated based on obstacle distance and safety threshold; and for dynamic collision risk-related states, pulse events can be generated based on predicted collision time and relative velocity. Thus, the pulse state sequence can reflect the priority and risk level of path planning states in the time dimension.

[0074] Furthermore, embodiments of this application may also determine the pulse trigger intensity of each path state unit within a preset simulation time window based on the pulse coding weight corresponding to each path state unit; determine the pulse trigger time step of each path state unit within a preset simulation time window based on the pulse trigger timing corresponding to each path state unit; and generate a pulse state sequence formed by arranging multiple pulse events in chronological order based on the pulse trigger intensity and the pulse trigger time step.

[0075] Specifically, the preset simulation time window can include multiple discrete time steps. For each path state unit, its pulse trigger intensity at one or more time steps can be determined according to the pulse coding weights, and the corresponding trigger time step can be determined according to the pulse trigger timing sequence. Then, the pulse events generated by different path state units are arranged in chronological order to form a pulse state sequence. This sequence can be input into the pulse neural computation layer so that subsequent membrane potential updates are performed in the order of event occurrence.

[0076] S105, the pulse state sequence is input into the pulse neural computation layer. Within the preset simulation time window, the membrane potential of the corresponding spiking neurons in the pulse neural computation layer is cumulatively updated according to the pulse events in the pulse state sequence. The pulse output state of the spiking neurons is determined according to the updated membrane potential state. The membrane potential cumulative update is an event-triggered state update performed on the corresponding spiking neurons when there is a pulse event input.

[0077] In this embodiment, the spiking neural computation layer may include multiple spiking neurons, each of which can be used to receive pulse events corresponding to one or more path state units. The membrane potential state of the spiking neuron is used to represent its cumulative response to the input pulse event. When a pulse event is input to the target spiking neuron, the target spiking neuron can update its membrane potential state according to the weight corresponding to the pulse event.

[0078] Event-triggered state updates are a key mechanism for reducing computational complexity in this application. Unlike continuous neural networks that perform multiply-accumulate calculations on all nodes at each time step, this embodiment only performs membrane potential accumulation updates on the corresponding spiking neurons when a spiking event input is present. For spiking neurons without spiking event input, the membrane potential state of the previous time step can be maintained, or a low-complexity update can be performed according to a preset membrane potential decay rule. This reduces a significant amount of unnecessary computation.

[0079] The pulse output state can be determined based on whether the membrane potential state reaches a preset membrane potential threshold. When the membrane potential state of a spiking neuron reaches or exceeds the threshold, it can output a pulse event and reset or reduce the membrane potential state according to a preset reset rule. This pulse output state can represent the significant impact of the path state on the readout of subsequent actions. For example, the output pulse of a spiking neuron related to obstacle avoidance may indicate that there is a strong obstacle avoidance requirement in the current environment.

[0080] Furthermore, embodiments of this application may also detect whether there is a pulse event input to the target spiking neuron in the pulse state sequence at the current time step of a preset simulation time window; if there is a pulse event input to the target spiking neuron, perform membrane potential accumulation update on the target spiking neuron based on the pulse weight corresponding to the pulse event and the membrane potential state of the target spiking neuron in the previous time step; if there is no pulse event input to the target spiking neuron, maintain the membrane potential state of the target spiking neuron in the previous time step, or update the state of the target spiking neuron based on a preset membrane potential decay rule.

[0081] Specifically, the preset membrane potential decay rule can be determined based on a time constant, decay ratio, or fixed decay amount. If there is no input pulse event in the current time step, the target spiking neuron does not need to perform complex calculations; it only needs to maintain the membrane potential state or perform a simple decay operation. If there is an input pulse event, it is updated cumulatively based on the input weights and the membrane potential state of the previous time step. This allows computational resources to be concentrated on path state units where state changes occur.

[0082] Furthermore, embodiments of this application may also compare the updated membrane potential state with a preset membrane potential threshold; if the updated membrane potential state meets the pulse triggering condition, generate the pulse output state corresponding to the spiking neuron and perform a reset process on the membrane potential state of the spiking neuron; if the updated membrane potential state does not meet the pulse triggering condition, prohibit the spiking neuron from outputting pulse events.

[0083] Specifically, the preset membrane potential threshold can be set according to the importance of path decision factors. Spike neurons related to safety risks can have a lower threshold to respond quickly to dangerous situations; spike neurons related to target approach can have their thresholds set according to path stability requirements. The reset process can include resetting the membrane potential to its initial value, subtracting the threshold, or decaying it to a preset level, thereby preparing for the next impulse response.

[0084] S106. Based on the membrane potential state, pulse output state, and path decision context data of the spiking neuron, determine the action readout input data corresponding to multiple candidate movement actions.

[0085] In this embodiment, the candidate movement action can be a discrete action or a discretized result of a continuous action that the robot can execute at the current path node. For example, the candidate movement action can include moving forward, moving to the left front, moving to the right front, turning left, turning right, stopping, decelerating, or other preset movement directions. The action readout input data is used to represent the spiking neural computation result that needs to be read for a certain candidate movement action in the current path state.

[0086] The membrane potential state of a spiking neuron can reflect the cumulative impact of a spiking event within a preset simulation time window, while the spiking output state can reflect the strong response state upon reaching the trigger condition. Path decision context data can provide current path nodes, historical path nodes, candidate movement actions, environmental obstacle changes, and path feedback evaluation data. Therefore, action readout input data can simultaneously include current state calculation results and historical path decision information.

[0087] In one implementation, the membrane potential state, pulse output state, and contextual constraint information of spiking neurons associated with a candidate movement action can be combined to form the action readout input data corresponding to that candidate movement action. For example, for the action of "moving to the left front", the state of spiking neurons associated with left front sensor detection, safety angle, target direction deviation, and dynamic obstacle prediction risk can be read, and historical path nodes can be combined to determine whether the action will cause path jitter or repeated turning.

[0088] Furthermore, embodiments of this application may also determine multiple candidate movement actions and corresponding action constraints based on path decision context data; calculate the action response correlation between the membrane potential state and pulse output state of the spiking neuron and each candidate movement action; and determine the action readout input data corresponding to each candidate movement action from the membrane potential state and pulse output state of the spiking neuron based on the action response correlation and action constraints.

[0089] Specifically, action constraints can include target approach constraints, obstacle avoidance constraints, motion direction continuity constraints, velocity constraints, turning angle constraints, and safety distance constraints. Action response relevance can be determined based on the correspondence between the direction of a candidate movement action and the direction, risk type, or decision factor represented by the spiking neuron. This approach avoids directly using impulse responses unrelated to a particular candidate action for action evaluation, improving the relevance of the action readout input data.

[0090] S107, input the readout data of each action into the non-pulse readout neurons corresponding to the multiple candidate movement actions respectively, and calculate the membrane potential change data of each non-pulse readout neuron within the preset simulation time window.

[0091] In this embodiment, non-pulse readout neurons are used to read out the membrane potential of candidate movement actions. Unlike pulsating neurons, which express the event triggering state through pulse output, non-pulse readout neurons can represent the continuous evaluation state of candidate movement actions through membrane potential changes. This avoids the problem of unstable action evaluation caused by relying solely on the number of pulse outputs.

[0092] Multiple candidate movement actions can each correspond to multiple non-pulse readout neurons. For example, each candidate movement action can correspond to one non-pulse readout neuron. After receiving the corresponding movement readout input data, the non-pulse readout neuron generates membrane potential change data within a preset simulation time window. The membrane potential change data can include membrane potential values ​​at multiple time steps, peak membrane potential, mean membrane potential, integral membrane potential, or membrane potential change trend.

[0093] In one implementation, non-pulse readout neurons can weighted accumulate the membrane potential state, pulse output state, and contextual constraint information from the pulse neural computation layer, but do not output pulse events. Their membrane potential change curves can serve as the basis for evaluating action value. If a candidate movement action helps approach the target and has a low obstacle avoidance risk, the corresponding non-pulse readout neuron can obtain a higher membrane potential value; if a candidate action may lead to collision risk or directional jitter, the corresponding membrane potential value can be reduced or subject to constraint correction.

[0094] S108. Based on the membrane potential change data of each non-pulse readout neuron, calculate the action value correlation between each candidate movement action and the preset path decision factors, and constrain and correct the action value correlation according to the path decision context data to obtain the path action evaluation result corresponding to each candidate movement action.

[0095] In this embodiment, action value correlation is used to represent the relationship between candidate movement actions and path decision-making factors such as target approach, obstacle avoidance, continuity of movement direction, and dynamic collision risk. Membrane potential change data of non-pulse readout neurons can represent the comprehensive response of candidate movement actions within a preset simulation time window; therefore, action value correlation can be calculated based on membrane potential change data.

[0096] Path decision context data can be used to adjust the constraints on the relevance of action values. For example, if a candidate movement action can shorten the distance to the target but significantly approaches a dynamic obstacle, its path action evaluation result can be reduced based on dynamic collision risk constraints; conversely, if a candidate movement action is continuous with the historical path direction and has a low obstacle avoidance risk, its evaluation result can be improved. Through context constraint adjustment, path action evaluation results can be made more consistent with actual path planning needs.

[0097] Path action evaluation results can be numerical scores, ranking results, action priorities, action feasibility markers, or comprehensive evaluation vectors. These results are used to subsequently determine the target movement action from multiple candidate movements. In dynamic environments, path action evaluation results can be dynamically adjusted as environmental conditions change and path feedback evaluation data is updated, thereby improving the adaptability of path planning.

[0098] Furthermore, embodiments of this application may also obtain the membrane potential values ​​of each non-pulse readout neuron at multiple time steps within a preset simulation time window; determine the action value representation data corresponding to each candidate movement action based on the maximum membrane potential of each non-pulse readout neuron within the preset simulation time window; and calculate the action value correlation between each candidate movement action and preset path decision factors based on the action value representation data corresponding to each candidate movement action.

[0099] Specifically, the maximum membrane potential represents the highest response intensity obtained by a candidate movement action within a preset simulation time window. Determining the action value representation data based on the maximum membrane potential can reduce the impact of unstable instantaneous pulse output on action evaluation. In other implementations, the action value representation data can also be determined by combining the mean membrane potential, the integral membrane potential value, or the membrane potential value at the last time step, to adapt to the evaluation requirements of different path planning tasks.

[0100] Furthermore, embodiments of this application may also determine the target approach constraints, obstacle avoidance constraints, motion direction continuity constraints, and dynamic collision risk constraints corresponding to each candidate movement action based on path decision context data; perform weighted correction on the action value correlation corresponding to each candidate movement action based on the target approach constraints, obstacle avoidance constraints, motion direction continuity constraints, and dynamic collision risk constraints; and obtain the path action evaluation result corresponding to each candidate movement action based on the weighted correction of the action value correlation.

[0101] Specifically, the target approach constraint can be determined based on the distance change between the robot and the target position after the candidate movement is executed; the obstacle avoidance constraint can be determined based on the safe distance between the robot and the obstacle after the candidate movement is executed; the motion direction continuity constraint can be determined based on the angle change between the candidate movement and the historical movement direction; and the dynamic collision risk constraint can be determined based on the safe interval between the predicted position of the dynamic obstacle and the predicted position of the robot. By weighting and correcting the correlation of action value, the final path action evaluation result can take into account approaching the target, avoiding obstacles, and maintaining path smoothness.

[0102] S109, Based on the path action evaluation results corresponding to each candidate movement action, determine the target movement action from multiple candidate movement actions, and update the path node data of the target robot according to the target movement action.

[0103] In this embodiment, the target movement action can be any action among multiple candidate movement actions that meets preset action selection criteria. These preset action selection criteria may include having the highest evaluation result, an evaluation result exceeding a threshold, obstacle avoidance risk meeting safety conditions, target approach effect meeting requirements, or the lowest overall cost. After the target movement action is determined, the path node data of the target robot can be updated based on its movement direction, step size, or speed control.

[0104] Path node data can include information such as the current path node, the next path node, the path node sequence, node timestamps, node directions, and node safety assessments. By progressively updating the path node data, a path node sequence from the current position to the target position can be formed. This path node sequence can be used for robot motion control, as well as for subsequent path smoothing and collision verification.

[0105] In one implementation, if candidate movement actions include multiple discrete directions, the candidate movement action with the highest path action evaluation result and satisfying safety constraints can be determined as the target movement action. If the evaluation results of multiple candidate movement actions are similar, the action with better continuity with historical movement directions can be selected first to reduce path jitter. If none of the candidate movement actions meet the safety conditions, the action can be stopped, decelerated, or environmental state data can be reacquired.

[0106] Furthermore, embodiments of this application may also sort multiple candidate movement actions according to the path action evaluation results corresponding to each candidate movement action; determine the candidate movement action whose sorting result meets the preset action selection conditions as the target movement action; determine the next path node after the target robot moves from the current path node based on the movement direction and movement step size corresponding to the target movement action; and write the next path node into the path node sequence in the path decision context data to update the path node data of the target robot.

[0107] Specifically, the sorting can be done from highest to lowest based on the path action evaluation results, or it can be sorted after filtering according to risk priority. The next path node can be calculated based on the coordinates of the current path node, the direction of movement, and the step size. After the next path node is written into the path node sequence, the path decision context data can record the path changes that the robot has determined, providing a basis for subsequent path feedback evaluation and continuous processing.

[0108] S110: Based on the updated path node data and the environmental state change data in the target working environment, generate path feedback evaluation data corresponding to the updated path nodes, and update the path decision context data based on the path feedback evaluation data.

[0109] In this embodiment, path feedback evaluation data is used to evaluate the impact of target movement actions or planning on path quality. This feedback may include information such as distance changes, risk changes, changes in directional continuity, and path benefits. By generating path feedback evaluation data, the current path node update results can be fed back to the next round of path decision-making.

[0110] Environmental state change data can include dynamic obstacle position changes, static obstacle detection changes, sensor detection intensity changes, and changes in passable areas. By combining the updated path node data and environmental state change data, it can be determined whether the current path node brings the robot closer to the target, avoids obstacles, increases or decreases the risk of collision, and maintains the continuity of the movement direction.

[0111] In one implementation, path feedback evaluation data can be represented in the form of scores or labels. For example, if the updated path node is closer to the target and the safety distance requirement is met, the feedback can be a positive evaluation; if the updated path node is closer to a dynamic obstacle or causes a sharp change in path direction, the feedback can be a risk evaluation. This feedback evaluation data can be used to update the historical evaluation relationships in the path decision context data, making subsequent state correlation calculations and action evaluation constraints more consistent with current environmental changes.

[0112] Furthermore, embodiments of this application may also calculate the distance change data of the target robot relative to the target position based on the path node data before and after the update; calculate the obstacle avoidance risk change data of the target robot relative to the environmental obstacle based on the updated path node data and the position change data of the environmental obstacle; calculate the motion direction continuity evaluation data of the target robot based on the updated path node data, historical path node data, and target position data; and generate path feedback evaluation data based on the distance change data, obstacle avoidance risk change data, and motion direction continuity evaluation data.

[0113] Specifically, distance change data can indicate whether the robot is moving closer to the target location; obstacle avoidance risk change data can indicate whether the safe distance between the robot and obstacles has increased or decreased; and motion direction continuity evaluation data can indicate whether the current path node has undergone a sharp turn, reversal, or oscillation relative to historical path nodes. By combining the above data to generate path feedback evaluation data, the path benefits and path risks brought about by the target movement can be comprehensively evaluated.

[0114] Furthermore, embodiments of this application may also determine the path benefit evaluation result and path risk evaluation result corresponding to the current target movement action based on path feedback evaluation data; update the historical evaluation relationship between each candidate movement action and the preset path decision factors in the path decision context data based on the path benefit evaluation result and path risk evaluation result; and adjust the state correlation calculation result and action value correlation correction result in the next path node determination process based on the updated historical evaluation relationship.

[0115] Specifically, the path benefit evaluation results can include the degree of reduction in target distance, the degree of improvement in directional continuity, and the degree of reduction in path cost; the path risk evaluation results can include obstacle approach risk, dynamic collision risk, and path oscillation risk. By updating historical evaluation relationships, the method in this application can have closed-loop feedback capability, enabling the next round of path planning to refer to the actual effects of past actions and avoid repeatedly selecting candidate movement actions with high risk or poor effect.

[0116] S111, if the target robot has not reached the target position and the preset termination condition has not been met, the next path node is iteratively determined based on the updated path decision context data and the updated current path planning state data, so as to generate the target planning path based on the multiple path nodes determined from the current position to the target position.

[0117] In this embodiment, whether the target robot has reached the target location can be determined based on whether the distance between the robot's current location and the target location is less than a preset threshold. Preset termination conditions may include reaching the maximum number of iterations, exceeding the maximum planning time, failing to find a safe candidate action, the robot entering an emergency stop state, or the target task being canceled. If the target robot has not reached the target location and the preset termination conditions are not met, the next round of path node determination can continue.

[0118] The updated path decision context data can include the path node update results from the previous round and path feedback evaluation data. The updated current path planning state data can include the environmental state acquired by the target robot at the new path node or within the new planning cycle. By combining these two, processes such as state unit partitioning, state correlation calculation, pulse timing encoding, membrane potential update, non-pulse readout, and action evaluation can be performed to iteratively determine multiple path nodes.

[0119] Once multiple path nodes are determined, a target planned path can be generated based on the sequence of path nodes. The target planned path can be a discrete set of path nodes, or it can be formed into a continuous trajectory after trajectory continuity processing. For actual robot motion control, continuous trajectories are usually more conducive to velocity planning and attitude control, so further smoothing and collision verification can be performed after generating the initial path nodes.

[0120] Furthermore, embodiments of this application may also iteratively determine multiple path nodes based on updated path decision context data and updated current path planning state data, and generate an initial planned path according to the determination order of each path node; perform trajectory continuity processing on adjacent path nodes in the initial planned path to obtain continuous candidate trajectories; perform collision verification on the continuous candidate trajectories based on the spatial occupancy data and motion change data of environmental obstacle objects in the target working environment; and determine the continuous candidate trajectories as the target planned path if the continuous candidate trajectories meet the collision verification conditions.

[0121] Specifically, the initial planned path can be formed by connecting discrete path nodes in a predetermined order. Trajectory continuity processing can be used to eliminate abrupt changes in direction or excessive curvature between path nodes, making robot motion smoother. Collision checking can be used to determine whether continuous candidate trajectories conflict with the predicted occupied areas of static or dynamic obstacles. If the continuous candidate trajectories meet the collision checking conditions, they can be output as the target planned path; if they do not meet the collision checking conditions, path nodes can be reselected, the trajectory adjusted, or path planning can be re-executed.

[0122] Furthermore, embodiments of this application may also determine multiple path segments to be smoothed based on the positional change relationship between adjacent path nodes in the initial planned path; and perform curve smoothing processing on the initial planned path based on the connection direction and curvature continuity constraints between the multiple path segments to be smoothed to obtain a continuous candidate trajectory; wherein the continuous candidate trajectory passes through at least some of the multiple path nodes, or is located within the safe passage area corresponding to the multiple path nodes.

[0123] Specifically, curve smoothing can be achieved using spline curves, Bézier curves, polynomial curves, or other trajectory smoothing methods. Connectivity direction constraints can be used to limit abrupt changes in direction between adjacent path segments, while curvature continuity constraints can be used to reduce sharp changes during robot turning. Continuous candidate trajectories do not necessarily need to strictly pass through all discrete path nodes, as long as they are within the safe passage area corresponding to the path node and satisfy the motion constraints.

[0124] Furthermore, embodiments of this application may also perform trajectory sampling on continuous candidate trajectories according to a preset spatial sampling interval to obtain multiple trajectory sampling points; determine the predicted spatial occupancy area of ​​the environmental obstacle object at the corresponding time based on the predicted arrival time corresponding to each trajectory sampling point; calculate the safety interval between each trajectory sampling point and the corresponding predicted spatial occupancy area; and determine that the continuous candidate trajectories meet the collision verification conditions if each safety interval meets the preset safety conditions.

[0125] Specifically, the preset spatial sampling interval can be determined based on the robot's size, movement speed, and environmental complexity. For dynamic obstacles, their spatial occupancy area at different times can be predicted based on their current position, velocity direction, and motion model. If the safety interval between the trajectory sampling point and the corresponding predicted spatial occupancy area is greater than the preset safety threshold, it indicates that the continuous candidate trajectories meet the safety requirements in the prediction time dimension, and it can be determined that they meet the collision verification conditions.

[0126] Third Embodiment This embodiment combines Figure 2 This application describes the robot path planning device based on pulse membrane potential readout provided in this application. Figure 2 This is a schematic diagram of an embodiment of the robot path planning device based on pulsed membrane potential readout provided in this application. This device embodiment corresponds to the method embodiment described above; therefore, for parts not described in detail in the device embodiment, please refer to the relevant descriptions in the method embodiment.

[0127] like Figure 2 As shown, the robot path planning device based on pulse membrane potential readout provided in this application embodiment may include: The state data acquisition module 201 can be used to acquire the current path planning state data of the target robot in the target working environment. The current path planning state data is used to characterize the spatial and motion relationships between the target robot, the target location object, the environmental obstacle object, and the sensing and detection object.

[0128] Specifically, the state data acquisition module 201 can communicate with the robot localization unit, sensor components, or environmental perception components to acquire data on the robot's current position, target position, obstacle position, sensor detection results, and the relative relationships between the robot and the target and obstacles. The state data acquisition module 201 can send the current path planning state data to the context and state unit determination module 202, and can provide environmental state change-related data to the path feedback update module 210 or the target path generation module 211.

[0129] The context and state unit determination module 202 can be used to generate path decision context data based on the current path planning state data, and divide the current path planning state data into state units according to the path decision dimension to obtain multiple path state units. The path decision context data is used to record the correspondence between the current path node, historical path node, candidate movement action, environmental obstacle change status and path feedback evaluation data.

[0130] Specifically, the context and state unit determination module 202 can establish the correspondence between the current path planning state data output by the state data acquisition module 201, the current path nodes, historical path nodes, candidate movement actions, and obstacle change states, and divide the complex state data into multiple path state units. The path state units output by this module can be transmitted to the state correlation calculation module 203, and the path decision context data can be transmitted to the action readout input determination module 206, the action evaluation result determination module 208, and the target path generation module 211.

[0131] Furthermore, in this embodiment, the current path planning status data includes the robot's current position data, target position data, static obstacle position data, dynamic obstacle position data, obstacle movement direction data, multi-directional sensor detection data, relative geometric data between the robot and the target object, and relative risk data between the robot and environmental obstacles.

[0132] Specifically, the context and state unit determination module 202 can normalize, group, or store the above-mentioned different types of data in association, so that subsequent modules can read target approach information, obstacle risk information, dynamic obstacle prediction information, and sensing direction detection information respectively.

[0133] The state correlation calculation module 203 can be used to calculate the state correlation between each path state unit and the preset path decision factors, and determine the pulse coding weight and pulse triggering timing corresponding to each path state unit based on the state correlation.

[0134] Specifically, the state correlation calculation module 203 can receive multiple path state units output by the context and state unit determination module 202, and determine the degree of influence of each path state unit in path decision-making based on factors such as target approach, obstacle avoidance, continuity of movement direction, and dynamic collision risk. The state correlation calculation module 203 can output pulse coding weights and pulse triggering timing to the pulse timing coding module 204.

[0135] Furthermore, in this embodiment, the state correlation calculation module 203 is used to calculate the initial state correlation between each path state unit and the preset path decision factors based on the spatial position change data, motion trend change data and sensor detection intensity data corresponding to each path state unit; to perform context correction on the initial state correlation based on the historical path nodes and path feedback evaluation data in the path decision context data to obtain the corrected state correlation; and to determine the pulse coding weight and pulse triggering timing of each path state unit in the pulse timing coding process based on the corrected state correlation.

[0136] Specifically, the state correlation calculation module 203 can combine the path decision context data updated by the path feedback update module 210 to correct the initial state correlation, so that the pulse coding weight and pulse triggering timing can reflect the actual feedback results of the previous round of path node selection.

[0137] The pulse timing coding module 204 can be used to perform pulse timing coding on each path state unit according to the pulse coding weight and pulse triggering timing to obtain the pulse state sequence corresponding to the current path planning state data.

[0138] Specifically, the pulse timing encoding module 204 can determine the pulse trigger intensity based on the pulse encoding weight output by the state correlation calculation module 203, determine the pulse trigger time step based on the pulse trigger timing, and generate a pulse state sequence in chronological order. This pulse state sequence can be used as input to the pulse membrane potential calculation module 205.

[0139] Furthermore, in this embodiment, the pulse timing coding module 204 is used to determine the pulse trigger intensity of each path state unit within a preset simulation time window based on the pulse coding weight corresponding to each path state unit; determine the pulse trigger time step of each path state unit within a preset simulation time window based on the pulse trigger timing corresponding to each path state unit; and generate a pulse state sequence formed by arranging multiple pulse events in chronological order based on the pulse trigger intensity and the pulse trigger time step.

[0140] Specifically, the pulse timing coding module 204 can map path state units such as target approach, obstacle proximity, and dynamic obstacle movement into pulse events, and enable state units with higher importance or higher risk to generate earlier or stronger pulse inputs within the simulation time window.

[0141] The pulse membrane potential calculation module 205 can be used to input the pulse state sequence into the pulse neural computation layer, and perform membrane potential accumulation and update on the corresponding spiking neurons in the pulse neural computation layer according to the pulse events in the pulse state sequence within a preset simulation time window. The pulse output state of the spiking neurons is determined according to the updated membrane potential state. The membrane potential accumulation and update is an event-triggered state update performed on the corresponding spiking neurons when there is a pulse event input.

[0142] Specifically, the pulse membrane potential calculation module 205 can receive the pulse state sequence output by the pulse timing encoding module 204 and detect whether there is a pulse event input to the target spiking neuron at each time step. If it exists, it performs membrane potential accumulation update; if it does not exist, it maintains the membrane potential state or performs decay processing. The pulse membrane potential calculation module 205 can transmit the membrane potential state and pulse output state to the action readout input determination module 206.

[0143] Furthermore, in this embodiment, the pulse membrane potential calculation module 205 is used to detect whether there is a pulse event input to the target spiking neuron in the pulse state sequence at the current time step of the preset simulation time window; if there is a pulse event input to the target spiking neuron, the target spiking neuron is updated by accumulating the membrane potential based on the pulse weight corresponding to the pulse event and the membrane potential state of the target spiking neuron in the previous time step; if there is no pulse event input to the target spiking neuron, the membrane potential state of the target spiking neuron in the previous time step is maintained, or the state of the target spiking neuron is updated based on the preset membrane potential decay rule.

[0144] Specifically, through the event-triggered update method described above, the pulse membrane potential calculation module 205 can reduce invalid calculations in the absence of pulse input, thereby reducing the computational load and energy consumption of the robot body.

[0145] Furthermore, in this embodiment, the pulse membrane potential calculation module 205 is used to compare the updated membrane potential state with a preset membrane potential threshold; if the updated membrane potential state meets the pulse triggering condition, the pulse output state corresponding to the spiking neuron is generated, and the membrane potential state of the spiking neuron is reset; if the updated membrane potential state does not meet the pulse triggering condition, the output pulse event of the spiking neuron is prohibited.

[0146] Specifically, the pulse membrane potential calculation module 205 can set different membrane potential thresholds according to different path decision factors to achieve differentiated responses to states such as safety risks, target approach and motion continuity.

[0147] The action readout input determination module 206 can be used to determine the action readout input data corresponding to multiple candidate movement actions based on the membrane potential state, pulse output state and path decision context data of the spiking neuron.

[0148] Specifically, the action readout input determination module 206 can receive the membrane potential state and pulse output state output by the pulse membrane potential calculation module 205, and combine them with the path decision context data provided by the context and state unit determination module 202 or the path feedback update module 210 to determine the action readout input data corresponding to each candidate movement action. The action readout input data can be transmitted to the non-pulse membrane potential readout module 207.

[0149] Furthermore, in this embodiment, the action readout input determination module 206 is used to determine multiple candidate movement actions and corresponding action constraints based on path decision context data; calculate the action response correlation between the membrane potential state and pulse output state of the spiking neuron and each candidate movement action; and determine the action readout input data corresponding to each candidate movement action from the membrane potential state and pulse output state of the spiking neuron based on the action response correlation and action constraints.

[0150] Specifically, the action readout input determination module 206 can select the relevant spiking neuron state as input based on the direction, step size and safety constraints of the candidate movement action, thereby avoiding interference from irrelevant states on the evaluation of the candidate action.

[0151] The non-pulse membrane potential readout module 207 can be used to input the readout input data of each action into the non-pulse readout neurons corresponding to multiple candidate movement actions, and calculate the membrane potential change data of each non-pulse readout neuron within a preset simulation time window.

[0152] Specifically, the non-pulse membrane potential readout module 207 can set up a corresponding non-pulse readout neuron for each candidate movement action, and express the evaluation response of the candidate action through non-pulse membrane potential change data. The membrane potential change data output by the non-pulse membrane potential readout module 207 can be transmitted to the action evaluation result determination module 208.

[0153] Furthermore, in this embodiment, the non-pulse membrane potential readout module 207 is used to obtain the membrane potential values ​​of each non-pulse readout neuron at multiple time steps within a preset simulation time window; based on the maximum membrane potential of each non-pulse readout neuron within the preset simulation time window, determine the action value representation data corresponding to each candidate movement action; and calculate the action value correlation between each candidate movement action and the preset path decision factors based on the action value representation data corresponding to each candidate movement action.

[0154] Specifically, by reading the membrane potential change data of non-pulse readout neurons, it is possible to avoid the fluctuations in action evaluation caused by relying solely on the number of pulse firings, thereby improving the stability of candidate movement action evaluation.

[0155] The action evaluation result determination module 208 can be used to calculate the action value correlation between each candidate movement action and the preset path decision factors based on the membrane potential change data of each non-pulse readout neuron, and to constrain and correct the action value correlation according to the path decision context data to obtain the path action evaluation result corresponding to each candidate movement action.

[0156] Specifically, the action evaluation result determination module 208 can receive the membrane potential change data output by the non-pulse membrane potential readout module 207, and perform action value correlation correction in conjunction with path decision context data. The corrected path action evaluation result can be transmitted to the path node update module 209 to determine the target movement action.

[0157] Furthermore, in this embodiment, the action evaluation result determination module 208 is used to determine the target approach constraint, obstacle avoidance constraint, motion direction continuity constraint, and dynamic collision risk constraint corresponding to each candidate movement action based on the path decision context data; and to perform weighted correction on the action value correlation corresponding to each candidate movement action based on the target approach constraint, obstacle avoidance constraint, motion direction continuity constraint, and dynamic collision risk constraint; and to obtain the path action evaluation result corresponding to each candidate movement action based on the weighted correction of the action value correlation.

[0158] Specifically, the motion evaluation result determination module 208 can improve the quality of path planning by weighting the final motion evaluation result to take into account target approach efficiency, obstacle avoidance safety, path smoothness and dynamic collision risk.

[0159] The path node update module 209 can be used to determine the target movement action from multiple candidate movement actions based on the path movement evaluation results corresponding to each candidate movement action, and update the path node data of the target robot according to the target movement action.

[0160] Specifically, the path node update module 209 can receive the path action evaluation results output by the action evaluation result determination module 208, select candidate movement actions that meet the action selection conditions as target movement actions, and calculate the next path node based on the direction and step size of the target movement action. The path node update module 209 can send the updated path node data to the path feedback update module 210 and the target path generation module 211.

[0161] Furthermore, in this embodiment, the path node update module 209 is used to sort multiple candidate movement actions according to the path action evaluation results corresponding to each candidate movement action; determine the candidate movement action whose sorting result meets the preset action selection conditions as the target movement action; determine the next path node after the target robot moves from the current path node based on the movement direction and movement step size corresponding to the target movement action; and write the next path node into the path node sequence in the path decision context data to update the path node data of the target robot.

[0162] Specifically, the path node update module 209 can determine the target movement action by sorting and safety screening, and make the path node sequence continuously record the generation process of the robot's planned path.

[0163] The path feedback update module 210 can be used to generate path feedback evaluation data corresponding to the updated path nodes based on the updated path node data and the environmental state change data in the target working environment, and update the path decision context data based on the path feedback evaluation data.

[0164] Specifically, the path feedback update module 210 can receive path node data output by the path node update module 209, and combine it with environmental state change data provided by the state data acquisition module 201 or other sensing data sources to calculate path feedback evaluation data. The path feedback update module 210 can feed back the updated path decision context data to the context and state unit determination module 202, the action readout input determination module 206, the action evaluation result determination module 208, and the target path generation module 211.

[0165] Furthermore, in this embodiment, the path feedback update module 210 is used to calculate the distance change data of the target robot relative to the target location object based on the path node data before and after the update; calculate the obstacle avoidance risk change data of the target robot relative to the environmental obstacle object based on the updated path node data and the location change data of the environmental obstacle object; calculate the motion direction continuity evaluation data of the target robot based on the updated path node data, historical path node data and the target location object; and generate path feedback evaluation data based on the distance change data, obstacle avoidance risk change data and motion direction continuity evaluation data.

[0166] Specifically, the path feedback update module 210 can reflect the path benefits and path risks of the target movement through feedback evaluation data, so that the next round of path decision-making can make use of the effects of historical decisions.

[0167] The target path generation module 211 can be used to iteratively determine the next path node based on the updated path decision context data and the updated current path planning state data when the target robot has not reached the target position and the preset termination condition has not been met, so as to generate a target planning path based on multiple path nodes determined from the current position to the target position.

[0168] Specifically, the target path generation module 211 can receive data transmitted by the status data acquisition module 201, the context and status unit determination module 202, the path node update module 209, and the path feedback update module 210. If the target location is not reached and the termination condition is not met, the target path generation module 211 can trigger the next round of path node determination; if the target arrival condition or termination condition is met, the target path generation module 211 can generate a target planned path based on the determined path nodes.

[0169] Furthermore, in this embodiment, the target path generation module 211 is used to iteratively determine multiple path nodes based on the updated path decision context data and the updated current path planning state data, and generate an initial planned path according to the determination order of each path node; perform trajectory continuity processing on adjacent path nodes in the initial planned path to obtain continuous candidate trajectories; perform collision verification on the continuous candidate trajectories based on the spatial occupancy data and motion change data of environmental obstacle objects in the target working environment; and determine the continuous candidate trajectories as the target planned path if the continuous candidate trajectories meet the collision verification conditions.

[0170] Specifically, the target path generation module 211 can smooth the path node sequence and perform collision verification by combining the space occupancy data of static and dynamic obstacles, thereby outputting a target planning path that can be used for robot motion control.

[0171] Furthermore, in this embodiment of the application, the target path generation module 211 is used to determine multiple path segments to be smoothed based on the positional change relationship between adjacent path nodes in the initial planned path; and to perform curve smoothing processing on the initial planned path based on the connection direction and curvature continuity constraints between the multiple path segments to be smoothed to obtain a continuous candidate trajectory; wherein the continuous candidate trajectory passes through at least some of the multiple path nodes, or is located within the safe passage area corresponding to the multiple path nodes.

[0172] Specifically, the target path generation module 211 can use curve smoothing to reduce abrupt changes in direction between path nodes, thereby improving the stability of the robot during actual operation.

[0173] Furthermore, in this embodiment of the application, the target path generation module 211 is used to sample the continuous candidate trajectories according to a preset spatial sampling interval to obtain multiple trajectory sampling points; determine the predicted spatial occupancy area of ​​the environmental obstacle object at the corresponding time according to the predicted arrival time corresponding to each trajectory sampling point; calculate the safety interval between each trajectory sampling point and the corresponding predicted spatial occupancy area; and determine that the continuous candidate trajectories meet the collision verification conditions when each safety interval meets the preset safety conditions.

[0174] Specifically, the target path generation module 211 can perform safety verification on continuous candidate trajectories in both spatial and temporal dimensions, making it particularly suitable for work environments with dynamic obstacles.

[0175] Fourth embodiment The above describes the internal functions and structure of a robot path planning device based on pulsed membrane potential readout, which can be implemented as an electronic device. Figure 3 A schematic diagram illustrating the structure of an embodiment of the electronic device provided in this application. (See attached diagram.) Figure 3 As shown, the electronic device includes a memory 31 and a processor 32.

[0176] Memory 31 is used to store programs. In addition to the programs described above, memory 31 can also be configured to store various other data to support operation on the electronic device. Examples of this data include instructions for any application or method used to operate on the electronic device, contact data, phonebook data, messages, pictures, videos, etc.

[0177] The memory 31 can be implemented by any type of volatile or non-volatile storage device or a combination thereof, such as static random access memory (SRAM), electrically erasable programmable read-only memory (EEPROM), erasable programmable read-only memory (EPROM), programmable read-only memory (PROM), read-only memory (ROM), magnetic storage, flash memory, magnetic disk or optical disk.

[0178] Processor 32 is not limited to a central processing unit (CPU), but may also be a graphics processing unit (GPU), a field-programmable gate array (FPGA), an embedded neural network processor (NPU), or an artificial intelligence (AI) chip. Processor 32 is coupled to memory 31 and executes the program stored in memory 31. When the program runs, it executes the robot path planning method based on pulse membrane potential readout described above.

[0179] Furthermore, such as Figure 3 As shown, the electronic device may also include other components such as a communication component 33, a power supply component 34, an audio component 35, and a display 36. Figure 3 The diagram only shows some components and does not mean that the electronic device includes only these components. Figure 3 The components shown.

[0180] Communication component 33 is configured to facilitate wired or wireless communication between electronic devices and other devices. The electronic devices can access wireless networks based on communication standards, such as WiFi, 3G, 4G, or 5G, or combinations thereof. In one exemplary embodiment, communication component 33 receives broadcast signals or broadcast-related information from an external broadcast management system via a broadcast channel. In one exemplary embodiment, communication component 33 also includes a near-field communication (NFC) module to facilitate short-range communication. For example, the NFC module may be implemented based on radio frequency identification (RFID) technology, Infrared Data Association (IrDA) technology, ultra-wideband (UWB) technology, Bluetooth (BT) technology, and other technologies.

[0181] Power supply component 34 provides power to various components of the electronic device. Power supply component 34 may include a power management system, one or more power supplies, and other components associated with generating, managing, and distributing power to the electronic device.

[0182] Audio component 35 is configured to output and / or input audio signals. For example, audio component 35 includes a microphone (MIC) configured to receive external audio signals when the electronic device is in an operating mode, such as call mode, recording mode, and voice recognition mode. The received audio signals may be further stored in memory 31 or transmitted via communication component 33. In some embodiments, audio component 35 also includes a speaker for outputting audio signals.

[0183] Display 36 includes a screen, which may include a liquid crystal display (LCD) and a touch panel (TP). If the screen includes a touch panel, the screen can be implemented as a touchscreen to receive input signals from a user. The touch panel includes one or more touch sensors to sense touches, swipes, and gestures on the touch panel. The touch sensors can detect not only the boundaries of touch or swipe actions, but also the duration and pressure associated with the touch or swipe operation.

[0184] Those skilled in the art will understand that all or part of the steps of the above-described method embodiments can be implemented by hardware related to program instructions. The aforementioned program can be stored in a computer-readable storage medium. When executed, the program performs the steps of the above-described method embodiments; and the aforementioned storage medium includes various media capable of storing program code, such as ROM, RAM, magnetic disks, or optical disks.

[0185] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of this application, and are not intended to limit them. Although this application 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 embodiments of this application.

Claims

1. A robot path planning method based on pulsed membrane potential readout, characterized in that, include: The current path planning state data of the target robot in the target working environment is obtained. The current path planning state data is used to characterize the spatial and motion relationships between the target robot, the target location object, the environmental obstacle object, and the sensing and detection object. Path decision context data is generated based on the current path planning state data, and the current path planning state data is divided into state units according to the path decision dimension to obtain multiple path state units. The path decision context data is used to record the correspondence between the current path node, historical path node, candidate movement action, environmental obstacle change status, and path feedback evaluation data. Calculate the state correlation between each path state unit and the preset path decision factors, and determine the pulse coding weight and pulse triggering timing corresponding to each path state unit based on the state correlation; Based on the pulse coding weight and the pulse triggering timing, pulse timing coding is performed on each path state unit to obtain a pulse state sequence corresponding to the current path planning state data; The pulse state sequence is input into the spiking neural computation layer. Within a preset simulation time window, the membrane potential of the corresponding spiking neuron in the spiking neural computation layer is cumulatively updated according to the pulse events in the pulse state sequence. The pulse output state of the spiking neuron is determined according to the updated membrane potential state. The membrane potential cumulative update is an event-triggered state update performed on the corresponding spiking neuron when a pulse event is input. Based on the membrane potential state and pulse output state of the spiking neuron and the path decision context data, determine the action readout input data corresponding to multiple candidate movement actions respectively; The action readout input data is respectively input into the non-pulse readout neurons corresponding to the plurality of candidate movement actions, and the membrane potential change data of each non-pulse readout neuron within the preset simulation time window is calculated. Based on the membrane potential change data of each of the non-pulse readout neurons, the action value correlation between each of the candidate movement actions and the preset path decision factors is calculated, and the action value correlation is constrained and corrected according to the path decision context data to obtain the path action evaluation result corresponding to each of the candidate movement actions. Based on the path action evaluation results corresponding to each of the candidate movement actions, a target movement action is determined from the multiple candidate movement actions, and the path node data of the target robot is updated according to the target movement action; Based on the updated path node data and the environmental state change data in the target work environment, path feedback evaluation data corresponding to the updated path nodes is generated, and the path decision context data is updated based on the path feedback evaluation data. If the target robot fails to reach the target position and the preset termination condition is not met, the next path node is iteratively determined based on the updated path decision context data and the updated current path planning state data, so as to generate a target planning path based on multiple path nodes determined from the current position to the target position.

2. The method according to claim 1, characterized in that, The process of iteratively determining the next path node based on updated path decision context data and updated current path planning state data, to generate a target planned path based on multiple path nodes determined from the current position to the target position, includes: Based on the updated path decision context data and the updated current path planning state data, multiple path nodes are iteratively determined, and an initial planned path is generated according to the determination order of each path node. Trajectory continuity processing is performed on adjacent path nodes in the initial planned path to obtain continuous candidate trajectories; Based on the spatial occupancy data and motion change data of environmental obstacles in the target work environment, collision verification is performed on the continuous candidate trajectories. If the continuous candidate trajectories meet the collision verification conditions, the continuous candidate trajectories are determined as the target planning path.

3. The method according to claim 1, characterized in that, The step of calculating the state correlation between each path state unit and the preset path decision factors, and determining the pulse coding weight and pulse triggering timing corresponding to each path state unit based on the state correlation, includes: Based on the spatial position change data, motion trend change data and sensor detection intensity data corresponding to each path state unit, the initial state correlation between each path state unit and the preset path decision factor is calculated. Based on the historical path nodes and path feedback evaluation data in the path decision context data, the initial state correlation is modified in context to obtain the modified state correlation. The pulse coding weight and pulse triggering timing of each path state unit in the pulse timing coding process are determined based on the corrected state correlation.

4. The method according to claim 1, characterized in that, The step of performing pulse timing encoding on each path state unit according to the pulse encoding weight and the pulse triggering timing to obtain a pulse state sequence corresponding to the current path planning state data includes: Based on the pulse coding weights corresponding to each path state unit, the pulse triggering intensity of each path state unit within the preset simulation time window is determined. Based on the pulse trigger timing corresponding to each path state unit, determine the pulse trigger time step of each path state unit within the preset simulation time window; Based on the pulse trigger intensity and the pulse trigger time step, a pulse state sequence is generated, which is formed by arranging multiple pulse events in chronological order.

5. The method according to claim 1, characterized in that, The step of accumulating and updating the membrane potential of the corresponding spiking neurons in the spiking neural computation layer based on the spiking events in the spiking state sequence within a preset simulation time window includes: At the current time step of the preset simulation time window, detect whether there is a pulse event input to the target spiking neuron in the pulse state sequence; In the presence of a pulse event input to the target spiking neuron, the membrane potential of the target spiking neuron is cumulatively updated based on the pulse weight corresponding to the pulse event and the membrane potential state of the target spiking neuron in the previous time step. In the absence of a pulse event input to the target spiking neuron, the membrane potential state of the target spiking neuron at the previous time step is maintained, or the state of the target spiking neuron is updated based on a preset membrane potential decay rule.

6. The method according to claim 1, characterized in that, Determining the pulse output state of the spiking neuron based on the updated membrane potential state includes: The updated membrane potential state is compared with the preset membrane potential threshold. If the updated membrane potential state meets the pulse triggering condition, the pulse output state corresponding to the spiking neuron is generated, and the membrane potential state of the spiking neuron is reset. If the updated membrane potential state does not meet the pulse triggering condition, the spiking neuron is prohibited from outputting pulse events.

7. The method according to claim 1, characterized in that, The determination of action readout input data corresponding to multiple candidate movement actions based on the membrane potential state, pulse output state, and path decision context data of the spiking neuron includes: Based on the path decision context data, multiple candidate movement actions and corresponding action constraints for each candidate movement action are determined. Calculate the correlation between the membrane potential state and pulse output state of the spiking neuron and the action response of each candidate movement action; Based on the action response correlation and the action constraint, action readout input data corresponding to each of the candidate movement actions are determined from the membrane potential state and pulse output state of the spiking neuron.

8. The method according to claim 1, characterized in that, The step of constraining and correcting the correlation of action value based on the path decision context data to obtain the path action evaluation result corresponding to each candidate movement action includes: Based on the path decision context data, determine the target approach constraint, obstacle avoidance constraint, motion direction continuity constraint, and dynamic collision risk constraint corresponding to each candidate movement action; Based on the target approach constraint, the obstacle avoidance constraint, the motion direction continuity constraint, and the dynamic collision risk constraint, the action value correlation corresponding to each of the candidate movement actions is weighted and corrected. Based on the weighted and corrected action value correlation, the path action evaluation results corresponding to each candidate movement action are obtained.

9. A robot path planning device based on pulsed membrane potential readout, characterized in that, include: The status data acquisition module is used to acquire the current path planning status data of the target robot in the target working environment. The current path planning status data is used to characterize the spatial and motion relationships between the target robot, the target location object, the environmental obstacle object, and the sensing and detection object. The context and state unit determination module is used to generate path decision context data based on the current path planning state data, and divide the current path planning state data into state units according to the path decision dimension to obtain multiple path state units. The path decision context data is used to record the correspondence between the current path node, historical path node, candidate movement action, environmental obstacle change state, and path feedback evaluation data. The state correlation calculation module is used to calculate the state correlation between each path state unit and the preset path decision factors, and to determine the pulse coding weight and pulse triggering timing corresponding to each path state unit based on the state correlation. The pulse timing coding module is used to perform pulse timing coding on each of the path state units according to the pulse coding weight and the pulse triggering timing to obtain a pulse state sequence corresponding to the current path planning state data; The pulse membrane potential calculation module is used to input the pulse state sequence into the pulse neural computation layer, perform cumulative update of the membrane potential of the corresponding spiking neurons in the pulse neural computation layer according to the pulse events in the pulse state sequence within a preset simulation time window, and determine the pulse output state of the spiking neurons according to the updated membrane potential state. The cumulative update of the membrane potential is an event-triggered state update performed on the corresponding spiking neurons when there is a pulse event input. The action readout input determination module is used to determine the action readout input data corresponding to multiple candidate movement actions based on the membrane potential state, pulse output state of the spiking neuron and the path decision context data. The non-pulse membrane potential readout module is used to input the readout input data of each of the aforementioned actions into the non-pulse readout neurons corresponding to the plurality of candidate movement actions, and to calculate the membrane potential change data of each of the aforementioned non-pulse readout neurons within the preset simulation time window; The action evaluation result determination module is used to calculate the action value correlation between each candidate movement action and the preset path decision factor based on the membrane potential change data of each of the non-pulse readout neurons, and to constrain and correct the action value correlation according to the path decision context data to obtain the path action evaluation result corresponding to each candidate movement action. The path node update module is used to determine the target movement action from multiple candidate movement actions based on the path movement evaluation results corresponding to each candidate movement action, and update the path node data of the target robot according to the target movement action; The path feedback update module is used to generate path feedback evaluation data corresponding to the updated path nodes based on the updated path node data and the environmental state change data in the target work environment, and update the path decision context data based on the path feedback evaluation data. The target path generation module is used to iteratively determine the next path node based on the updated path decision context data and the updated current path planning state data when the target robot has not reached the target position and has not met the preset termination condition, so as to generate a target planning path based on multiple path nodes determined from the current position to the target position.

10. An electronic device, characterized in that, include: Memory, used to store programs; A processor is configured to run the program stored in the memory to perform the robot path planning method based on pulse membrane potential readout as described in any one of claims 1 to 8.