A method for socially navigating a mobile robot

Abstract

The application discloses a kind of mobile robot social navigation method, comprising the following steps: global path planning is carried out based on static environment map, and global reference path is generated;When detecting dynamic pedestrian, start local social navigation, and local social navigation includes: learning and extracting social cost function from human demonstration behavior by inverse reinforcement learning;Using model predictive control method, according to social cost function, robot dynamics constraint and environment constraint, local control trajectory is generated by rolling optimization in limited prediction time domain;Through reinforcement learning regulator, the control cost weight in model predictive control is dynamically adjusted according to real-time environment state, to realize the balance of safety, sociality and navigation efficiency.

Description

Technical Field

[0001] This invention relates to the field of mobile robot technology, and more specifically to a social navigation method for mobile robots. Background Technology

[0002] Currently, social navigation methods for mobile robots in human-robot collaborative environments mainly fall into two categories. One category relies on traditional optimization methods such as Model Predictive Control (MPC), which depend on manually designed cost functions and fixed parameters. This requires extensive domain knowledge to characterize social norms, and the models have poor generalization ability, making them ill-suited for complex interactions in dynamic crowds. The other category directly applies end-to-end deep reinforcement learning (RL) models to navigation tasks. However, these methods have inherent flaws when facing real-world human-robot collaborative environments: First, unlike structured environments with explicit physical constraints, social navigation requires understanding implicit rules of human behavior. The fixed parameters of traditional MPC struggle to capture such contextual semantics, leading to rigid robot behavior that does not conform to social expectations. Second, interactions conforming to social norms account for a very small percentage of navigation decisions, creating a sparse reward problem. This results in low training efficiency for reinforcement learning models, difficulty in policy convergence, and a lack of safety constraints in the learned policies. Summary of the Invention

[0003] Purpose of the Invention: The purpose of this invention is to provide a social navigation method for mobile robots. Based on a static environment map, a global reference path is generated using a global path planning algorithm. When the robot detects a dynamic pedestrian, it initiates local social navigation based on the fusion of MPC and RL. This framework extracts the social cost function from human demonstrations through inverse reinforcement learning, constructs a social behavior model, and provides predictive control for the model. A reinforcement learning regulator dynamically adjusts the MPC parameters according to the real-time environment to achieve a balance between safety and sociality. Finally, MPC optimization generates a local trajectory that conforms to social norms, thus solving the problems existing in the background technology.

[0004] Technical solution: The social navigation method for mobile robots described in this invention includes the following steps:

[0005] (1) Perform global path planning based on the static environment map to generate a global reference path;

[0006] (2) When a dynamic pedestrian is detected, local social navigation is initiated, which includes: learning and extracting the social cost function from human demonstration behavior through inverse reinforcement learning; and using model predictive control method to integrate the social cost function, robot dynamics constraints and environmental constraints to generate local control trajectory through rolling optimization within the finite prediction time domain.

[0007] (3) By using a reinforcement learning regulator, the weights of various control costs in the model predictive control method are dynamically adjusted according to the real-time environmental state.

[0008] Furthermore, in step (2), the social cost function is learned from a dataset of exemplary trajectories containing human social interaction behaviors through an inverse reinforcement learning algorithm, and is used to introduce social behavior constraints in trajectory optimization.

[0009] Furthermore, in step (2), the model predictive control method simultaneously considers trajectory tracking accuracy, control input smoothness, and social cost constraints extracted by inverse reinforcement learning during optimization, and solves the optimal control sequence and executes the first control command in each control cycle.

[0010] Furthermore, in step (3), the reinforcement learning regulator generates control parameter adjustment amounts based on the real-time acquired environmental state information and the near-end policy optimization algorithm, and dynamically updates the weights corresponding to trajectory tracking error, control smoothness and social constraints in the objective function of the model predictive control method optimization.

[0011] Furthermore, the reinforcement learning regulator is trained offline using a multi-dimensional reward function that includes goal advancement rewards, collision penalties, trajectory smoothing rewards, and social interaction penalties. The trained policy network is then subjected to forward inference during the online deployment phase to adjust parameters online.

[0012] Furthermore, the method also includes: acquiring environmental information through multi-sensor fusion to construct a global state vector containing the states of the robot and pedestrians; predicting the future trajectory of pedestrians based on a time-series prediction model, and using the prediction results for the decision-making process of local social navigation.

[0013] The social navigation system for mobile robots described in this invention includes:

[0014] Global Reference Path Module: Used for global path planning based on a static environment map, generating global reference paths;

[0015] Local social navigation module: When a dynamic pedestrian is detected, local social navigation is initiated. Local social navigation includes: learning and extracting the social cost function from human demonstration behavior through inverse reinforcement learning; and using model predictive control method to generate a local control trajectory by rolling optimization within a finite prediction time domain based on the social cost function, robot dynamics constraints, and environmental constraints.

[0016] Adjustment module: Used to dynamically adjust the weights of each control cost in model predictive control based on the real-time environmental state through a reinforcement learning regulator, so as to achieve a balance between safety, sociality and navigation efficiency.

[0017] An electronic device according to the present invention includes a memory and a processor, wherein the memory stores a computer program, and the processor executes the program to perform any of the steps of the method described herein.

[0018] The present invention provides a computer-readable storage medium storing a computer program, wherein the program, when executed by a processor, comprises the steps of any of the methods described herein.

[0019] Beneficial effects: Compared with the prior art, the present invention has the following advantages: By limiting reinforcement learning to only adjust the model predictive control parameters instead of directly generating control commands, and combining it with the social cost function extracted by inverse reinforcement learning, the present invention constructs a safe, stable and socially compliant mobile robot navigation method. The advantages are as follows: (1) By extracting the social cost function from human demonstration behavior through the inverse reinforcement learning module and using it as the constraint term of the model predictive control, the robot is guided to generate navigation behavior that conforms to social norms in a dynamic crowd environment, which improves the naturalness and safety of human-computer interaction; (2) By adjusting the weight parameters in the model predictive control objective function online through the reinforcement learning adjustment module, the instability problem caused by the direct control of the robot by reinforcement learning is avoided, which improves the safety and controllability of the system; (3) By generating control commands uniformly through the model predictive control module, the trajectory accuracy, motion smoothness and social constraints are comprehensively considered while satisfying the dynamic constraints, and stable navigation in complex dynamic environments is achieved; (4) By adopting the strategy of "offline learning + online reasoning", the system does not need to perform online learning during the deployment stage, which reduces the computational complexity and improves the real-time performance and engineering feasibility. Attached Figure Description

[0020] Figure 1 This is a flowchart of the present invention;

[0021] Figure 2 This is a flowchart of the social cost function learning process based on inverse reinforcement learning of the present invention;

[0022] Figure 3 This is a flowchart of the strategy optimization and parameter adjustment based on PPO of the present invention. Detailed Implementation

[0023] The technical solution of the present invention will be further described below with reference to the accompanying drawings.

[0024] like Figure 1-3As shown in the figure, this invention provides a social navigation method for mobile robots, including the following steps: The figure illustrates the robot social navigation process based on the combination of IRL, PPO, and MPC. First, the system uses sensors (such as RGB-D cameras, LiDAR, IMU, etc.) to perceive the surrounding environment and dynamic pedestrians in real time, and uses multi-sensor fusion to form a unified state expression for subsequent learning and control module sharing; at the same time, it combines target point information and environmental interaction features to construct an environmental state vector (e.g., including target point distance). pedestrian density Relative velocity social distance Features such as these provide input for online decision-making. In the offline phase, the IRL module learns the social cost function based on human demonstration behavior data. The cost function remains fixed during the online deployment phase and is only provided to the MPC module as a social behavior constraint; the PPO module completes policy training in the offline simulation environment to obtain the policy network. When deployed online, only the forward inference output parameter adjustment is performed. This is used for online adjustment of trajectory error weights, control increment weights, and social cost weights in MPC. During online operation, the system first calculates reference paths and generates reference states on a static map using global path planning algorithms (such as A* and Dijkstra's algorithm). Subsequently, within each control cycle, a closed-loop iteration of "environmental perception and prediction → MPC rolling optimization → execution control → environmental feedback" is executed: when no dynamic pedestrian is detected, MPC performs standard trajectory tracking and smooth control; when a dynamic pedestrian is detected, MPC incorporates offline IRL learning into the rolling optimization. And combined with the policy network output obtained from offline PPO training The MPC weights are adjusted online to form an adaptive control objective function. When a dynamic pedestrian is detected, the optimization objective of MPC can be expressed as: This optimization objective ensures that the system, while performing tasks, not only meets the requirements of trajectory accuracy and motion smoothness but also adheres to social constraints. This invention uses real-time feedback for dynamic adjustments, ensuring the robot smoothly and safely reaches the target point in dynamic crowd environments. Inputs include environmental data, reference path, social cost function, and control parameters adjusted by PPO. Outputs are control commands generated by the MPC module and dynamically adjusted social constraints. The output is the control command (linear velocity) generated by the MPC module. With angular velocity or the corresponding optimal control sequence (and their corresponding trajectories).

[0025] Step 1: Design environmental perception methods and state variables

[0026] This step is the design of the perception and state interface for "online closed loop": determining the sensor, fusion method, global state vector and pedestrian prediction output, so that the subsequent offline IRL (step 2), offline PPO (step 3) and online MPC (step 4) can all use the same set of state representation and data interface.

[0027] The robot senses and predicts pedestrians, obstacles, and dynamic environmental features around it, utilizing multi-sensor data (including RGB-D cameras, LiDAR, and IMU) to perceive the environment and estimating future pedestrian trajectories through a temporal prediction model, thereby providing decision support for subsequent social navigation. This includes the following steps:

[0028] RGB-D cameras (such as RealSense, Kinect, and Azure Kinect) are used to acquire environmental images and depth information. Depth images provide the distance between each pedestrian and the robot, which helps to construct a depth distribution map and spatial layout for each pedestrian. LiDAR is used to acquire obstacle information from the surrounding environment, helping to identify dynamic and static obstacles and enhancing environmental perception, especially pedestrian perception. IMUs provide dynamic information for the robot, such as acceleration and angular velocity, for precise control and path planning, ensuring the robot performs tasks stably.

[0029] Constructing the global state vector: A global state vector is constructed through multi-sensor fusion to represent the real-time state of the robot and all pedestrians, including position and velocity. The specific formula is as follows:

[0030]

[0031] in: , These represent the robot at time [time]. pose and velocity; , Indicates the first The position and speed of an individual pedestrian;

[0032] This represents the number of dynamic pedestrians in the scene.

[0033] Pedestrian trajectory prediction: This feature uses time-series prediction models (such as LSTM, GRU, or Transformer) to predict the future location of pedestrians based on historical data. By analyzing pedestrian movement patterns, the model can estimate the pedestrian's trajectory within the next few seconds. The prediction result is the future trajectory of each pedestrian, including their position and speed information. The system can automatically trigger appropriate social navigation strategies when dynamic pedestrians are detected.

[0034] Depth and Location Fusion: Backend Fusion Methods: Pedestrian detection and region of interest (RoI) extraction are performed using image segmentation algorithms (such as YOLOv8). The mean or median depth values ​​of the corresponding segmentation mask are then calculated for distance estimation and filtering. Network-Level Fusion Methods: RGB and depth images are used as additional input channels. RGB and depth features are extracted separately through two branches, and finally fused together for joint detection and localization. This method better utilizes the complementarity of RGB and depth information, improving detection accuracy.

[0035] Output:

[0036] Global State Output: The module outputs a global state vector. It contains the robot's status and the real-time status of all pedestrians.

[0037] Trajectory prediction output: Predicts the future trajectory of each pedestrian, typically including their position and velocity information for the next 1-5 seconds. This output provides the necessary decision input for subsequent social navigation.

[0038] Step 2: Learning the social cost function based on inverse reinforcement learning (performed offline):

[0039] This step uses a human demonstration dataset offline to learn social parameters through a maximum likelihood / trajectory probability model, obtaining social weights and a social cost function. In the online deployment phase, no further learning is performed; only the parameters are... As input to the social behavior constraints of MPC.

[0040] like Figure 2 As shown, the implicit cost structure of social navigation is learned from human demonstration behavior through inverse reinforcement learning (IRL), and this cost function is provided to the model predictive control (MPC) module to provide social behavioral constraints for navigation decisions. Input data and learning objectives:

[0041] Input dataset: The module receives a human navigation demonstration dataset:

[0042]

[0043] in, These are human demonstration trajectories, including typical social interaction behaviors such as avoidance, intersection, and yielding. From time Step to The state sequence. State vector. This typically includes information such as the location and speed of robots and pedestrians. It is the corresponding action sequence, that is, at each time step In this context, it refers to the actions taken by the robot. For example, the robot might choose to move in a certain direction or perform an avoidance maneuver.

[0044] Learning objective: To learn the optimal cost parameters by maximizing the probability of expert behavior occurring. The goal is to find a set of parameters This maximizes the probability distribution of the demonstration trajectory. The specific formula is:

[0045]

[0046] Among them, trajectory The probability distribution is defined as:

[0047]

[0048] in, : Represents trajectory The social cost function; : Reflects the quality of the trajectory (the lower the cost, the higher the probability). The denominator is a normalization constant to ensure the effectiveness of the probability distribution.

[0049] Through the above process, the IRL module learns... This refers to the social cost function implicit in human demonstration behavior. The social cost function is composed of the following: Based on the learning results of the IRL module, the social cost function can be expressed as:

[0050]

[0051] in: This represents the distance cost between the robot and pedestrians, and is used to ensure that the robot and pedestrians maintain an appropriate safe distance. : Represents the direction consistency cost, used to constrain the coordination between the robot and the pedestrian's movement direction; : Represents the speed matching cost, used to ensure that the speed difference between robots and pedestrians remains within a socially acceptable range; This indicates the cost of avoidance behavior. This item guides the robot to make reasonable interaction sequences, such as avoidance.

[0052] parameter , , , The optimal weight coefficients obtained from IRL learning reflect the importance of each social element in the overall cost. Output results and interface: The output includes a social weight matrix and a social cost function, defined as follows:

[0053]

[0054] The final output is: social weight matrix : Used to characterize the importance of different social interaction elements; social cost function : Social constraints used to characterize navigation behavior. During the system deployment phase, the IRL module outputs... With the cost function fixed, it simplifies to:

[0055]

[0056] This function is passed to the Model Predictive Control (MPC) module via a parameter interface for subsequent online invocation and dynamic optimization of social behavior constraints. During system deployment, the inverse reinforcement learning module ceases online learning; its output social cost function parameters remain fixed and are used as social behavior constraint terms for the MPC module to invoke.

[0057] Step 3: PPO Training (Offline):

[0058] The PPO policy network is trained in an offline simulation environment through a process of "policy network—environment interaction sampling—multi-dimensional reward—pruning optimization—convergence check." During online deployment, only forward inference output is performed. The weights will no longer be updated.

[0059] like Figure 3 As shown, the PPO module is based on the MPC rolling optimization framework and predicts the step size. The control weights are adaptively adjusted online, enabling real-time correction and stable optimization of control parameters in dynamic environments. A proximal policy optimization (PPO) algorithm is employed to dynamically adjust MPC control parameters based on environmental changes during robot operation, achieving adaptive social navigation through a two-layer "policy-control" optimization.

[0060] Input and output definitions:

[0061] Module input environment state vector:

[0062]

[0063] in: Distance from the robot to the target point; Pedestrian density; Relative velocity; Social distance metric. Input state vector. Used in the PPO policy network to determine how to adjust the control parameters of MPC.

[0064] PPO policy network output parameter adjustment vector:

[0065]

[0066] in: The state error weight adjustment affects the trajectory accuracy component in the MPC optimization objective. : Controlling the incremental weight adjustment affects the stationarity component of the MPC optimization objective; The adjustment of social constraint weights affects the social cost component of the MPC optimization objective. The PPO module outputs parameter increments in real time. The optimization weights of MPC are dynamically adjusted to achieve adaptive adjustment and optimization of MPC control performance.

[0067] The PPO module receives reference states (such as the output of global path planning) from the MPC module. The input state vector consists of real-time environmental data (e.g., pedestrian location, speed, and density) from the environmental perception and pedestrian state prediction module, as well as data from the environment perception and pedestrian state prediction module. The PPO dynamically adjusts the MPC control parameters based on these inputs. The adjustment parameters output by the PPO module... This information is then passed to the MPC module. After receiving these adjustment parameters, the MPC module re-optimizes its control strategy based on the new control weights to ensure that the robot's behavior meets trajectory accuracy requirements, operates smoothly, and satisfies social constraints, ensuring that its interactions with those around it conform to social norms.

[0068] The updated parameters are passed to the MPC module for dynamically adjusting and optimizing the objective function.

[0069]

[0070] in: These represent the online adjustment amounts for trajectory error weights, control increment weights, and social constraint weights, respectively. The PPO outputs these increments in real time based on changes in environmental conditions to achieve adaptive correction of MPC control performance. The updated weights are used to resolve the MPC optimization problem, thereby ensuring a dynamic balance between tracking accuracy, control smoothness, and social friendliness in different scenarios. This optimization objective updates the policy parameters through gradient descent. This improves decision-making performance while ensuring strategy stability. The reward function is designed as follows: the immediate reward of PPO is composed of a weighted average of multiple metrics.

[0071]

[0072] in: : Goal-advancing rewards encourage robots to move closer to their goals; Collision penalty: When the robot collides with a pedestrian or obstacle, it gives a large negative reward. Rewards for smooth trajectory, penalties for high acceleration or frequent turns; Social penalties are imposed when a robot enters a pedestrian's personal space or fails to avoid them properly.

[0073] When the accumulated social costs exceed the threshold At this time, the system triggers a policy update to enhance the adaptability of human-computer interaction. This mechanism ensures that the robot can maintain smooth, safe, and socially friendly navigation behavior even in high-density crowds or sudden interaction situations. Policy optimization mechanism: PPO limits the policy update magnitude by pruning the probability ratio of the objective function. Its optimization objective is defined as:

[0074]

[0075] in: : Probability ratio between the old and new strategies; The advantage function represents the advantage of the current strategy relative to the old strategy. This formula updates the constraint factor to ensure stability during training and prevent excessive policy changes. It guarantees that each policy update will not cause the new policy to perform worse than the old one, thus avoiding excessive fluctuations and ensuring stable policy convergence.

[0076] PPO Training Mechanism: This module optimizes the PPO policy network through offline training to obtain the optimal policy parameters. After training, the deployment phase only performs forward inference of the policy network, without updating the weights, enabling online adjustment of MPC parameters. This significantly reduces online computational overhead and achieves lightweight real-time adaptive control. The parameter correction amount output by PPO is... The system transmits data in real time to the MPC via the system interface, which then re-optimizes the control strategy based on the updated control weights. The MPC combines the reference path output from global path planning with a rolling optimization method to dynamically adjust the trajectory, avoiding obstacles, yielding to pedestrians, and adhering to social norms. Finally, the PPO outputs correction values ​​in real time, achieving closed-loop control of "perception-learning-optimization-execution," ensuring efficient, safe, and socially friendly navigation of the robot in dynamic environments.

[0077] Step 4: Actual online operation, including the following steps:

[0078] When running online, step 2 outputs As a social constraint, the trained PPO policy network output in step 3 is used for output. Each control cycle executes the following steps: Step 4.1 Global Planning → Step 4.2 Environmental Perception and Prediction → Step 4.3 MPC (two cases) → Execute Control → Environmental Feedback → Determine if Target Reached → Loop.

[0079] Step 4.1 Global Path Planning: In a known static map environment, a feasible global reference path is generated based on the robot's current position and the target point's position. The module employs a global path planning algorithm, such as the improved A* heuristic search algorithm, which optimizes the path by comprehensively considering distance cost and a heuristic function. Its objective function is:

[0080]

[0081] in, From the starting point to the node The cumulative cost, This represents the heuristic distance estimate from the node to the target point. The global path set output by the algorithm is denoted as:

[0082]

[0083] This path serves as a reference input for the local control module.

[0084] in, It is the starting point position. This refers to the target point location, while the path set contains all the key nodes from the starting point to the end point. This path set will serve as the reference input for subsequent local control modules, ensuring that the robot can move along the predetermined path and making dynamic adjustments through the local control modules to cope with real-time environmental changes.

[0085] Input and output:

[0086] Input: The robot's current position and the target point position.

[0087] Output: Global path set , which serves as the reference input for local control.

[0088] Global path planning works closely with subsequent environmental perception and local control. During robot movement, environmental perception continuously provides updated obstacle information and dynamic pedestrian status. When obstacles or pedestrians change, local control adjusts according to the current path. If there is a significant conflict between the global path and the current environment, local control can also replan the path or perform local obstacle avoidance to ensure safety and goal achievement.

[0089] Step 4.2 Environmental Awareness (Executed Online)

[0090] During online operation, this step executes the sensing link designed in "Step 1" (RGB-D, LiDAR, IMU fusion), and outputs: global state vector. The system predicts the future trajectory of pedestrians (position and speed information in the next 1-5 seconds) and provides a basis for determining "dynamic pedestrian detected" in step 4.3.

[0091] Step 4.3 MPC Scrolling Optimization (two cases: whether or not dynamic pedestrians are detected)

[0092] Model Predictive Control (MPC), based on the principle of rolling optimization, solves for the optimal control input that satisfies system dynamics and safety constraints within a finite prediction time domain, achieving high-precision tracking and smooth control of the robot trajectory. This module aims to minimize the cost function, comprehensively considering trajectory error, control increment, and social cost constraints. During online operation, MPC simultaneously receives the social cost function learned from offline IRL training and the parameter tuning strategy trained from offline PPO training, thus achieving closed-loop optimization control of "social constraints + online adaptive parameter tuning".

[0093] Input and output:

[0094] Input: The module receives reference states from the global path planning module. And the social cost function from the Inverse Reinforcement Learning (IRL) module. and the robot's current state. and control input (linear velocity) and angular velocity The policy network trained by PPO only performs forward inference in the online phase, based on the environment state vector. and output parameter adjustment vector ,in These are used to correct the trajectory error weight, control increment weight, and social cost weight for online MPC correction, respectively.

[0095] Output: The module outputs the optimal control input sequence. Each control input [This is at each prediction step] The optimal control input is calculated below. The system executes the first control input based on the rolling optimization result. And it will perform rolling optimization in the next control cycle to ensure real-time closed-loop control.

[0096] Dynamics Model and Constraints

[0097] The system state update model is as follows:

[0098]

[0099] And it satisfies the following constraints:

[0100]

[0101] in: Robots and the First The distance between pedestrians; : Safety distance threshold; : Social distance dynamic compensation coefficient.

[0102] Multi-objective optimization: The MPC module considers the following objectives to optimize robot behavior:

[0103] Trajectory accuracy: By constraining state deviations, this ensures the robot accurately tracks the target path. This objective is achieved by... This ensures that the robot gets as close as possible to the reference path.

[0104] Motion smoothness: By constraining the rate of change of input, abrupt changes in acceleration and angular velocity are suppressed, thereby improving trajectory smoothness and robot operation comfort. This objective is achieved through... Implementation, in which To control the increment.

[0105] Social behavioral norms: Introducing the social cost function learned by the IRL module. This aims to regulate robot behavior in dynamic human-robot interaction environments, ensuring it conforms to social inclusion and safety standards. This project is undertaken by... Implementation, in which This is the weighting coefficient for social costs.

[0106] No dynamic pedestrians detected (No) - Standard MPC

[0107] When the environmental perception module determines that no dynamic pedestrians have been detected (e.g., the number of pedestrians in the scene), The system executes standard MPC trajectory tracking and smoothing control, at which point the social behavior norms item is not enabled (equivalent to...). (Or ignore the social aspect).

[0108] The objective function to be optimized is:

[0109]

[0110] in: Let k be the robot's state vector when predicting step size k. The reference state output by the global path planning module; For control input; To control the increment; These are the state error weight matrix and the control smoothing weight matrix, respectively.

[0111] Dynamic pedestrian detected (yes) — Adaptive MPC with IRL social constraints + PPO online parameter tuning

[0112] When the environmental perception module detects a dynamic pedestrian (e.g.) During the online control phase, the social cost function obtained and fixed through offline IRL learning is invoked. Simultaneously, the policy network trained offline using PPO is invoked to adaptively adjust the MPC weights online: the PPO policy network uses the environment state vector constructed in the previous stage. Input and output parameter adjustment vectors Based on this, the MPC optimization objective function with social constraints and after PPO parameter tuning is:

[0113]

[0114] in : The social cost function learned by the IRL module; , , The weights obtained after online parameter tuning for PPO are used to dynamically balance the strength of social constraints while ensuring tracking accuracy and motion smoothness, so that the robot can maintain safe and socially friendly navigation behavior in pedestrian interaction scenarios.

[0115] This reflects the coupling effect between velocity and safety space. The optimization problem is solved in real time by a nonlinear constrained optimization solver (such as qpOASES or OSQP). The optimal control sequence output by the solution is:

[0116]

[0117] The system only executes the first control input:

[0118]

[0119] And it performs rolling optimization in the next control cycle to achieve real-time closed-loop control.

[0120] Step 4.4 repeats steps 4.2–4.3 until the target point is reached (closed-loop execution).

[0121] The online operation performs closed-loop iterations in control cycles. Each control cycle first executes step 4.2, the environmental perception and pedestrian state prediction module, which updates the global state vector using RGB-D, LiDAR, and IMU fusion. It then outputs the pedestrian trajectory prediction results for the next 1-5 seconds; subsequently, it proceeds to step 4.3, the MPC rolling optimization module, which determines whether dynamic pedestrians have been detected based on the perception results and selects the corresponding optimization branch: when no dynamic pedestrians are detected (e.g. When a moving pedestrian is detected, standard MPC trajectory tracking and smoothing control is performed; when a moving pedestrian is detected (such as...), standard MPC trajectory tracking and smoothing control is performed. The social cost function, obtained and fixed through offline IRL learning, is invoked at that time. Meanwhile, the policy network trained offline by PPO performs forward inference and outputs during the online phase. This is used to correct the MPC weights online before performing rolling optimization. The MPC solution yields the optimal control sequence. After that, the system only executes the first control input. The system drives the robot to move and generates new environmental feedback and state updates. The system then determines whether the target point has been reached: if the termination condition is met (e.g., ...). If the condition is met, the online operation ends; otherwise, the next control cycle begins, and the closed-loop process of step 4.2 Environmental Perception and Prediction → step 4.3 MPC Rolling Optimization → Execution Control → Environmental Feedback continues until the robot finally reaches the target point.

Claims

1. A social navigation method for mobile robots, characterized in that, Includes the following steps: (1) Perform global path planning based on the static environment map to generate a global reference path; (2) When a dynamic pedestrian is detected, local social navigation is initiated, which includes: learning and extracting the social cost function from human demonstration behavior through inverse reinforcement learning; using model predictive control method to integrate the social cost function, robot dynamics constraints and environmental constraints, and rolling optimization to generate local control trajectory within the finite prediction time domain; the social cost function is learned from the demonstration trajectory dataset containing human social interaction behavior through inverse reinforcement learning algorithm, and is used to introduce social behavior constraints in trajectory optimization; the model predictive control method considers trajectory tracking accuracy, control input smoothness and social cost constraints extracted by inverse reinforcement learning simultaneously during optimization, and rolls to solve the optimal control sequence and executes the first control instruction in each control cycle; (3) By using a reinforcement learning regulator, the weights of various control costs in the model predictive control method are dynamically adjusted according to the real-time environmental state; The reinforcement learning regulator generates control parameter adjustments based on real-time environmental state information and a proximal policy optimization algorithm, and dynamically updates the weights corresponding to trajectory tracking error, control smoothness, and social constraints in the objective function of the model predictive control method.

2. The social navigation method for mobile robots according to claim 1, characterized in that, The reinforcement learning regulator is trained offline using a multi-dimensional reward function that includes goal advancement rewards, collision penalties, trajectory smoothing rewards, and social interaction penalties. The trained policy network is then subjected to forward inference during the online deployment phase to adjust parameters online.

3. The social navigation method for mobile robots according to claim 1, characterized in that, The method further includes: acquiring environmental information through multi-sensor fusion, constructing a global state vector containing the states of the robot and pedestrians; predicting the future trajectory of pedestrians based on a time-series prediction model, and using the prediction results for the decision-making process of local social navigation.

4. A social navigation system for mobile robots, characterized in that, The method described in any one of claims 1-3 is used to implement the method, comprising: The Global Reference Path module is used to perform global path planning based on a static environment map and generate a global reference path. The local social navigation module is used to initiate local social navigation when a dynamic pedestrian is detected. It includes: an inverse reinforcement learning unit for learning and extracting the social cost function from human demonstration behavior; and a model prediction control unit for fusing the social cost function, robot dynamics constraints, and environmental constraints to generate a local control trajectory through rolling optimization within a finite prediction time domain. The adjustment module is used to dynamically adjust the weights of trajectory tracking error, control smoothness, and social constraints in the model prediction control unit based on the real-time environmental state through a reinforcement learning regulator.

5. An electronic device, characterized in that, It includes a memory and a processor, the memory storing a computer program, and the processor executing the program to implement the steps of the method according to any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The device contains a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1-3.

Images