Mobile robot indoor autonomous exploration hierarchical planning method based on dynamic prediction

Abstract

The invention discloses a layered planning method for indoor autonomous exploration of a mobile robot based on dynamic prediction, and the method comprises the steps: S1, dividing an exploration environment where the robot is located into a plurality of local subspaces, and marking the state of each subspace as an unexplored state, an explored state, an explored state or a semi-explored state according to the data of a sensor and a semantic detection result; s2, establishing a global topological graph connected with each subspace, and planning a global access path by using a traveling salesman problem model; s3, integrating a pedestrian trajectory prediction model in the selected current subspace, and evaluating and selecting a local viewpoint capable of maximizing information gain and minimizing dynamic collision risk based on a prediction result; the method has the beneficial effects that a semi-exploration state and a self-adaptive revisit mechanism are introduced, and when the door is possibly opened, the check is turned back, so that the dead angle of a map is eliminated, and the exploration completeness is improved; pedestrian trajectory prediction is incorporated into viewpoint scoring, and people deadlock is avoided.

Description

Technical Field

[0001] This invention relates to the intersection of mobile robot technology, artificial intelligence and autonomous navigation technology, specifically a hierarchical planning method for autonomous indoor exploration of mobile robots based on dynamic prediction. Background Technology

[0002] Autonomous exploration is one of the core capabilities of mobile robots to achieve intelligent operations. Its goal is to actively perceive the unknown environment using onboard sensors (such as LiDAR and RGB-D depth cameras) without any prior map information, plan movement paths to maximize the acquisition of environmental information, and ultimately construct a complete environmental map. This technology has irreplaceable value in fields such as disaster relief (e.g., search and rescue inside collapsed buildings), extraterrestrial exploration (e.g., Mars rover navigation), underground mine exploration, and indoor services (e.g., shopping mall cleaning and hospital supply delivery).

[0003] Current autonomous exploration methods are mainly divided into frontier-based methods, sampling-based methods, and hierarchical planning methods that combine both (such as TARE Planner and FAEL). However, existing technologies face two major challenges in practical applications, resulting in poor performance in dynamic indoor environments: First, most existing exploration algorithms assume a static environment. When robots encounter dynamic obstacles such as pedestrians, they often rely solely on low-level local obstacle avoidance algorithms (such as DWA) for passive responses. The planning layer lacks the ability to predict the future trajectory of pedestrians, causing the robot's selected viewpoint to be frequently blocked by the pedestrian's future path. This forces the robot to frequently stop abruptly, rotate in place, or get trapped in crowds, greatly reducing exploration efficiency.

[0004] Indoor environments contain numerous "movable objects," such as doors, chairs, and carts. Existing algorithms typically mark closed doors as permanent obstacles (walls) and permanently remove the area behind them from the exploration graph. Even if the door is subsequently opened, the robot cannot automatically revisit the area because it has already abandoned the node in the internal topology map. This leads to incomplete map construction, which is unacceptable for tasks with extremely high coverage requirements, such as search and rescue or security.

[0005] Therefore, there is an urgent need for an autonomous exploration planning method that can proactively predict pedestrian dynamics and intelligently handle changes in the state of movable objects. Summary of the Invention

[0006] The purpose of this invention is to provide a hierarchical planning method for autonomous indoor exploration of mobile robots based on dynamic prediction, which aims to solve the problems of low exploration efficiency and insufficient coverage in existing technologies in dynamic and semi-static environments.

[0007] To achieve the above objectives, a hierarchical planning method for autonomous indoor exploration of mobile robots based on dynamic prediction is proposed, comprising the following steps: S1. Divide the robot's exploration environment into several local subspaces, and mark the state of each subspace as unexplored, exploring, explored, or partially explored based on sensor data and semantic detection results. The partially explored state refers to the area containing exploration conveniences that are currently blocked by movable objects. S2. Establish a global topology graph connecting all subspaces and plan global access paths using the traveling salesman problem model; when calculating the path cost to a subspace in a semi-explored state, dynamic weights based on time decay and environment triggering are introduced for correction. S3. Within the selected current subspace, integrate the pedestrian trajectory prediction model, evaluate and select the local viewpoint that maximizes information gain and minimizes dynamic collision risk based on the prediction results. S4. Generate a smooth trajectory based on the selected viewpoint sequence and control the robot to execute it.

[0008] Preferably, the movable object in the semi-exploration state includes a door, chair, or cart; the subspace in the semi-exploration state has a timestamp indicating that it was first discovered.

[0009] Preferably, in step S2, the distance cost from the current node i to the node j in the semi-exploration state is... The cost of the correction Wherein, the weight α is a function based on time decay and environmental triggering: ,in, The weight is related to the distance between the robot and the target and the duration (∆t) of the object being occluded. When a pedestrian is detected passing through the occluder, the weight drops rapidly, triggering a global replanning.

[0010] Preferably, the state transition index The calculation formula is: , in, This is the robot's current position. Γ is the centroid of the semi-exploration subspace; ∆t is the duration from the first discovery to the present; Γ is the directional indicator factor; θ is the deviation angle between the robot's orientation and the target; Ξ is the pedestrian interaction marker; Υ is the spatial proximity indicator factor; λ is a large constant used to trigger state switching; β and δ are adjustment coefficients used to balance the weights of various factors.

[0011] Preferably, when the semantic perception system detects that a pedestrian has entered the semi-exploration subspace or interacted with a movable object in the area, the pedestrian interaction flag Ξ is set to 1, causing the weight α to rapidly approach 1, thereby triggering the global planner to prioritize the inclusion of the subspace in the access sequence.

[0012] Preferably, the pedestrian trajectory prediction model in step S3 adopts a neural network based on an encoder-decoder structure. Its input includes the robot state, the local occupancy grid map and the historical trajectories of surrounding pedestrians, and the output is the velocity vector sequence of pedestrians for a future period of time. The prediction model adopts a self-supervised continuous learning mechanism to fine-tune the network parameters online using the error between prediction and observation.

[0013] Preferably, in step S3, a reward function that includes dynamic collision penalty is used. To evaluate candidate viewpoint i, where For information gain; Collision risk calculated based on predicted pedestrian speed sequences; and These are the weighting coefficients.

[0014] Preferably, the collision risk The calculation is based on a speed obstacle model, utilizing predicted future pedestrian speed sequences. For viewpoint i, assume the robot moves at a speed of... Head to predict collision time , , in, This is the pedestrian's position at time τ, calculated based on the prediction model; if there is a collision risk, the penalty term is calculated as follows: , in, For estimated arrival time; This is the predicted collision time; if there is no risk of collision, .

[0015] Preferably, in step S5, a smooth trajectory is generated using a B-spline curve and executed in conjunction with a model prediction controller; during execution, a reactive obstacle avoidance based on gradient velocity obstacles is used as a safety redundancy.

[0016] Compared with existing technologies, the beneficial effects of this invention are as follows: By introducing a "semi-exploration" state and an adaptive revisit mechanism, this invention enables the robot to check back when a door might be open (such as when someone passes by or after a period of time), just like a human, thereby eliminating map blind spots and improving the completeness of exploration; by incorporating pedestrian trajectory prediction into viewpoint scoring, this invention enables the robot to predict congestion and proactively plan "off-peak" routes, avoiding getting stuck in crowds, significantly improving average movement speed, and enhancing exploration efficiency in dynamic environments; the prediction model adopts a self-supervised continuous learning architecture, which can adapt to the characteristics of pedestrian flow in different places without manual data labeling, and has strong adaptability. Attached Figure Description

[0017] Figure 1 This is a flowchart of the method of the present invention; Figure 2 This is a schematic diagram of the hierarchical planning system framework proposed in this invention; Figure 3 This is a schematic diagram illustrating the principle of the adaptive cost matrix in global planning. Figure 4 These are illustrations depicting effects observed in embodiments of the present invention. Figure 5 This is a comparative data graph showing the exploration volume and time efficiency of the method of this invention and the prior art (TARE, FAEL). Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] Please see Figure 1-5 This invention provides a technical solution: a hierarchical planning method for autonomous indoor exploration of mobile robots based on dynamic prediction, comprising the following steps: S1. Divide the robot's exploration environment into several local subspaces, and mark the state of each subspace as unexplored, exploring, explored, or semi-explored based on sensor data and semantic detection results. The semi-explored state refers to an area containing exploration conveniences that are currently blocked by movable objects. Movable objects in the semi-explored state include doors, chairs, or carts. The subspace in the semi-explored state has a timestamp of its first discovery.

[0020] S2. Establish a global topology graph connecting all subspaces and plan global access paths using the Traveling Salesman Problem model; when calculating the path cost to a subspace in a semi-explored state, dynamic weights based on time decay and environment triggering are introduced for correction; the distance cost from the current node i to the node j in a semi-explored state is calculated. The cost of the correction Wherein, the weight α is a function based on time decay and environmental triggering: ,in, The weights are related to the distance between the robot and the target, and the duration (∆t) of the object being occluded. When a pedestrian is detected passing through the occlusion, the weights decrease rapidly, triggering a global replanning; state transition metrics The calculation formula is: , in, This is the robot's current position. Let α be the centroid of the semi-exploration subspace; ∆t be the duration from the first discovery to the present; Γ be the directional indicator factor; θ be the deviation angle between the robot's orientation and the target; Ξ be the pedestrian interaction flag; Υ be the spatial proximity indicator factor; λ be a large constant used to trigger state switching; β and δ are adjustment coefficients used to balance the weights of each factor. When the semantic perception system detects that a pedestrian has entered the semi-exploration subspace or interacted with a movable object in the area, the pedestrian interaction flag Ξ is set to 1, causing the weight α to rapidly approach 1, thereby triggering the global planner to prioritize the inclusion of this subspace in the access sequence.

[0021] S3. Within the selected current subspace, integrate the pedestrian trajectory prediction model, evaluate and select the local viewpoint that maximizes information gain and minimizes dynamic collision risk based on the prediction results; use a neural network based on an encoder-decoder structure, whose inputs include the robot state, the local occupancy grid map, and the historical trajectories of surrounding pedestrians, and whose output is the velocity vector sequence of pedestrians over a future period; the prediction model adopts a self-supervised continuous learning mechanism, using the error between prediction and observation to fine-tune the network parameters online; and employs a reward function that includes dynamic collision penalties. To evaluate candidate viewpoint i, where For information gain; Collision risk calculated based on predicted pedestrian speed sequences; and These are the weighting coefficients; Collision risk The calculation is based on a speed obstacle model, utilizing predicted future pedestrian speed sequences. For viewpoint i, assume the robot moves at a speed of... Head to predict collision time , , in, This is the pedestrian's position at time τ, calculated based on the prediction model; if there is a collision risk, the penalty term is calculated as follows: , in, For estimated arrival time; This is the predicted collision time; if there is no risk of collision, .

[0022] S4. Generate a smooth trajectory based on the selected viewpoint sequence and control the robot to execute it; use B-spline curves to generate a smooth trajectory and combine it with a model prediction controller to execute it; during execution, use gradient velocity-based reactive obstacle avoidance as a safety redundancy.

[0023] A mobile robot indoor autonomous exploration system for implementing any of the above methods includes an environment perception and modeling module, a global planning module, a local planning module, and a motion control module; The environmental perception and modeling module is used to perform sensor data acquisition, semantic recognition, synchronous localization and map building, and maintain an environmental model containing different state subspaces; the global planning module is used to implement global path planning based on adaptive state transition cost; the local planning module is used to implement local viewpoint selection and path planning based on adaptive pedestrian prediction; and the motion control module is used to perform trajectory generation, tracking and emergency obstacle avoidance.

[0024] It should be noted that this embodiment uses a mobile robot platform equipped with a 3D LiDAR (Livox MID360) and an RGB-D depth camera (RealSense D435i); The modeling module uses the SLAM algorithm (DLIO) to estimate the robot pose in real time. The system maintains a global point cloud map and a low-resolution voxel grid map for planning. It uses a semantically aware object detection algorithm (YOLOv11) to identify specific objects (people, doors, chairs) within the field of view in real time. The system dynamically maintains a list of subspaces. When a new frontier is detected, it checks whether the frontier is located within or behind a 3D bounding box identified as a "movable object." If so, the area is marked as a "semi-explored subspace," and the first discovery timestamp is recorded. Otherwise, mark it as the regular "Exploring Neutron Space"; The core task of the global planning module is to solve a generalized Traveling Salesman Problem (TSP) to determine the optimal order for visiting various subspaces; suppose there are currently several subspaces to be visited (including those under exploration and those in partial exploration). A distance matrix is ​​defined, where each element represents the path distance from a node. For the "partially explored" subspace (denoted as node ), this invention modifies its distance cost: , Introduce the reciprocal of the state transition probability as a weight: , The formula for determining the robot's willingness to revisit the semi-explored area is as follows: , in, Let be the Euclidean distance between the robot's current position and the centroid of the subspace. The closer the distance, the smaller the denominator. The larger α is, the larger α becomes (to prevent frequent oscillations without a clear trigger); Δt is a time factor, the current time minus As time progresses, the denominator increases, α gradually decreases, and the probability of the simulated door potentially opening over time increases; Γ is a Boolean value, indicating whether the centroid of the "semi-exploration" subspace is within or outside the predefined field of view in the robot's forward direction; θ is the angular deviation between the robot's orientation and the target area; Ξ (critical trigger factor) is the pedestrian interaction flag. If the vision system detects a pedestrian entering the semi-exploration area or interacting with them (such as opening a door), Ξ is set to 1; otherwise, it is 0; λ is set to a large positive number (e.g., λ=10000), at which point the denominator increases dramatically. →0, α→1, which instantly restores the access cost of the region to the real physical distance, and the TSP planner will immediately put the region into the priority access sequence; Υ is a Boolean value, which is 1 if the semi-explored region is the closest to the current connected region, and 0 otherwise; β and δ are adjustment coefficients used to balance the weights of various factors; The local planning module operates within the selected current subspace at a frequency of 10Hz-20Hz; the model architecture employs an encoder-decoder structure; inputs include the robot's past frame velocity, local occupancy grid map, and the relative positions and velocity sequences of surrounding pedestrians; the interaction module utilizes an attention mechanism to extract human-to-human and human-to-robot interaction features; outputs the velocity sequence of each surrounding pedestrian in the next second... Self-supervised update: The system caches historical prediction results. At time 1, it retrieves the prediction made at time 2 and calculates the error (Loss) by comparing it with the currently observed pedestrian state. It then fine-tunes the network weights online through backpropagation to adapt them to the current crowd behavior patterns (e.g., the rushing mode in a hospital vs. the strolling mode in a shopping mall). In the prediction-based viewpoint reward function, the system uniformly samples candidate viewpoints in the local free space and calculates a score for each viewpoint. , in, (Information gain) is calculated by subtracting the number of unknown grid cells that the viewpoint can cover from the number of cells that overlap with the selected viewpoint. (Dynamic collision penalty) constructs dynamic velocity obstacles using predicted pedestrian velocity sequences. For the viewpoint, assuming the robot travels at a certain velocity, the collision time is predicted as follows: , in The pedestrian's position at a given time is estimated based on the prediction model. If there is a collision risk, the penalty term is calculated as follows: If there is no risk of collision, , This mechanism ensures that the selected best viewpoint not only sees the most unknown areas, but also that the robot is unlikely to be blocked by the future movements of pedestrians while traveling to that viewpoint. When the motion control module is running, after selecting the optimal viewpoint sequence, it generates a smooth trajectory using B-spline curves and sends it to the underlying MPC (Model Predictive Control) controller for execution. During execution, if an unpredictable obstacle is encountered, gradient velocity obstacle avoidance (GVO) is activated as a last line of defense.

[0025] Example 1: Dynamic Exploration of the Shopping Mall Environment and Revisit of the Area Behind the Doors Scenario Description: A shopping mall cleaning robot (model: CleanBot Pro) is performing fully automated mall map building and cleaning route planning tasks late at night. The mall has a complex structure, containing multiple shops, some of which are closed, and a small number of security personnel are still moving around.

[0026] Implementation process: ① The robot starts from the mall lobby and uses LiDAR and a depth camera to create a SLAM map. When it approaches a closed glass door of a clothing store, the YOLO detection module identifies it as a "door." The robot detects a large area of ​​unknown space behind the door (the LiDAR point cloud breaks at the door, forming a boundary), but cannot enter due to the door's obstruction. The system marks this area as a "semi-explored subspace," recording the centroid position and timestamp. .

[0027] ② When planning subsequent paths, the global planner calculates the cost of going to this "semi-explored" clothing store. Initially, due to the distance and the fact that it has just been discovered, the dynamic weight α is large, amplifying the access cost. Therefore, the robot decides to prioritize exploring other completely open public areas.

[0028] ③ Approximately 20 minutes later, a security guard patrolled the area and pushed open the door of the clothing store to conduct an inspection. The robot's vision system detected the interaction between the "pedestrian" and the "door" (the Ξ flag was set to 1). This event immediately triggered an update to the global cost matrix, causing the cost weight α for going to the clothing store to drop sharply to near zero.

[0029] ④ The global planner replans based on the new cost matrix, including the clothing store area in the high-priority access sequence. The robot then adjusts its route, enters the opened store, and completes the full map construction and cleaning path planning for that area. Ultimately, the mall map achieves near 100% coverage, with no blind spots.

[0030] This embodiment demonstrates the core capability of the present invention in handling the "movable object obstruction" problem. Instead of permanently abandoning the area behind the closed door, the robot, like a human, "waits for the right moment" or "seizes the opportunity" to revisit the area through a "semi-exploration" state marker and a "pedestrian interaction trigger" mechanism, thus ensuring the completeness of the map.

[0031] Example 2: Efficient traversal of hospital corridors during peak hours Scenario Description: A hospital logistics delivery robot (model: MedTransporter) needs to travel from the pharmacy through a busy corridor to the inpatient department during the morning rush hour. The corridor is frequently used by medical staff, patients, and their families, and contains many dynamic obstacles.

[0032] Implementation process: ① Before the robot entered the corridor, its integrated self-supervised pedestrian prediction model had already learned online and initially adapted to the behavioral patterns of people in the hospital environment who walk relatively fast and whose trajectories are relatively purposeful (such as walking towards a ward or nurse station).

[0033] ② Within the "exploring" subspace of the corridor, the local planner samples multiple candidate viewpoints (such as several intermediate locations at the front of the corridor). For each viewpoint, the planner not only calculates the new area it can observe, but also... More importantly, it calls upon the predictive model to predict the trajectories of pedestrians in the surrounding area within the next few seconds.

[0034] ③ The prediction model output shows that if the robot travels straight along the normal path, it will intersect with the trajectory of a nurse walking quickly towards the intersection after 3 seconds, and there are carts converging from the side, which may create a temporary "deadlock". Based on this prediction, the collision penalty item... This has a significant negative impact on the rating of the straight viewpoint.

[0035] ④ Conversely, another viewpoint approach that pauses slightly and moves slowly to the right, while having a slightly lower information gain, predicts that it can perfectly avoid the future paths of all pedestrians. The value is very low. Therefore, this scheme scores poorly in the overall evaluation. China won.

[0036] ⑤ The local planner selects the "move slowly to the right" viewpoint and generates a smooth, yielding B-spline trajectory. The robot gracefully blends into the crowd, avoiding sudden stops, spinning in place, or confrontations with pedestrians, and passes through the corridor efficiently and safely.

[0037] This embodiment highlights the efficiency advantages of the present invention in highly dynamic environments. By embedding pedestrian trajectory prediction depth into the evaluation system of local planning, the robot possesses predictive capabilities, enabling it to proactively plan socially friendly and efficient paths, avoiding passive reactions and congestion, and significantly improving the average speed and smoothness of task execution.

[0038] Example 3: Persistent Monitoring and Adaptive Testing in Integrated Building Security Patrols Scenario Description: A security patrol robot (model: SecPatrol AI) is responsible for continuous autonomous patrols of a research and development office building at night. The building contains numerous meeting rooms and laboratories, with tables and chairs frequently moved around, and occasional staff working overtime.

[0039] Implementation process: ① During the inspection, the robot successively discovered that several meeting room doors were closed and some laboratory entrances had movable equipment cabinets placed there. The system marked these areas as their respective "semi-exploration subspaces" and continuously maintained a list containing multiple such areas, each with its own independent ∆t (duration not revisited).

[0040] ② For a meeting room that has been closed for several hours (∆t is large), although no pedestrian interaction was detected, its dynamic weight α has gradually decreased over time. After completing a round of inspections of other open areas, the global planner automatically generates an inspection route to the door of the meeting room because the access cost of this meeting room has been reduced to a reasonable level.

[0041] ③ The robot arrives outside the conference room. The local planner activates, attempting to find a viewpoint through the door crack or glass to observe the interior. Simultaneously, the prediction model continues to run, assessing the surrounding environment (such as whether anyone is approaching at the end of the corridor). The robot selects a safe pose with the greatest information acquisition for a brief stop. After confirming that the door remains closed and there are no abnormalities inside, the robot updates the subspace state to "semi-exploration" and continues with subsequent tasks.

[0042] ④ When the robot patrols the open office area, the predictive model detects the movement trajectories of two people indicating they are heading towards a "semi-explored" storage room containing miscellaneous items. The Ξ flag is triggered. The global planner immediately adjusts the plan, prioritizing the robot to follow (maintaining a safe distance) to the vicinity of this area. After the people open the door and leave, the robot immediately enters the storage room that has completed its state transition (α≈1), completes the exploration of the area, and marks it as "explored" after confirming its safety.

[0043] This embodiment demonstrates the persistence, adaptability, and multi-objective management capabilities of the system of the present invention. It can not only handle single-triggered events but also simulate the logic of human "periodic review" through a "time decay" mechanism and adapt to different dynamic patterns through a continuously learning predictive model. The system can intelligently manage multiple potential points of interest, ensuring regular coverage while responding agilely to emergencies, making it suitable for long-term, large-scale security patrol tasks.

[0044] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A hierarchical planning method for autonomous indoor exploration of mobile robots based on dynamic prediction, characterized in that, Includes the following steps: S1. Divide the robot's exploration environment into several local subspaces, and mark the state of each subspace as unexplored, exploring, explored, or partially explored based on sensor data and semantic detection results. The partially explored state refers to the area containing exploration conveniences that are currently blocked by movable objects. S2. Establish a global topology graph connecting all subspaces and plan global access paths using the traveling salesman problem model; when calculating the path cost to a subspace in a semi-explored state, dynamic weights based on time decay and environment triggering are introduced for correction. S3. Within the selected current subspace, integrate the pedestrian trajectory prediction model, evaluate and select the local viewpoint that maximizes information gain and minimizes dynamic collision risk based on the prediction results. S4. Generate a smooth trajectory based on the selected viewpoint sequence and control the robot to execute it.

2. The hierarchical planning method for indoor autonomous exploration of mobile robots based on dynamic prediction according to claim 1, characterized in that, The movable objects in the semi-exploration state include doors, chairs, or carts; the subspace in the semi-exploration state has a timestamp indicating that it was first discovered.

3. The hierarchical planning method for indoor autonomous exploration of mobile robots based on dynamic prediction according to claim 1, characterized in that, In step S2, the distance cost from the current node i to the node j in the semi-exploration state is... The cost of the correction Wherein, the weight α is a function based on time decay and environmental triggering: ,in, The weight is related to the distance between the robot and the target and the duration of the object being occluded. When a pedestrian is detected passing through the occluded object, the weight drops rapidly, triggering a global replanning.

4. The hierarchical planning method for indoor autonomous exploration of mobile robots based on dynamic prediction according to claim 3, characterized in that, The state transition index The calculation formula is: ， in, This is the robot's current position. Γ is the centroid of the semi-exploration subspace; ∆t is the duration from the first discovery to the present; Γ is the directional indicator factor; θ is the deviation angle between the robot's orientation and the target; Ξ is the pedestrian interaction marker; Υ is the spatial proximity indicator factor; λ is a large constant used to trigger state switching; β and δ are adjustment coefficients used to balance the weights of various factors.

5. A hierarchical planning method for indoor autonomous exploration of mobile robots based on dynamic prediction, as described in claim 4, is characterized in that... When the semantic perception system detects that a pedestrian has entered the semi-exploration subspace or interacted with a movable object in the area, the pedestrian interaction flag Ξ is set to 1, causing the weight α to rapidly approach 1, thereby triggering the global planner to prioritize the inclusion of the subspace in the access sequence.

6. A hierarchical planning method for autonomous indoor exploration of mobile robots based on dynamic prediction, as described in claim 1, is characterized in that... The pedestrian trajectory prediction model in step S3 adopts a neural network based on an encoder-decoder structure. Its input includes the robot state, the local occupancy grid map and the historical trajectories of surrounding pedestrians, and the output is the velocity vector sequence of pedestrians for a future period of time. The prediction model adopts a self-supervised continuous learning mechanism to fine-tune the network parameters online using the error between prediction and observation.

7. A hierarchical planning method for autonomous indoor exploration of mobile robots based on dynamic prediction, as described in claim 1, is characterized in that... In step S3, a reward function that includes dynamic collision penalty is used. To evaluate candidate viewpoint i, where For information gain; Collision risk calculated based on predicted pedestrian speed sequences; and These are the weighting coefficients.

8. A hierarchical planning method for indoor autonomous exploration of mobile robots based on dynamic prediction, as described in claim 7, is characterized in that... The collision risk The calculation is based on a speed obstacle model, utilizing predicted future pedestrian speed sequences. For viewpoint i, assume the robot moves at a speed of... Head to predict collision time , ， in, This is the pedestrian's position at time τ, calculated based on the prediction model; if there is a collision risk, the penalty term is calculated as follows: ， in, To estimate arrival time; This is the predicted collision time; if there is no risk of collision, .

9. A hierarchical planning method for indoor autonomous exploration of mobile robots based on dynamic prediction, as described in claim 1, is characterized in that... In step S5, a smooth trajectory is generated using B-spline curves and executed in conjunction with a model prediction controller; during execution, reactive obstacle avoidance based on gradient velocity obstacles is used as a safety redundancy.

Images