Hospital scene-oriented mobile robot layered obstacle avoidance method

Abstract

The invention discloses a hospital scene-oriented layered obstacle avoidance method for a mobile robot, and aims to solve the problem of insufficient adaptability in a dynamic dense environment in the prior art. According to the method, a set of'perception-decision-control 'collaborative hierarchical navigation framework is constructed. In a sensing layer, an improved ORB-SLAM3 is adopted to realize the construction of a'global sparse and local dense 'hybrid map, and both the positioning precision and the calculation real-time performance are considered. In a decision-making layer, a three-layer progressive architecture of'global A * path planning-local TEB trajectory optimization-upper layer TD3 reinforcement learning self-adaptive parameter adjustment 'is designed; wherein the TD3 intelligent agent can dynamically adjust the cost function weight of the TEB algorithm according to the real-time environment state, so that the robot has the human-like decision-making ability of autonomously balancing between efficiency and safety. The whole system is integrated based on ROS and is verified through simulation and a physical platform. According to the method, the navigation robustness, safety and intelligence of the mobile robot in high-dynamic scenes such as hospitals and the like are remarkably improved.

Description

Technical fields:

[0001] This invention relates to a layered obstacle avoidance method for mobile robots in hospital settings. Background technology:

[0002] Currently, autonomous navigation of mobile robots in hospital settings heavily relies on pre-built environmental maps and real-time positioning technology, playing an increasingly important role in drug delivery, specimen transport, and material scheduling. However, in the unique environment of hospitals—characterized by high population density, complex layouts, and frequent dynamic obstacles—traditional path planning algorithms, dependent on static environmental assumptions, struggle to predict pedestrian intentions and dynamic obstacle trajectories, leading to issues such as rigid decision-making, collision avoidance failures, or path oscillations. If the navigation system fails, the robot may stall, collide with people, or even inadvertently enter sensitive areas, impacting not only logistical efficiency but also potentially endangering patient safety, disrupting medical order, and causing unforeseen medical risks. Existing technologies have proposed various improvement methods, such as introducing rule-based local obstacle avoidance strategies to handle sudden obstacles and using deep reinforcement learning to train adaptive navigation strategies. However, existing technologies still suffer from the following problems:

[0003] 1. Traditional path planning algorithms lack adaptability and intelligence in the face of dynamic obstacles;

[0004] 2. Pure reinforcement learning methods, however, suffer from drawbacks in training efficiency, global orientation, and safety and interpretability.

[0005] Therefore, existing technologies struggle to construct mobile robot navigation systems that combine robust environmental perception, global path optimization, adaptive local obstacle avoidance, and interpretable system decisions. There is an urgent need for a multi-module collaborative navigation framework that integrates lightweight visual SLAM, hierarchical path planning, and online reinforcement learning decision-making to improve the autonomy, safety, and operational efficiency of robots in high-dynamic and high-safety-requirement scenarios such as hospitals. Summary of the Invention:

[0006] This invention provides a hierarchical obstacle avoidance method for mobile robots in hospital settings. The system constructs a hierarchical architecture that coordinates perception, decision-making, and control, integrating lightweight visual SLAM, hierarchical path planning, and online reinforcement learning decision-making. This aims to achieve robust, efficient, and autonomous navigation for robots in dense, dynamic, and highly safety-critical hospital environments.

[0007] The core of this invention lies in a hierarchical navigation framework with multi-module collaboration, which mainly includes:

[0008] (1) Perception and Mapping Module: This module is built on an improved ORB-SLAM3 framework and adopts a hybrid map strategy of "global sparseness and local density". The system uses RGB-D camera and IMU data for tight coupling processing to maintain a lightweight sparse feature point map at the global level for long-term, stable drift-free localization and loop closure detection; at the same time, a densely occupied grid map of the local area is generated in real time in front of the robot to provide fine spatial structure information for subsequent obstacle avoidance decisions. This design effectively balances map accuracy and computational efficiency, and enhances the system's localization robustness in environments with frequent personnel traffic.

[0009] (2) Hierarchical Planning and Decision Module: This module adopts a three-layer progressive decision architecture, which combines global guidance, local optimization and adaptive learning.

[0010] Global path planning layer: Based on the global sparse feature map, the optimal static path from the starting point to the target point is planned using the A* search algorithm, providing macro-level task guidance for the robot.

[0011] Local trajectory optimization layer: The Time Elastic Band (TEB) algorithm is employed, using the aforementioned global path as a reference and combining real-time obstacle information from the local dense map to perform online trajectory optimization and dynamic obstacle avoidance. The TEB algorithm comprehensively considers the robot's kinematic constraints, path following accuracy, and obstacle avoidance safety.

[0012] Adaptive Learning Decision Layer: To overcome the limitations of the fixed-weight strategy in traditional TEB algorithms under complex dynamic scenarios, this layer innovatively introduces the Twin Delayed Deep Deterministic Policy Gradient (TD3) deep reinforcement learning algorithm as the upper-layer decision-maker. This decision-maker takes real-time perceived environmental states (such as the distance to the nearest obstacle, robot speed, path deviation, obstacle density, etc.) as input, and dynamically adjusts the coefficients of various constraints in the TEB cost function (such as obstacle avoidance weights, path following weights, and speed weights) online through a trained policy network. This enables the robot to autonomously make optimal trade-offs between efficiency and safety based on the current scene's congestion level and dynamic characteristics, achieving human-like flexible decision-making capabilities.

[0013] (3) System Integration and Execution Module: The entire system is integrated within the Robot Operating System (ROS). Navigation commands are ultimately converted into wheel speed commands for the robot via the underlying controller, driving the robot to execute the planned trajectory safely and smoothly. The system development and training process is conducted in the Gazebo simulation environment. A high-fidelity hospital scene model is built for large-scale algorithm training and verification, and then the system is migrated and deployed to a physical robot platform for field testing.

[0014] This invention provides an innovative systemic solution for the safe, efficient, and autonomous operation of mobile robots in complex and dynamic indoor environments such as hospitals through the deep integration of the aforementioned multiple technologies. Attached image description:

[0015] Figure 1 This is a technical roadmap of the present invention.

[0016] Figure 2 System overall flowchart

[0017] Figure 3 For autonomous vehicles used for deployment and testing Detailed implementation method:

[0019] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings.

[0020] This invention constructs a hierarchical technical approach, such as... Figure 1 As shown in the diagram. At the environmental perception level, the system integrates RGB-D camera and IMU data, relying on the ORB-SLAM3 framework to achieve lightweight map construction, generating a global sparse feature map to support accurate positioning. Simultaneously, it constructs a local dense map in real-time within the navigation area ahead, providing detailed spatial information for obstacle avoidance decisions. At the path planning level, a layered architecture is adopted: the A* algorithm handles global path planning, the TEB algorithm optimizes local trajectories, and the TD3 reinforcement learning decision-maker endows the system with intelligent obstacle avoidance capabilities in dynamic scenes. The overall flowchart is shown below. Figure 2 As shown.

[0021] 1. Overall system hardware and software setup

[0022] The hardware platform is based on NVIDIA Jetson Orin series high-performance computing units, such as... Figure 2 As shown, the system is equipped with a multi-line lidar, a 360° blind spot radar, a forward-looking camera, a millimeter-wave radar, and a high-precision differential navigation system, forming a multi-layered, redundant environmental perception and positioning sensor array. Various sensors are integrated and data is forwarded through an onboard industrial router and in-vehicle network. The data is then fused and processed by the computing unit to form a real-time, accurate understanding of the vehicle's surroundings, providing input for the decision-making and planning module.

[0023] The software platform is built on the Robot Operating System (ROS) and extends and customizes its functionality using the open-source autonomous driving framework Autoware as its core. The system software layer covers the entire algorithm chain, from sensor-driven systems, data preprocessing, perception fusion, localization and mapping, trajectory planning to vehicle control. The development and training environment is encapsulated using Docker containers, ensuring system portability and dependency consistency. It provides a Runtime Manager graphical interface, Rviz visualization tools, and a complete secondary development interface (SDK), supporting users in conveniently debugging algorithms, testing functions, and integrating the system in both simulation and real-world vehicle environments.

[0024] 2. Perception and Mapping Module

[0025] The implementation of the perception and mapping module begins with the startup and configuration of the visual-inertial SLAM system. After the robot starts, a customized ORB-SLAM3 node is run first. This system fuses visual and inertial information by enabling a tight coupling mode between RGB-D and the IMU and configuring precise spatiotemporal calibration parameters (including time offset and extrinsic parameters). To balance computational efficiency and system robustness, key parameters are carefully adjusted; for example, the number of feature points extracted is set to approximately 2000, and the image pyramid scale is optimized. Simultaneously, by setting keyframe insertion thresholds and map point culling strategies, the size of the global sparse map is effectively controlled, ensuring the system's efficiency and stability during long-term operation.

[0026] Based on the stable operation of the SLAM system and its provision of real-time pose, a hybrid map generation strategy of "global sparse - local dense" was implemented. The common view and essential map maintained internally by ORB-SLAM3 constitute a lightweight global sparse map, which is persistently stored and loaded during each task initialization, providing the robot with a long-term, consistent prior environment skeleton. Simultaneously, the system created an independent local_map ROS node, which subscribes to depth images and high-precision pose topics in real time. For each frame of input depth map, the node uses the current precise pose to convert it into a local point cloud in the world coordinate system, which is then accumulated and rasterized into a local occupancy grid map centered on the robot with a radius of approximately 5 meters. This locally dense map is updated and published at a frequency of 10Hz, providing fine-grained spatial occupancy information specifically for subsequent local path planning and real-time obstacle avoidance. This "on-demand density" architecture cleverly avoids the enormous computational overhead of globally dense reconstruction of the entire environment, achieving an optimal balance between navigation accuracy and system real-time performance.

[0027] 3. Hierarchical Planning and Decision-Making Module

[0028] The task of global path planning is to plan a macroscopically optimal path from the starting point to the target point for the robot in a known static environment. This planning process is based on a global sparse feature map constructed using ORB-SLAM3, which provides the topological structure and geometric constraints of the environment. This system uses the classic A* (A-Star) algorithm as the global planner. The A* algorithm is a typical heuristic search technique that draws on the strengths of Dijkstra's Best First (path-first) method. The core cost function is as follows:

[0029] f(n)=g(n)+h(n) (1)

[0030] Where f(n) represents the total movement cost of the grid where the robot is currently located; g(n) represents the total movement cost from the grid where the robot is currently located to the initial target location; and h(n) represents the total movement cost from the grid where the robot is currently located to the target location.

[0031] In the ROS move_base framework, the global planner is configured as AStarPlanner from the global_planner plugin. The global sparse feature map provided by ORB-SLAM3 is spatially discretized and converted into a static cost map with adjustable resolution (e.g., 0.05 meters per pixel). Known walls and fixed obstacles are set as lethal (LETHAL) regions, while free areas are set as free. Upon receiving a navigation target, AStarPlanner runs on this static cost map. Its heuristic function h(n) uses Euclidean distance to ensure search efficiency, and the planned path is published to the local planner as a series of ordered global waypoints.

[0032] In the local trajectory optimization layer, the system employs the Time Elastic Band (TEB) algorithm to perform real-time trajectory optimization and tracking of the received global waypoints. The specific process is as follows: Let the pose sequence of n robots be denoted as:

[0033] Q = {s i} i＝0...n n∈N (2)

[0034] Where s i This represents the i-th pose of the robot in the pose sequence:

[0035]

[0036] (x i ,y i ) represents the robot's position coordinates on a two-dimensional grid map, β i This represents the robot's orientation on the global map. Let the pose be s.i and position s i+1 The time interval between them is ΔT i Then we can obtain a time series consisting of n-1 time intervals:

[0037] τ={ΔT i} i＝0...n-1 (4)

[0038] Each time interval represents the time taken for the robot to transition from its current pose to the next pose in sequence Q. The time series and the robot pose sequence are combined to obtain a new tuple.

[0039] B:=(Q,τ) (5)

[0040] The key to the time elastic band algorithm is to adjust and optimize the tuple B using weights and various motion models, and finally find an optimal pose-time series.

[0041]

[0042] Among them, f k (B) represents the objective function considering a single constraint, γ k Let f(B) represent the weights corresponding to the objective function, and let f(B) represent the cost function of path planning. Our goal is to find f(B) such that the cost function is minimized.

[0043]

[0044] Finally, by configuring various kinematic constraints and optimization weights of the TEB algorithm (such as path following, obstacle avoidance, speed, etc.), it is ensured that the generated local trajectory is smooth, safe, and conforms to vehicle dynamics characteristics.

[0045] To enhance intelligence in dynamic environments, the module innovatively integrates an adaptive learning decision layer, the core of which is a decision-maker built based on the TD3 deep reinforcement learning algorithm. This decision-maker is trained offline: in a simulation environment, the algorithm learns to dynamically adjust the weights of various cost functions in the TEB local planner based on real-time perceived local environmental states (such as obstacle distance, path deviation, dynamic target density, etc.). The trained policy network is then deployed to the actual vehicle, forming an intelligent decision node called adaptive_planner. When this node runs online, it can automatically and in real time fine-tune local planning behavior according to actual road conditions, enabling the robot to make more flexible and efficient local motion decisions in dynamic and complex hospital scenarios while ensuring safety.

[0046] The specific embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A layered obstacle avoidance method for mobile robots in hospital settings, characterized in that, Includes the following steps: Step 1: Construct a hybrid map containing global sparse features and local dense information using visual SLAM; Step 2: Based on the A* algorithm, plan a global path from the starting point to the target point on the hybrid map; Step 3: Based on the Time Elastic Band (TEB) algorithm, perform real-time trajectory optimization and dynamic obstacle avoidance on the global path in the local layer of the hybrid map; Step 4: Based on the Twin Delayed Deep Deterministic Policy Gradient (TD3) reinforcement learning algorithm, dynamically adjust the cost function weights of the TEB algorithm according to the real-time perceived environmental state to achieve a balance between path tracking, obstacle avoidance safety and motion efficiency.

2. The method according to claim 1, characterized in that, The visual SLAM uses the ORB-SLAM3 framework to achieve global sparse localization and local dense mapping.

3. The method according to claim 1, characterized in that, In the step of dynamically adjusting the cost function weights, the action output of the TD3 algorithm is directly mapped to the weight coefficients of the constraint terms such as path following, obstacle avoidance, speed and acceleration in the TEB algorithm.

4. The method according to claim 1, characterized in that, The state inputs of the TD3 reinforcement learning algorithm include at least one of the following: the distance to the nearest obstacle, the robot's linear velocity and angular velocity, the lateral deviation from the global path, and the density of surrounding dynamic obstacles.

5. The method according to claim 1, characterized in that, The reward function of the TD3 reinforcement learning algorithm is a composite function, which includes at least: a positive reward for successfully reaching the target point, a negative penalty for a collision, and a progressive reward to encourage the robot to move closer to the target and maintain a safe distance.

6. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the program is executed by the processor, it implements the method as described in any one of claims 2 to 5.

Images