Small robot with intelligent anti-navigation function

Abstract

The invention provides a small robot with intelligent anti-navigation, and the method comprises the steps: collecting a fisheye image or video stream of an indoor environment in real time, and obtaining wide-field visual data close to the ground; inputting each frame of collected image picture into a YOLO model, identifying a target position, a static obstacle and dynamic interference, and outputting an image with a marked target category and position; the NoMaD model analyzes the distortion characteristics and environment characteristics of the fisheye image in real time, and locates the relative relation between the robot and the target; setting a final position, an anti-capture target and an indoor space constraint, dynamically generating a non-fixed optimal path, and avoiding an unreachable area; and generating a non-fixed optimal avoidance path, preventing the captured party from pre-judging the trajectory, and outputting an image or video stream marking a planned route and a robot motion control message stream. The technical problems that in the prior art, robot navigation needs to scan the environment in advance or depends on manual marking, a dynamic unknown environment cannot be handled, and the capability limitation of a robot body is neglected are solved.

Description

Technical Field

[0001] This invention relates to the field of robot navigation technology, and more specifically to a small robot with intelligent anti-navigation capabilities. Background Technology

[0002] Traditional robot navigation relies on SLAM (Simultaneous Localization and Mapping) or pre-planned paths. It constructs an environmental map using LiDAR / visual sensors and then uses path algorithms to avoid obstacles. However, it suffers from core defects such as passive adaptability, single-dimensional target orientation, and lack of reverse cognition.

[0003] 1. Passive adaptability: Requires pre-scanning of the environment or reliance on manual labeling, unable to cope with dynamic and unknown environments; 2. Single-dimensional goal orientation: Only optimizing the "shortest path" or "minimum energy consumption" while ignoring the limitations of the robot's own capabilities leads to failure in complex terrain due to mechanical configuration mismatch; 3. Lack of reverse cognition: Existing technology only solves the problem of "how to reach the goal", but cannot answer the question of "which areas are inaccessible to the robot", resulting in the inability to self-heal when the mission fails. Summary of the Invention

[0004] The purpose of this invention is to overcome the above-mentioned technical deficiencies and provide a small robot with intelligent anti-navigation capabilities, which solves the technical problems in the prior art where robot navigation requires pre-scanning of the environment or relies on manual marking, cannot cope with dynamic unknown environments, and ignores the limitations of the robot's own capabilities.

[0005] To achieve the above-mentioned technical objectives, in a first aspect, the present invention provides a robot with embodied intelligent anti-navigation capabilities, employing a robot navigation method with embodied intelligent anti-navigation capabilities, characterized by comprising the following steps: The low-angle wide-angle camera captures fisheye images or video streams of the indoor environment in real time, obtaining wide-field visual data close to the ground. Each frame of the captured image is input into the YOLO model to identify the target location, static obstacles, and dynamic interference, and outputs an image with the target category and location marked. The NoMaD model analyzes the distortion characteristics and environmental features of fisheye images in real time to determine the relative relationship between the robot and the target. Set the endpoint location, "anti-capture" target, and indoor space constraints to dynamically generate non-fixed optimal paths and avoid unreachable areas; It generates non-fixed optimal avoidance paths to prevent the target from predicting the trajectory and outputs two types of core results: first, an image or video stream marking the planned route to intuitively present the direction of travel; and second, a robot motion control message stream.

[0006] Compared with the prior art, the beneficial effects of the present invention include: This invention overcomes the limitations of traditional navigation systems that rely on external sensors such as LiDAR, UWB, or GPS, achieving purely vision-driven low-view navigation and anti-navigation. It enables bidirectional intelligent behavior in both navigation and anti-navigation, introducing an anti-navigation mechanism to give the robot recognition capabilities. Simultaneously, it employs low-power, high-efficiency embedded vision algorithms, optimizing the algorithm structure to address the robot's computational limitations, allowing it to run in real-time on a micro-embedded platform, reducing reliance on external computing power and offering high deployment flexibility. Referring to biomimetic behavior-inspired motion control strategies and drawing on the rapid exploration and escape behavior models of small animals, it combines visual feedback to form a non-deterministic maneuvering strategy, effectively avoiding predicted paths or fixed navigation patterns and improving the success rate of anti-navigation. Furthermore, it constructs an extensible anti-navigation learning framework, enabling the robot to autonomously optimize its anti-navigation strategy during training through machine learning, achieving adaptive anti-interference and experience transfer capabilities.

[0007] According to some embodiments of the present invention, before real-time acquisition of fisheye images or video streams of an indoor environment, the following steps are included: Using the video data collected by the low-angle wide-angle camera as input, the NoMaD model is transferred and trained, enabling the NoMaD model to accurately learn obstacles and passable areas in indoor scenes.

[0008] According to some embodiments of the present invention, after outputting the robot motion control message stream, the following steps are included: The robot's motion system receives the control message stream output by the NoMaD model, drives the actuator to complete the corresponding actions, and realizes path finding and obstacle avoidance.

[0009] According to some embodiments of the present invention, circumventing inaccessible areas includes the following steps: Input the robot's body parameters into the NoMaD model, and mark the robot's inaccessible areas based on obstacles in the indoor scene.

[0010] According to some embodiments of the present invention, the robot's body parameters include: robot size, robot endurance, minimum turning radius, maximum travel speed, acceleration / braking response time, obstacle crossing height, and climbing angle threshold.

[0011] According to some embodiments of the present invention, the control message flow includes: left turn command, right turn command, acceleration command, deceleration command, and steering command.

[0012] According to some embodiments of the present invention, the mechanical mouse utilizes the ROS v2 mechanical control framework to drive the actuator to complete corresponding actions, including the following steps: During navigation, the ROS v2 decision module publishes a topic containing standard messages about linear velocity and angular velocity; The control conversion node subscribes to the topic and uses the differential drive kinematic model to solve the abstract command into specific left and right wheel speeds; The specific left and right wheel speeds are sent to the underlying motor driver via a hardware interface, driving the actual motor to move.

[0013] In a second aspect, the present invention provides a robot with embodied intelligent anti-navigation, which applies the robot navigation method with embodied intelligent anti-navigation as described in any one of the first aspects.

[0014] Additional aspects and advantages of the invention will be set forth in part in the description which follows, and in part will be obvious from the description, or may be learned by practice of the invention. Attached Figure Description

[0015] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the description of the embodiments taken in conjunction with the following drawings, wherein the abstract drawings are to be completely consistent with one of the drawings in the specification: Figure 1 This is a schematic diagram of the operation process of an intelligent robot under normal conditions, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the operation process of an intelligent robot under abnormal conditions, provided in one embodiment of the present invention. Figure 3 This is a schematic diagram of key working nodes of an intelligent miniature robot provided in one embodiment of the present invention. Detailed Implementation

[0016] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the invention.

[0017] It should be noted that although functional modules are divided in the system diagram and the logical order is shown in the flowchart, in some cases, the steps shown or described may be performed in a different order than the module division in the system or the order in the flowchart. The terms "first," "second," etc., in the specification, claims, and the aforementioned drawings are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence.

[0018] This invention proposes an embodied intelligent anti-navigation technology, breaking through the traditional "map-driven" passive navigation mode and pioneering a "capability-driven" bidirectional coupling mechanism. This improves upon the limitations of wide-angle lenses used by robots close to the ground, which suffer from rich near-field detail but distortion at distant points, and a wide global field of view but difficulty in determining depth of field. By inputting robot body parameters in reverse, the technology prioritizes accessibility assessment, pre-excluding unreachable areas and reducing collisions in task planning. Simultaneously, an offline "anti-navigation library" is built to identify unreachable areas, significantly improving the success rate and safety of operations in dynamic and unknown environments.

[0019] Reference Figures 1 to 3 , Figure 1 This is a schematic diagram of the operation process of an intelligent robot under normal conditions, according to an embodiment of the present invention. Figure 2 This is a schematic diagram of the operation process of an intelligent robot under abnormal conditions, provided in one embodiment of the present invention. Figure 3 This is a schematic diagram of key working nodes of an intelligent miniature robot provided in one embodiment of the present invention.

[0020] In one embodiment, the robot with embodied intelligent anti-navigation has a low-angle wide-angle camera installed at its front end and applies an embodied intelligent anti-navigation robot navigation method, including the following steps: real-time acquisition of fisheye images or video streams of the indoor environment through the low-angle wide-angle camera to obtain wide-field visual data close to the ground; inputting each frame of the acquired image into the YOLO model to identify the target position, static obstacles, and dynamic interference, and outputting an image with marked target category and position; the NoMaD model analyzes the distortion characteristics and environmental features of the fisheye image in real time to locate the relative relationship between the robot and the target; setting the endpoint position, the "anti-capture" target, and indoor space constraints, dynamically generating a non-fixed optimal path to avoid inaccessible areas; generating a non-fixed optimal avoidance path to prevent the captured party from predicting the trajectory, and outputting two types of core results: one is an image or video stream marking the planned route, intuitively presenting the direction of travel; the other is a robot motion control message stream.

[0021] 1. YOLO detection algorithm: YOLO is a single-stage detection algorithm that directly detects and predicts the location and category of multiple targets from the entire image at once.

[0022] Unlike traditional object detection, the YOLO lightweight detection algorithm used in this application requires fewer model parameters and consumes less computational resources, while maintaining accuracy despite high running speed. Compared to other models, YOLO is suited to the computational needs of small robots and is ideal for real-time deployment on the terminal.

[0023] Transfer training is performed on the official pre-trained model to improve its accuracy in detecting targets. Each frame of the input image or video stream is used as input to the transferred-trained model, and the output box selects the target location and labels the image or video stream with the corresponding target category.

[0024] This experimental model, designed for real-world environments, explores the application of embodied intelligence in small robots. Due to their small size, these robots possess a low field of view and a wide field of vision. Using this algorithm with purely visual input, the robot can simulate a real biological mouse, continuously interacting and coupling its detection results and subsequent responses with the environment, thus realizing the application of embodied intelligence in small robots.

[0025] 2. NoMaD Navigation Algorithm. NoMaD (Navigation with Goal Masked Diffusion) is an automated navigation system based on reinforcement learning and optimization algorithms, designed to solve automated path planning and decision-making problems in complex and dynamic environments. The core model is the "diffusion model," and the key technique is "goal masking," enabling the robot not only to move towards a clear goal but also to effectively explore when the goal is unclear. The system can intelligently select the optimal path without human intervention and effectively cope with unknown or changing environmental challenges. The application of the NoMaD algorithm to this intelligent robot primarily involves adapting to indoor scenes through transfer training and combining real-time visual input to achieve autonomous navigation, obstacle avoidance, and anti-capture decision-making.

[0026] For optimization of robot transfer training, transfer training was conducted based on the NoMaD official open-source pre-trained model. The training data focused on indoor environmental characteristics, combining the imaging distortion characteristics and field of view of the robot's fisheye camera to enable the model to accurately learn key information such as obstacles and passable areas in indoor scenes. The training objective was clearly aimed at "anti-capture" requirements, not only optimizing the path planning capabilities of conventional navigation and obstacle avoidance but also strengthening the model's perception of dynamic threats and non-deterministic path decisions, adapting to the flexible maneuverability required for anti-capture.

[0027] Real-time fisheye image analysis: The robot acquires real-time images or video streams solely through a fisheye camera, using each frame containing the target location (such as the escape endpoint) as input to the NoMaD model after transfer training. The core module of the NoMaD algorithm analyzes fisheye images in real time, leveraging the environmental awareness learned through machine learning to quickly identify static obstacles and dynamic disturbances in the indoor environment, while simultaneously determining its relative position to the target, without relying on external sensors such as LiDAR or GPS.

[0028] The NoMaD algorithm, which focuses on anti-capture-oriented path planning and decision output, dynamically adjusts the path based on preset anti-capture targets and constraints, such as indoor space boundaries and obstacle limitations, combined with real-time visual feedback. This enables the robot to make autonomous decisions, generating non-fixed optimal avoidance paths to prevent the target from predicting its trajectory. It also outputs two core results: first, an image / video stream marking the planned route, visually presenting the direction of travel; and second, a robot motion control message stream, including specific commands such as left, right, acceleration, and turning.

[0029] Embodied Intelligence in Practice: Motion Control and Real-Time Adaptation. The robot's motion system receives control message streams from NoMaD, driving actuators to complete corresponding actions and achieving pathfinding and obstacle avoidance. Continuous Optimization Through Real-Time Feedback Closed-Loop: Environmental changes in each frame of the fisheye camera image, such as changes in the tracking source's position or sudden obstacles, are re-inputted into the model. NoMaD rapidly iterates its decisions and dynamically adjusts motion commands, ensuring continuous anti-capture navigation capabilities in complex indoor environments.

[0030] 3. The ROS v2 framework for mechanical control: ROS v2 is an open-source framework designed specifically for robot control that requires high real-time performance and high reliability. It is especially important in the field of mechanical control in complex environments with multiple tasks and multiple nodes.

[0031] Unlike other robot control software frameworks, ROS, with its node-based distributed computing architecture, has unique advantages in mechanical control.

[0032] After the camera is triggered, the ROS v2 topic publishes the raw image for YOLO to detect and achieve perception. During navigation, the ROS v2 decision module publishes a topic containing standard messages including linear and angular velocities; subsequently, a dedicated control conversion node subscribes to this topic and uses kinematic models such as differential drive to solve the abstract commands into specific left and right wheel speeds; finally, the converted commands are sent to the underlying motor driver through a hardware interface to drive the actual motor movement, completing a reliable closed-loop control from high-level decision-making to low-level execution.

[0033] Real-time constraint adjustment of motion addresses the challenges of low stability and poor anti-interference capabilities in small robots due to their small size and mass, which are difficult to overcome in conventional designs. To address these difficulties, dynamic feasibility constraints are introduced into the mechanical control based on detection results using modules such as the local planner in ROS v2 to update and optimize the path. Real-time fusion of sensor data, combined with detection algorithms, dynamically modifies the path to provide environmental constraints. This real-time adjustment of velocity and path resolves the sacrifice of complex environmental adaptability inherent in one-dimensional goal-oriented robots, while reducing the impact of uncertainties in the robot's parameters. This achieves the high degrees of freedom and flexibility required for small robots while also ensuring stability and anti-interference capabilities.

[0034] The robot employs a low-angle, wide-field camera mounted on its front end to capture environmental visual data close to the ground. Unlike traditional high-angle cameras and multi-sensor fusion solutions, this approach relies on image input for autonomous localization, path planning, and obstacle avoidance without depending on LiDAR or GPS. The low-angle image acquisition enhances its ability to recognize complex environments, obstacle details, and shadow variations. An intelligent anti-navigation strategy mechanism allows the system to generate an avoidance path based on visual scene reconstruction and trajectory anti-prediction algorithms when external signal interference is detected, enabling proactive countermeasures and escape from externally induced navigation.

[0035] The beneficial effects of this embodiment: 1. Fewer collisions: Pre-simulation replaces real trial and error, and sets up handling procedures for abnormal working conditions, greatly reducing the occurrence of drops and collisions; 2. Short recalculation time: "Capability-driven" decision-making directly removes inaccessible areas from the map, reducing route planning time; 3. Small blind spot range: The offline anti-navigation library continuously learns from various cases, and the robot has a long-term memory of places it cannot pass through, so it will no longer accidentally enter them when performing repetitive tasks; 4. Fast response speed: Rapid judgment allows the robot to brake suddenly in the event of a change in terrain, shortening the braking distance, improving the task completion rate and reducing energy consumption under the same computing power.

[0036] In one embodiment, a robot navigation method with integrated intelligent anti-navigation, wherein a low-angle wide-angle camera is installed at the front end of the robot, includes the following steps: using video data collected by the low-angle wide-angle camera as input to perform transfer training on a NoMaD model, enabling the NoMaD model to accurately learn obstacles and passable areas in an indoor scene; acquiring fisheye images or video streams of the indoor environment in real time through the low-angle wide-angle camera to obtain wide-field visual data close to the ground; inputting each frame of the acquired image into a YOLO model to identify the target position, static obstacles, and dynamic interference, and outputting an image with labeled target category and position; the NoMaD model analyzing the distortion characteristics and environmental features of the fisheye image in real time to determine the relative relationship between the robot and the target; setting the endpoint position, the "anti-capture" target, and indoor space constraints, dynamically generating a non-fixed optimal path to avoid inaccessible areas; generating a non-fixed optimal avoidance path to prevent the captured party from predicting the trajectory, and outputting two types of core results: one is an image or video stream that marks the planned route, intuitively presenting the direction of travel; the other is a robot motion control message stream.

[0037] Furthermore, after outputting the robot motion control message stream, the process includes the following steps: the robot's motion system receives the control message stream output by the NoMaD model, drives the actuator to complete the corresponding action, and realizes path finding and obstacle avoidance.

[0038] The process of avoiding inaccessible areas includes the following steps: inputting robot body parameters into the NoMaD model, and marking inaccessible areas of the robot based on obstacles in the indoor scene. The robot body parameters include: robot size, robot endurance, minimum turning radius, maximum travel speed, acceleration / braking response time, obstacle crossing height, and climbing angle threshold.

[0039] Based on parameters such as robot size and minimum turning radius, it can accurately determine whether areas such as narrow passages and sharp turns are passable, avoiding the planning of "theoretically feasible but practically impassable" paths and reducing the risk of collisions and getting stuck. The obstacle-crossing height and climbing angle thresholds are directly related to terrain adaptability, marking inaccessible areas such as steps and steep slopes in advance to prevent robots from falling or being damaged due to forced passage, thus reducing the mission failure rate.

[0040] By pre-defining reachable boundaries using ontological parameters, the NoMaD model eliminates the need for path calculations in areas beyond its capabilities, directly removing inaccessible regions, reducing path recalculation time, and improving navigation response speed. By combining range, maximum speed, and acceleration / braking response time, path planning can balance efficiency and energy consumption, avoiding the planning of long-distance paths exceeding range or the energy waste caused by frequent acceleration and braking, thus preventing extended operation time.

[0041] By incorporating ontological parameters as fixed constraints into the model, NoMaD can quickly identify unreachable areas in new indoor environments, such as changing obstacle positions or temporary passage closures, without requiring re-adaptation of robot capabilities, resulting in more efficient decision-making. To meet anti-navigation requirements, in non-fixed path planning, the "capability boundaries" defined by ontological parameters prevent the robot from falling into traps that appear escapeable but are actually impassable, improving the success rate of anti-capture and anti-interference.

[0042] The control message flow includes leftward commands, rightward commands, acceleration commands, deceleration commands, and turning commands. The robotic mouse utilizes the ROS v2 mechanical control framework to drive the actuators to complete corresponding actions. The steps include: during navigation, the ROSv2 decision module publishes a topic containing standard messages of linear and angular velocities; the control conversion node subscribes to this topic and uses a differential drive kinematic model to calculate the abstract commands into specific left and right wheel speeds; these specific left and right wheel speeds are sent to the underlying motor driver via a hardware interface to drive the actual motor movement. The standard message topics of linear and angular velocities provide a unified benchmark for control, avoiding command ambiguity and ensuring consistency between decision intent and execution action. The differential drive model specifically calculates the left and right wheel speeds, perfectly matching directional commands such as "left / right / turn" and speed commands such as "accelerate / decelerate," ensuring precise robot action response without stuttering or deviation.

[0043] In one embodiment, the robot with embodied intelligent anti-navigation uses the robot navigation method with embodied intelligent anti-navigation as described above.

[0044] The above is a detailed description of the preferred embodiments of the present invention. However, the present invention is not limited to the above embodiments. Those skilled in the art can make various equivalent modifications or substitutions without departing from the spirit of the present invention. All such equivalent modifications or substitutions are included within the scope defined by the claims of the present invention.

[0045] The specific embodiments of the present invention described above do not constitute a limitation on the scope of protection of the present invention. Any other corresponding changes and modifications made in accordance with the technical concept of the present invention should be included within the scope of protection of the claims of the present invention.

Claims

1. A small robot with body-aware anti-navigation, the front end of the robot is equipped with a low-view wide-angle camera, and the robot navigation method using body-aware anti-navigation is characterized in that, The method comprises the steps of: Real-time collection of fisheye images or video streams of indoor environment by the low-view wide-angle camera to obtain wide-view visual data close to the ground; Input of each collected image frame into a YOLO model to identify target position, static obstacles and dynamic interference, and output of images with marked target categories and positions; Real-time analysis of distortion characteristics and environmental features of the fisheye images by the NoMaD model to locate the relative relationship between the robot and the target; Setting of end position, "anti-capture" target and indoor space constraints, dynamic generation of a non-fixed optimal path, and avoidance of unreachable areas; Generation of a non-fixed optimal avoidance path to avoid the prediction of the trajectory by the capturing party, and output of two types of core results: one is an image or video stream marking the planned route to intuitively present the direction of travel, and the other is a robot motion control message stream.

2. The body-aware backchannelling robot of claim 1, wherein, Before real-time collection of fisheye images or video streams of the indoor environment, the method comprises the steps of: Using video data collected by the low-view wide-angle camera as input to perform transfer training on the NoMaD model, so that the NoMaD model accurately learns the obstacles and passable areas of the indoor scene.

3. The method of claim 2, wherein, After outputting the robot motion control message stream, the method comprises the steps of: The motion system of the robot receives the control message stream output by the NoMaD model to drive the execution mechanism to complete corresponding actions and achieve path finding and obstacle avoidance.

4. The body-aware backchannelling robot of claim 2, wherein, Avoidance of unreachable areas, comprising the steps of: Input of robot body parameters into the NoMaD model and marking of the unreachable areas of the robot based on the obstacles of the indoor scene.

5. The body-aware anti-navigating robot of claim 4, wherein, The body parameters of the robot include robot size, robot endurance, minimum turning radius, maximum travel speed, acceleration / deceleration response time, obstacle climbing height and climbing angle threshold.

6. The body-aware backchannelling robot of claim 2, wherein, The control message stream includes left instruction, right instruction, acceleration instruction, deceleration instruction and steering instruction.

7. The body-aware backchannelling robot of claim 2, wherein, The mechanical mouse applies a mechanical control ROS v2 framework to drive the execution mechanism to complete corresponding actions, comprising the steps of: In the execution of navigation, the decision module of ROS v2 publishes a standard message topic containing linear velocity and angular velocity; The control conversion node subscribes to the topic and uses a differential drive kinematics model to calculate specific left and right wheel speeds from abstract commands; The specific left and right wheel speeds are sent to the underlying motor driver through a hardware interface to drive the real motor to move.

Images