A method for navigation of legged robots targeting objects within a custom scene

Abstract

This invention discloses a method for bipedal robot navigation targeting objects within a custom scene, belonging to the field of robotics. The method includes: constructing a simulation environment highly consistent with the local real-world scene, and performing data acquisition and model training within this environment; recognizing objects through semantic segmentation and target discrimination; designing an adaptive switching mechanism between exploration mode and target-oriented navigation mode; generating continuous navigation trajectories based on a diffusion navigation decision model; and driving the robot to perform movements using a reinforcement learning gait control method. This invention also relates to supporting methods for simulation scene construction, data acquisition, segmentation model training, and bipedal robot control. This invention can improve the robot's navigation capability for specific objects within a local custom scene without relying on large-scale labeled data, and is suitable for autonomous navigation tasks in complex environments such as homes and offices.

Description

Technical Field

[0001] This invention belongs to the field of robotics technology, specifically relating to a method for navigating a legged robot targeting objects within a custom scene. Background Technology

[0002] Navigation is a fundamental capability required by robots. Object navigation refers to the specific instance where a robot searches for and reaches a target object in its environment, using the object's name as the navigation target information. This task integrates multiple capabilities such as environmental perception, semantic understanding, and path planning, requiring the robot to continuously explore unknown environments and accurately identify the target object when it approaches it.

[0003] Existing object navigation methods are mainly divided into two categories: end-to-end learning methods and modular methods. End-to-end learning methods typically use imitation learning or reinforcement learning to directly map the robot's observation information into control actions, and use large-scale data training to enable the robot to learn to complete object navigation tasks. However, this method usually faces two problems: first, navigation targets are often limited to categories predefined in the training phase; second, there are differences between the training environment and the actual application environment, and navigation performance is prone to degrade when the robot performs tasks in new scenarios. To improve the generalization ability of the model, it is often necessary to build a larger-scale dataset and increase the number of training scenarios, which will significantly increase the cost of data annotation and training resources.

[0004] Recent research has attempted to incorporate large-scale language models to assist robot navigation decisions. These methods typically utilize visual language models to semantically recognize target objects and rooms in the current environment, converting the perception results into text descriptions. These descriptions are then input into large-scale language models for common-sense reasoning to predict the likely locations of target objects, thereby guiding the robot's next action. While this approach overcomes the limitations of traditional methods on fixed target categories to some extent, it often incurs high computational costs due to the complex language reasoning process involved, making it difficult to meet the real-time requirements of navigation tasks.

[0005] Modular approaches typically preserve a geometric map rich in semantic detail. This method leverages a top-down semantic map constructed from environmental observations as its foundation, learning prior knowledge through a model to predict the reachability relationships or potential exploration boundaries between each location on the map and the target. Subsequently, a navigation strategy is formulated based on the predictions, generating directional long-term goal points to guide the agent in efficiently exploring the environment and approaching the target object. The core of this method lies in combining semantic information with spatial reasoning, optimizing path planning by predicting the correlation between unknown regions and the target. Unlike end-to-end methods, modular approaches decompose the object navigation task into modules such as mapping, long-term goal strategy, and path planning.

[0006] However, existing methods rely on large amounts of training data to achieve general navigation performance across different navigation scenarios. These models still face challenges when dealing with local scenarios, especially when training data is limited or scene structures differ significantly, leading to a decline in robot navigation performance. Therefore, developing robot navigation methods that can adapt to local scenarios has become a pressing technical problem. Summary of the Invention

[0007] To address the aforementioned technical problems, this invention provides a legged robot navigation method for objects within a custom scene. By constructing a simulation environment highly consistent with the local real scene, data is collected in this environment and a segmentation model and a navigation decision model are trained. This enables the legged robot to adaptively complete the navigation task for specific objects within a custom scene without the need for large-scale labeled data.

[0008] To achieve the above objectives, the present invention adopts the following technical solution:

[0009] A first aspect of the present invention provides a method for navigation of a legged robot oriented towards a custom scene object target, the method comprising:

[0010] Step 1: Semantic perception and target discrimination;

[0011] The robot acquires perceptual information from RGB and depth images using a camera. A pre-trained semantic segmentation model then performs semantic parsing on this information, extracting pixel-level semantic distributions of various objects within its field of view. Combined with the input target semantic category information, the robot determines the existence of target objects within the current observation space. Subsequently, this pixel-level positional distribution is mapped to a unified intermediate state representation, implicitly containing the target object's positional distribution information in both image and 3D space, thus enabling the conversion between semantic input and navigation decision model input.

[0012] Step 2: Adaptive mode switching mechanism;

[0013] Based on the target object identification results, a switching mechanism is constructed between exploration mode and target-oriented navigation mode. When the target object is not detected, the system drives the robot to perform exploratory navigation, focusing on avoiding obstacles while expanding the observable space coverage to find the navigation target; when the target object is detected, target-oriented navigation is triggered, generating spatial representation information of the target object based on the segmentation results, and navigating to the vicinity of the designated object target.

[0014] Step 3: Generation of perception-driven navigation strategies;

[0015] When in target-oriented navigation mode, the intermediate state output in step one is used as the navigation target condition. Combined with the pixel-level semantic distribution output in step two, a diffusion navigation decision model based on multimodal perception information is constructed. The spatial relationship between the robot's current state and the target state is modeled using fused RGB image and depth image information to generate a continuous navigation trajectory to reach the target object.

[0016] Step 4: Motion control and task determination mechanism;

[0017] The continuous navigation trajectory generated by the navigation strategy is converted into speed commands and input into the legged robot motion control model. The model generates control commands for the robot's joint movements and constructs a task judgment criterion based on target distance or spatial constraints. When the robot's state meets the preset target approach conditions, the task is judged as successful; otherwise, the task is judged as failed.

[0018] Preferably, the present invention provides a method for constructing a custom scene in a simulation environment, the method comprising:

[0019] By performing proportional constraint modeling on the geometric structure of a 3D scene and combining it with an object model library to quickly generate a basic scene, the scene is further optimized for texture attributes and enhanced for visual details. Based on this, each object in the scene is structurally decomposed and independently identified, and the processed scene is converted into a unified standard format to adapt to the simulation platform. In the simulation environment, the geometric attributes, material attributes, and hierarchical relationships of the scene are consistently mapped, and physical attribute parameters are introduced for each object to achieve realistic physical interaction modeling. Simultaneously, lighting and perception-related parameters are optimized, thereby constructing a robot navigation simulation scene that combines geometric consistency, semantic resolvability, and physical interactivity.

[0020] Preferably, the present invention provides a data acquisition method based on a custom scenario in a simulation environment, the method comprising:

[0021] By discretizing the 3D space of the scene and extracting passable areas based on robot structural parameters, and expanding obstacle areas according to kinematic constraints to form safe space markers, a set of initial and target states is constructed using a random sampling strategy. This, combined with the spatial distribution information of the target object, generates a navigation task oriented towards semantic objectives. Furthermore, path planning and trajectory continuity methods are used to generate navigation trajectories that satisfy motion constraints, and redundant sampling and failure compensation mechanisms are employed to improve the success rate of trajectory generation. During simulation execution, multimodal data, including robot motion state, visual perception information, and target semantic information, are collected simultaneously. Trajectory consistency evaluation is used to filter the data, resulting in a high-quality robot navigation dataset.

[0022] Preferably, the present invention provides a segmentation model training method based on a custom scene in a simulation environment, the method comprising:

[0023] Based on the constructed 3D simulation scene with semantic structural information and the corresponding navigation trajectory data, a scene-aware dataset in a unified format is generated. This dataset constructs pixel-level annotations that simultaneously contain category and instance-distinguishing information by semantically classifying and encoding each object in the scene, achieving a refined representation of target objects in space. Furthermore, by establishing a mapping relationship between the input perception data and the corresponding annotations, the model is trained using supervised learning, enabling it to achieve semantic recognition and spatial localization of target objects under specific scene distributions.

[0024] Furthermore, a data closed-loop optimization mechanism oriented towards performance feedback is introduced. Based on the model's performance on low-performance instances, targeted data sampling and enhancement are adaptively performed in the simulation environment. The data distribution is dynamically supplemented and adjusted through an iterative process of "data construction - model training - performance evaluation - data reconstruction". This breaks through the traditional static dataset training mode, realizes the co-evolution of data distribution and model learning state, and improves the robustness and generalization ability of the model in complex scenarios.

[0025] Preferably, the present invention provides a reinforcement learning gait control method based on a bipedal robot platform, the method comprising:

[0026] A 3D model of the bipedal robot's physical structure was created using CAD software, and the model file was exported. A robot description file containing links, joints, and their inertial, visual, and collision attributes was written, constructing a simulation training platform readable by the training environment. Based on this, a reinforcement learning training mechanism was built using the policy gradient theorem. A clipping mechanism was introduced to constrain the policy update magnitude, and an optimization value function was used to improve the accuracy of state value estimation. An entropy regularization term was added to encourage policy exploration. In a simulation environment with uneven ground, the neural network model was trained through a cyclical process of "interactive data collection - calculation of the advantage function - sampling and updating parameters," enabling stable gait control in the simulation environment. After training, the policy model from the simulation environment was converted into a policy model suitable for real-world deployment. The policy was loaded onto the real robot, and the sensor and motor control SDK was called to achieve real-world control of the robot's standing, forward movement, and turning gait actions.

[0027] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for navigating a legged robot oriented towards an object target within a custom scene.

[0028] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for navigating a legged robot oriented towards an object target within a custom scene.

[0029] The beneficial effects of this invention are as follows:

[0030] This invention provides a legged robot navigation method for objects in a custom scene. By constructing a simulation environment that is highly consistent with the local real scene, and completing data acquisition, segmentation model training and navigation strategy learning in this environment, the robot can adaptively complete the navigation task of objects in a specific scene without relying on large-scale manually labeled data, which significantly reduces the dependence on general training datasets and labeling costs.

[0031] This invention introduces an adaptive switching mechanism between exploration mode and goal-oriented navigation mode, combined with a trajectory generation method based on a diffusion model. This enables efficient environmental exploration even in the absence of a target, and rapidly generates diverse candidate navigation trajectories after a target appears, improving the flexibility and robustness of navigation decisions. Furthermore, by decoupling high-level trajectory planning from low-level reinforcement learning gait control, a complete link from semantic goals to joint control commands is achieved, ensuring the motion stability of the legged robot in complex environments.

[0032] Furthermore, this invention overcomes the limitations of traditional methods in terms of poor generalization ability and restricted navigation target categories in local scenarios by using structured decomposition of simulation scenarios, data closed-loop optimization, and targeted training of segmentation models. Attached Figure Description

[0033] Figure 1 This is an overall flowchart of a legged robot navigation method for objects in a custom scene according to the present invention.

[0034] Figure 2 This is a diagram of the overall model architecture of a legged robot navigation method for objects in a custom scene, as described in this invention.

[0035] Figure 3 This is a diagram of the high-level navigation model architecture based on the diffusion model in this invention;

[0036] Figure 4 This is a diagram of the motion control model architecture for a bipedal robot based on reinforcement learning in this invention. Detailed Implementation

[0037] The present invention will be further described below with reference to the accompanying drawings and embodiments.

[0038] Object navigation integrates perception and cognition processes. The agent needs to continuously explore the environment, plan the optimal path to the target, and accurately identify the target object upon approach. Therefore, object navigation is widely considered a challenging task. This invention focuses on its applicability to local scenarios, offering a different approach to address the poor generalization ability and ineffective performance of existing object navigation tasks in local environments.

[0039] To make the solution provided in this example clearer, combined with Figure 1 The various steps in this method are described in detail. This invention primarily focuses on the applicability of object navigation tasks to local scenarios, where the robot needs to complete the navigation task using given semantic object targets. To achieve this, a local scenario is first constructed in a simulation environment, and data is collected within the constructed scenario for model training, helping the robot better understand the information in the local scenario. Simultaneously, segmentation model data is collected and trained within the scenario to improve the segmentation model's accuracy in recognizing objects in the local scenario.

[0040] S1, Based on real-world information from the local scene, construct a simulation environment that is as close as possible to the actual scene, for use in subsequent training data collection and as an evaluation inference environment:

[0041] First, based on the spatial layout information of the real-world scene, an initial 3D scene is constructed using 3D modeling tools. By calling the object model library built into the modeling tools, the basic structure of the indoor environment is built, including walls, tables, chairs, cabinets, and other common furniture. During the construction process, proportional constraints are applied to the size and relative positional relationships of each object in the scene to ensure that the simulation scene maintains consistency with the real environment in terms of geometric scale and spatial distribution, thereby ensuring that the spatial features learned by the robot in the simulation environment have actual correspondence. For situations where the model library cannot provide object models that are completely identical to the real scene, a substitution modeling strategy based on geometric similarity is adopted. That is, models that are similar to the target object in terms of shape and spatial occupancy are selected for replacement. Simultaneously, differences in the appearance of the objects are compensated for through subsequent texture optimization processes to balance modeling efficiency and scene realism.

[0042] After completing the basic scene construction, the scene data is exported as a general 3D model format and imported into a 3D editing tool for further processing. Since the initially exported scene typically exists as a single mesh, it is necessary to perform structured decomposition on the different objects in the scene, dividing the overall model into multiple independent object units, and uniquely identifying and naming each object. This processing enables each object to be manipulated individually in subsequent stages and provides a foundation for the localization and identification of target objects in navigation tasks.

[0043] After object decomposition, the surface properties of each object are refined to improve the acquisition effect of the visual sensor. This process mainly includes UV mapping of the model, unfolding the 3D surface into a 2D texture space to establish the correspondence between geometric structure and texture coordinates; constructing object surface properties based on the material model, and further enhancing details by introducing texture images and adjusting their color, brightness, and detail information, using texture painting to directly modify local areas on the model surface, thereby improving the realism and diversity of the texture. After completing the above processing, the texture data is saved to ensure that the modified results can be correctly used in subsequent processes.

[0044] After optimizing the scene's geometry and texture, the processed 3D scene is imported into a simulation platform to construct a robot navigation environment with physical interaction capabilities. In the simulation environment, the material properties of all objects in the scene are first uniformly configured. By adjusting parameters such as base color, roughness, and metallicity, the scene's visual appearance is made more realistic. Simultaneously, scene resources are optimized, including mesh merging, material simplification, and instantiation management, to reduce rendering overhead and improve simulation efficiency. In terms of lighting modeling, a global illumination source is introduced to simulate ambient light distribution, ensuring uniform and natural lighting effects and improving the stability and realism of visual perception data.

[0045] Finally, to achieve physical interaction between the robot and its environment, physical property models are constructed for each object in the scene. Specifically, this includes adding rigid body properties to objects to enable them to respond to gravity and dynamic constraints, and configuring collider properties to allow them to collide with the robot or other objects during movement. For objects with relatively regular geometry in the scene, their colliders can be automatically generated based on their geometric shapes. Through this physical property modeling, the simulation scene is transformed from a static visual model into an interactive environment with realistic physical responses.

[0046] S2, based on the simulation environment built in S1, first constructs a grid map to estimate navigable and impassable areas, and then uses a classic navigation algorithm. Potential field navigation and The algorithm plans the navigation path, then deploys the pre-trained reinforcement learning model on the robot, and places the robot in a simulation environment to complete data collection.

[0047] First, a continuous spatial representation is discretized by constructing an occupancy grid map with a fixed voxel resolution, thereby achieving unified modeling of the environmental structure. The optimal voxel resolution is 0.05 meters, which strikes an effective balance between spatial representation accuracy and computational complexity, allowing for sufficient depiction of obstacle details while avoiding a significant increase in computational burden due to excessively high resolution. Building upon this, an Euclidean Signed Distance Field (ESDF) is introduced to further characterize the spatial structure. ESDF estimation is performed by calculating the minimum Euclidean distance from each voxel to the nearest obstacle boundary. Received, among which Indicates the current voxel's position coordinates. Represents any point on the boundary of the obstacle. Represents the set of all obstacle boundary points. Point With point The Euclidean distance between them. The distance field provides important cost information for subsequent path planning, enabling the planning algorithm to effectively avoid obstacles and maintain a safe distance.

[0048] Regarding spatial drivability determination, this invention introduces a constraint modeling mechanism related to the robot's structural characteristics. The sectional height is determined based on the robot's height parameter. Voxels with z-axis coordinates lower than the robot's height are defined as drivable regions, while voxels with z-axis coordinates higher than the robot's height are defined as obstacle regions. Furthermore, to further eliminate collision risks introduced by changes in the robot's size and posture during actual movement, a spatial expansion mechanism based on the robot's motion characteristics is constructed. By incorporating the robot's rotation radius into the environment model, the drivable region is constrained and contracted, and potentially interference-risk neighborhoods are uniformly mapped as obstacle regions. This expansion ensures that the robot will not collide with obstacles during rotation or turning.

[0049] For navigation task generation, a random sampling strategy is used to select the navigation starting point within the traversable area, and the navigation target point is set near the clustered target object to support the robot in performing object-oriented navigation tasks. The random sampling of the starting point is based on a uniform distribution function, generating candidate points within the two-dimensional projection range of the traversable area, and eliminating points whose distance from obstacles is less than a safety threshold to ensure the feasibility and safety of the initial pose. The setting of the target point needs to consider the spatial distribution characteristics of the target object: firstly, Euclidean clustering is performed on object instances of the same semantic category to obtain the center position of each instance. :

[0050] ,

[0051] in, Indicates the first The central location of each cluster This indicates the number of object instances contained in the cluster. Indicates the number of elements belonging to this cluster. The spatial coordinates of each object instance are then used for local sampling within a traversable area near the center, ensuring the target point is both close to the object and within robot reach. For multi-object scenarios, multiple candidate target points are generated for each instance to enrich the diversity of the training data.

[0052] In the trajectory generation stage, a heuristic search mechanism based on distance field information is introduced to achieve global path optimization by comprehensively considering path length and spatial safety. A set of coordinates is randomly selected from pre-sampled starting and target point pairs, and a discrete preliminary path is planned using a navigation algorithm. Its cost function can be expressed as: ,in This represents the actual path cost from the starting point to the current node. This represents a heuristic estimate of the distance from the current node to the target point. The cost map is derived from ESDF, and the cost of each grid cell is determined by its distance from obstacles, guiding the path to stay as far away from obstacles as possible. For each type of target object, a uniform sampling strategy is implemented to ensure a balanced number of data samples across different object categories, preventing the model from overfitting to common objects. Redundant sampling is also considered for potential path planning failures, with a maximum number of sampling attempts set. If a failure occurs, a new starting or target point is selected until a valid path is generated or the upper limit is reached. To address the potential discontinuities in discrete paths during actual execution, a trajectory smoothing method is used to convert discrete paths into continuous executable trajectories, ensuring they satisfy robot kinematic constraints and improving the stability of control execution.

[0053] After the navigation trajectory is constructed, the robot with the pre-trained control strategy is deployed in the simulation environment. The robot is driven by the strategy to perform navigation tasks, simultaneously collecting multimodal perception data. The collected data includes not only visual and depth information, but also the pose state of the robot and sensors, thus forming a complete data representation encompassing environmental perception and motion behavior. This data construction method can simultaneously reflect the coupling relationship of "perception-decision-execution," providing high-quality input for subsequent model training. By evaluating the consistency between the robot's actual execution trajectory and the reference trajectory, samples with large deviations are eliminated, thereby constructing a highly reliable dataset. This screening mechanism constrains the effectiveness of navigation behavior at the data level, avoiding the negative impact of low-quality samples on model training and further improving the overall system performance.

[0054] S3, Constructing a robot navigation system with a hierarchical decision-making framework, uses the data collected in S2 for training. The model construction and training process is as follows:

[0055] like Figure 2 The diagram illustrates the overall model architecture (high-level planning module + low-level control module). In terms of model architecture design, this invention proposes a robot navigation system based on a hierarchical decision framework, dividing the navigation process into two decoupled sub-processes: high-level planning and low-level execution. This hierarchical structure, by modeling decision problems at different levels, allows for independent optimization of global path generation and robot control execution, thereby reducing the complexity of policy learning and improving the system's generalization ability and stability in complex environments. The high-level planning module, based on the Transformer architecture, receives navigation targets and outputs candidate motion trajectories, embeds diffusion strategies to generate multimodal trajectory distributions, and uses an independent evaluation unit to perform quality screening and scoring of the generated trajectories. The system supports two navigation modes: targetless and target-oriented. Targetless navigation is used for autonomous exploration when the robot has not detected a navigation target. The low-level control module employs the Proximal Policy Optimization (PPO) algorithm, responsible for converting the planned trajectory into specific joint control commands for the robot, as detailed in section S5. In the inference phase, a pre-trained segmentation model is introduced to realize the conversion of semantic targets to pixel targets, and the actual planning is completed by combining two modes: targetless exploration and pixel target navigation.

[0056] Furthermore, such as Figure 3 The diagram illustrates the navigation model architecture. The high-level planning module uses RGB images, depth images, and navigation targets as multimodal inputs. Visual data is feature-encoded via a pre-trained DepthAnything module. Navigation targets are primarily guided by pixel-based navigation targets, with point targets introduced as supervisory signals during training. Pixel-based navigation targets are automatically calculated based on point targets and camera parameters. All features are fused across modalities and modeled sequentially by a Transformer encoder. For trajectory prediction, a diffusion model based on a conditional generation mechanism is used to model future motion trajectories. Specifically, by constructing a conditional U-Net architecture and combining it with a Denoising Diffusion Probabilistic Model (DDPM) scheduling strategy, the trajectory generation process is represented as an iterative denoising process that gradually evolves from a random noise distribution to a target trajectory distribution, thereby achieving the ability to model the multimodal distribution of the trajectory space. Compared to traditional trajectory prediction methods based on deterministic mappings, this method can characterize the uncertainty of path selection in complex environments and generate diverse candidate trajectories.

[0057] After obtaining the candidate trajectory set, an independent evaluation mechanism is introduced to assess the quality of each candidate trajectory. By constructing trajectory feasibility and safety constraints, the trajectory is comprehensively measured in terms of spatial accessibility, distance relationship with obstacles, and motion continuity. The generated candidate trajectories are then evaluated for safety and feasibility by the critic module, and a future trajectory sequence of length that meets the constraints is output.

[0058] To enhance the diversity and efficiency of training samples, this invention divides the original complete trajectory into multiple local subsequences by randomly sampling sub-trajectories, thereby achieving decomposition modeling of long-term navigation processes. This approach enables the model to learn trajectory evolution patterns at different time scales while avoiding strong dependence on the complete trajectory, thus improving data utilization and enhancing the model's generalization ability. Furthermore, the endpoint coordinates of the sub-trajectories are used as point target supervision signals for calculating trajectory loss changes. Further, to establish the correspondence between visual space and target space and align with the target segmentation model, this invention innovatively uses pixel targets. By mapping the relative pose changes of the endpoint coordinates at each time step relative to the robot's current position to the current RGB image plane, a corresponding pixel-level target representation is obtained. This method achieves a unified mapping mechanism from spatial targets to visual representations, enabling the model to perceive navigation targets in image space.

[0059] To support autonomous behavior modeling of robots in the absence of explicit targets, this invention introduces a target information masking mechanism during training. By masking the input target, the model learns environmental exploration strategies in the absence of target constraints. Point target navigation, pixel target navigation, and targetless exploration are uniformly modeled as a conditional decision problem. Different tasks are distinguished by task condition information, thus constructing a unified multi-task learning framework. Based on this unified modeling, a conditional masking mechanism modulates the model input, allowing the diffusion strategy to receive variable context information under different task conditions and generate corresponding action sequences accordingly. This mechanism essentially achieves explicit control over task conditions, enabling the same model to share representations and perform adaptive inference across multiple task modes.

[0060] During training, to achieve effective learning of the trajectory generation distribution, a noise prediction mechanism based on a diffusion model is used as the optimization objective. Specifically, the predicted noise is used... With real noise The mean squared error is used as the loss function for backpropagation, where Indicates the task category index. This represents the sequence number of the denoising step, and its loss function is expressed as:

[0061] ,

[0062] The initial noise-free expert action sequence is represented by MSE, which stands for Mean Squared Error. By constructing an error metric between the model's predicted noise and the actual noise, the model gradually approximates the real data distribution at different denoising stages, thereby achieving a mapping learning from the noise space to the trajectory space. Different task types are distinguished by task indices, and different denoising stages are characterized by time step indices. The loss function updates parameters by minimizing the mean squared error between the predicted noise and the actual noise.

[0063] S4. During navigation, the first step is to use a segmentation model to identify the navigation target. To build a perceptual foundation that supports semantically guided navigation, this invention proposes a dataset generation and segmentation model optimization paradigm for locally customized scenarios:

[0064] Based on the custom 3D simulation environment constructed in step S1 and the navigation trajectory generated in step S2, this invention utilizes a data generation engine to jointly model environmental perception data and motion trajectories, thereby constructing a standardized dataset consistent with the scene structure. This dataset uses RGB images as the primary observation form and establishes a correspondence between visual information and spatial targets through a semantic annotation mechanism. In terms of label design, by constructing semantic segmentation masks for each category of objects and assigning globally unique codes to different instances within the same category, accurate spatial differentiation of target objects is achieved, thus elevating the navigation problem from the category level to the instance level and improving the system's ability to represent complex scenes.

[0065] To address the issue of insufficient data distribution in limited-scale scenarios, this invention extends traditional data augmentation methods at the data generation mechanism level. Unlike image-space-based transformation operations, this invention directly operates on geometric and imaging parameters during the data generation process. By introducing perturbations in the trajectory and observation spaces, it achieves proactive expansion of the data distribution. Specifically, by introducing random perturbations into the robot's motion trajectory, the sampling path expands into a neighborhood in space. Simultaneously, perturbations are applied to the camera's intrinsic parameters, causing changes in the observation process in the projection space, thereby generating multi-view, multi-scale observation data. Furthermore, operations such as trajectory reversal enhance the symmetry and diversity of the path distribution.

[0066] To address the issues of uneven data distribution and label complexity, this invention further constructs a dynamic dataset adjustment mechanism. Since path-generated data naturally exhibits uneven distribution among different instances, directly dividing the training and validation sets makes it difficult to guarantee sufficient sample support for each instance during the validation phase. By statistically modeling the sample counts of each category and instance in both the training and validation sets, data distribution constraints are constructed, and cross-set data exchange operations are performed based on these constraints, thereby achieving adaptive completion of the validation set. In this scheme, the conditions for missing encoded instance data are defined as follows: Define constraints: .in Let i be the number of the i-th encoded instance in the validation set. Let be the number of the j-th encoded instance in the training set. Furthermore, a swap reward mechanism is designed to prioritize swapping image data containing labels of multiple encoded instances, thereby improving the image quality and label diversity of the validation set. Through these constraints and optimization mechanisms, a data distribution balance is achieved between the training and validation sets, thus providing a reliable basis for model performance evaluation.

[0067] At the model training level, this invention constructs a data-driven adaptive optimization mechanism. By coupling the dataset construction, model training, and performance evaluation processes, model performance can guide the data generation process in reverse. Specifically, by analyzing the performance metrics mIoU and mAP of the segmentation model on the validation set, the differences in learning performance among different instances are identified. mIoU (Mean Intersection over Union) measures the degree of overlap between the model's predicted region and the ground truth region.

[0068] ,

[0069] ,

[0070] In the formula, TP is the true positive, FP is the false positive, and FN is the false negative.

[0071] mAP (Mean Average Precision) measures the overall performance of a model at different confidence levels. The expression for the mean precision (AP) of a single class is:

[0072] ,

[0073] in, It is the index of the predicted samples sorted by confidence level, and also the sequence number of the confidence level output by the model. and Indicates the steps at step m and step m. Recall rate at the time of step Indicates the first Precision at each step. mAP is the average AP across all classes, defined as follows:

[0074] ,

[0075] For instances with insufficient performance, targeted data sampling and augmentation are re-performed in the simulation environment to dynamically supplement the data distribution. Through a closed-loop optimization process of "data construction – model training – performance evaluation – data reconstruction," the data distribution can be dynamically adjusted according to the model's learning state, thereby gradually improving the model's segmentation performance on each instance. This closed-loop mechanism breaks through the traditional static dataset training mode, enabling the system to have continuous optimization capabilities, especially in cases of uneven instance distribution and high scene complexity, significantly improving the model's robustness and generalization ability.

[0076] S5. To enable the robot to complete actual movements based on the path information output by the navigation model, it is necessary to train the control strategy of the legged robot to ensure stable execution of navigation actions. This invention uses a bipedal robot and, based on this, trains the robot's motion control model:

[0077] Traditional gait control methods require designing dedicated strategies and controllers for each motion behavior, necessitating extensive prior knowledge of robot structure and dynamics. Manual design and optimization are costly, and redesign is required for iterative model changes, resulting in poor versatility. To address these issues, this invention employs deep reinforcement learning for gait control. Given a robot description file, the model is first trained and tested in a simulation environment to obtain an effective neural network model suitable for the robot in the simulation environment. The trained model is then exported to ONNX format, converting it into a model suitable for real-world deployment, thus completing gait control in real-world scenarios. Figure 4 As shown, the robot motion control model architecture is illustrated.

[0078] First, a 3D physical model of the bipedal robot was created using CAD software. Links and joints were defined, and each link was exported as an STL file. This format uses multiple small triangular faces to form the 3D model. Then, a description file for the robot was written, defining the links, joints, and their properties. The links are composed of... <link> Element definition, containing <inertial> 、 <visual> 、 <collision>element. <inertial>The element defines the mass properties of the link, including the link's mass, the coordinates of its center of mass and its inertia tensor in the link's sub-coordinate system, which are calculated by CAD software. <visual>The element defines the visual properties of the link in the simulation environment. <collision>The element defines the collision properties of the link in the simulation environment; both are defined by an exported STL file. This description file can be read in the training environment.

[0079] Based on the principles of reinforcement learning, there is an expected cumulative reward:

[0080] ,

[0081] in, For state-action-reward sequence, For network parameters, For parameterized neural network strategies, Here, is a hyperparameter, represents the discount factor, and t represents the time step. The core objective is to maximize... According to the policy gradient theorem, the gradient of the objective function is:

[0082] ,

[0083] in, The advantage function is defined as the difference between action value and state value. Action value function (state) Take action below (expected cumulative reward) State value function (state) The expected cumulative reward. The advantage function is used to measure the "action". "How much better than the average" is the core signal for policy updates. This introduces probability ratios. This indicates the difference between the new strategy and the old strategy in state. Select action The probability ratio, thus the gradient objective can be transformed into:

[0084] ,

[0085] To address the issue of excessively large policy gradient update magnitudes, a clipping mechanism is introduced, and a clipping objective function is constructed:

[0086] ,

[0087] in, This indicates the editing operation, which will... Limited to Within the interval, This is a hyperparameter. Clipping operations can prevent overly aggressive or conservative strategy optimization, thus promoting stable updates.

[0088] Simultaneously, the state value function Optimization is performed to more accurately estimate state values, aiding in the calculation of the advantage function. The loss function for the value function is the mean squared error (MSE):

[0089] ,

[0090] in, This represents the sum of future rewards starting from time t. By minimizing this loss, the value function network can more accurately evaluate the value of a state, providing a more reliable advantage signal to the policy network.

[0091] To prevent the policy from converging prematurely to a local optimum, an entropy regularization term is introduced to encourage the policy to explore:

[0092] ,

[0093] in, Policy entropy measures the uncertainty of policy distribution. The larger the value, the greater the randomness of the policy and the stronger the exploratory ability; the smaller the value, the greater the certainty of the policy and the stronger the ability to make decisions based on existing knowledge.

[0094] Integrating the editing strategy loss, value function loss, and entropy regularization term, we obtain the total loss function:

[0095] ,

[0096] in, and These are hyperparameters, representing the value function loss weight and entropy regularization weight, respectively, used to balance the importance of each part. The negative sign indicates maximizing the editing strategy loss, i.e. minimizing the loss. .

[0097] The training algorithm process is as follows:

[0098] 1. Use the current strategy Interact with the environment to collect state-action-reward sequences;

[0099] 2. Calculate using Generalized Dominance Estimation (GAE) The formula is:

[0100] ,

[0101] in, For time difference residuals, Here, is a hyperparameter representing the GAE coefficient;

[0102] 3. Randomly sample small batches of data and calculate probability ratios. Calculate the total loss function and update the network parameters;

[0103] Repeat the above steps multiple times until the strategy converges.

[0104] The motion control training code is written in Python, the training framework is based on PyTorch, the simulation environment is provided by IsaacGym, and the system environment is Ubuntu 22.04. An environment class is defined to receive the training algorithm, containing interfaces for the robot model, spatial environment, observation, actions, physical stepping, rendering, and debugging. The spatial environment includes uneven terrain, which improves the robustness of gait control on different surfaces. The observation input is inertial measurement unit data, including acceleration and angular velocity, from which the robot's position data is calculated. The policy network output is the target position of each joint, which is converted into torque at the corresponding moment by the PD controller to drive the motor rotation. The reward function mainly includes the following terms: ① constant positive term, encouraging survival; ② base height term, encouraging appropriate height and stable posture; ③ forward velocity term, encouraging appropriate speed; ④ lateral velocity, vertical velocity, and angular velocity terms, constraining the corresponding speeds; ⑤ foot state term, constraining swing height and preventing slippage and impact; ⑥ joint velocity and torque terms, preventing exceeding limits; ⑦ gait phase term, used to form and improve gait rhythm. After 5000 iterations, the model converged. Tests were conducted under commands to stand, move forward, and turn; the movements were successfully completed without any falls. After training, the network model file was exported and deployed on a live robot. The robot control computing unit used an Ubuntu 22.04 system, deploying the motor and inertial measurement unit SDK to monitor sensor status and motor position data, receive motion commands, send motor control signals, and load the network model file to implement gait movements.

[0105] After completing the training of the navigation model, segmentation model, and robot control strategy, S6 can execute autonomous navigation tasks in either a simulation or real-world environment. The specific process of the autonomous navigation task is as follows:

[0106] During the task initialization phase, the system receives the semantic category information of the target object and uses it as input for global navigation. The robot acquires visual observation data of the current environment through its onboard RGB-D sensor and performs semantic parsing on the input image based on a semantic segmentation model. This establishes a correspondence between the environmental perception results and the target semantics, thereby determining the existence of the target.

[0107] Based on the above judgment results, the system constructs a state-driven behavior switching mechanism: when the target is not detected, the system is in the exploration state; when the target is detected, the system switches to the target-driven navigation state. Through this state division, a unified modeling is achieved between environment exploration under targetless conditions and path planning under target conditions.

[0108] In exploration mode, the system generates a target-free movement strategy based on environmental perception information. By perceiving the local spatial structure, it avoids obstacles, thereby expanding the environmental coverage area while ensuring safety. During exploration, the system continuously updates visual observations and repeatedly executes the target detection process, forming a closed-loop feedback mechanism between exploration behavior and target perception, thus enabling timely state switching when a target appears.

[0109] In navigation mode, the target region in the semantic segmentation result is mapped to a position representation in image space, and the correlation between visual targets and spatial motion is further constructed. Based on this, multimodal perception information is used as input, and a unified representation of the environment and target is obtained through feature extraction and cross-modal fusion mechanisms. Based on this representation, a conditional trajectory generation model is constructed, modeling the trajectory prediction problem as a conditional probability generation process. A diverse set of candidate trajectories is generated through progressive denoising, thereby characterizing the uncertainty of path selection in complex environments. To ensure the executability and safety of the trajectory, a constraint evaluation mechanism is introduced for the candidate trajectories. The trajectory is comprehensively measured from aspects such as spatial accessibility, obstacle distance, and trajectory continuity, and the optimal trajectory that satisfies the constraints is selected as the navigation output.

[0110] At the control execution level, a trajectory-to-control command mapping mechanism is constructed to transform the trajectory results planned at higher levels into speed control signals executable by the robot. Specifically, firstly, a trajectory tracking control method is used to map the reference trajectory into speed commands. Then, a policy learning-based control model further converts the speed commands into control variables in the robot's joint space, and combines this with the lower-level controller to achieve stable drive of the robot's motion. This hierarchical control approach achieves an effective connection between high-level path planning and low-level dynamics execution.

[0111] During the task evaluation phase, a termination determination mechanism based on spatial distance is constructed to measure the distance relationship between the robot's final position and the target position. When the distance is less than a preset threshold, the task is determined to be successful; otherwise, it is determined to be a failure, thereby achieving a quantitative evaluation of the navigation task results.

[0112] In a second aspect, the present invention provides an electronic device, comprising: one or more processors; and a memory for storing one or more programs; wherein, when the one or more programs are executed by the one or more processors, the one or more processors implement the aforementioned method for navigating a legged robot oriented towards an object target within a custom scene.

[0113] Thirdly, the present invention provides a computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, enable the processor to implement the aforementioned method for navigating a legged robot oriented towards an object target within a custom scene.

[0114] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above descriptions are merely specific embodiments of the present invention and are not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.< / collision> < / visual> < / inertial> < / collision> < / visual> < / inertial>

Claims

1. A method for navigating a legged robot targeting objects within a custom scene, characterized in that, include: Step 1: The robot acquires perceptual information from RGB and depth images using a camera. It then performs semantic parsing on the perceptual information using a pre-trained semantic segmentation model to extract the pixel-level semantic distribution of various objects in the field of view. Combined with the input target semantic category information, the robot determines the existence of the target object in the current observation space. If the target object exists, it outputs the pixel-level position distribution of the target object in the image. Step 2: Based on the result of the existence determination of the target object, construct a switching mechanism between the exploration mode and the target-oriented navigation mode. When the target object is not detected, drive the robot to execute the exploration mode; when the target object is detected, trigger the target-oriented navigation mode. Step 3: When in target-oriented navigation mode, the pixel-level position distribution is mapped to an intermediate state representation and used as a navigation target condition. A diffusion navigation decision model based on multimodal perception information is constructed. The spatial relationship between the robot's current state and the target state is modeled using fused RGB images and depth images to generate a continuous navigation trajectory to reach the target object. Step 4: Convert the continuous navigation trajectory into speed commands and input them into the legged robot motion control model to generate control commands for robot joint movement and drive the robot to move. When the robot's state meets the preset target approach conditions, the task is considered successful; otherwise, the task is considered a failure.

2. The legged robot navigation method for objects within a custom scene according to claim 1, characterized in that, The pre-trained semantic segmentation model is obtained by: constructing a 3D simulation scene with the same geometric scale and spatial distribution as the local real scene, performing structured decomposition and unique identification on each object in the scene, and generating pixel-level annotation representations containing semantic category labels and instance-level encodings. Based on the dataset containing RGB images and corresponding annotations collected from the 3D simulation scene, a semantic segmentation model is trained through supervised learning. A data closed-loop optimization mechanism oriented towards performance feedback is introduced. For the model's performance on low-performance instances, targeted data sampling and enhancement are adaptively performed in the simulation environment. The data distribution is dynamically supplemented and adjusted through an iterative process of "data construction - model training - performance evaluation - data reconstruction".

3. The legged robot navigation method for objects within a custom scene according to claim 2, characterized in that, The construction of the three-dimensional simulation scene specifically includes: constructing an initial three-dimensional scene based on the spatial layout information of the real scene using three-dimensional modeling tools; applying proportional constraints to the size and relative positional relationships of each object in the scene; adopting an alternative modeling strategy based on geometric structural similarity for object models that cannot be provided in the model library; performing structured decomposition after exporting the scene data; uniquely identifying and naming each object; optimizing the texture and enhancing the visual details of the surface properties; configuring material properties and lighting parameters after importing into the simulation platform; and adding rigid body properties and collision body properties to each object.

4. The legged robot navigation method for objects within a custom scene according to claim 1, characterized in that, In step two, the switching mechanism between the exploration mode and the target-oriented navigation mode involves the system continuously updating visual observations and repeatedly executing the target detection process in the exploration mode, so that the exploration behavior and target perception form a closed-loop feedback mechanism, thereby triggering the state switch in a timely manner when the target appears.

5. A legged robot navigation method for objects within a custom scene according to claim 1, characterized in that, In step three, the diffusion navigation decision model is trained using a multimodal navigation dataset. The method for collecting the navigation dataset includes: in the constructed 3D simulation scene, the passable area is discretized and spatially expanded; a random sampling strategy is used to generate a set of initial and target states; a path planning algorithm is used to generate a navigation trajectory; a pre-trained reinforcement learning gait control model is deployed on the robot to drive the robot to perform navigation tasks; the robot's motion state, visual perception information, and target semantic information are collected simultaneously to form a multimodal navigation dataset containing RGB images, depth images, robot pose, and trajectory labels; and the data is filtered through trajectory consistency evaluation.

6. A legged robot navigation method for objects within a custom scene according to claim 1, characterized in that, In step three, the diffusion navigation decision model uses a diffusion model based on a conditional generation mechanism to model the future trajectory. By constructing a conditional network architecture and combining it with a denoising diffusion probability model scheduling strategy, the trajectory generation process is represented as an iterative denoising process that gradually evolves from a random noise distribution to a target trajectory distribution, generating a diverse set of candidate trajectories.

7. A legged robot navigation method for objects within a custom scene according to claim 6, characterized in that, After obtaining the candidate trajectory set, an independent evaluation mechanism is introduced to evaluate the quality of each candidate trajectory. By constructing trajectory feasibility and safety constraints, the trajectory is comprehensively measured in terms of spatial accessibility, distance relationship with obstacles, and motion continuity. The optimal trajectory that meets the constraints is selected as the navigation output.

8. A legged robot navigation method for objects within a custom scene according to claim 6, characterized in that, The diffusion navigation decision model introduces a target information masking mechanism during training. By masking the input target, the model learns environmental exploration strategies in the absence of target constraints. It unifies the three task forms of point target navigation, pixel target navigation, and targetless exploration into a conditional decision problem. The model input is modulated through a conditional masking mechanism, enabling the model to receive variable context information and generate corresponding action sequences under different task conditions.

9. A legged robot navigation method for objects within a custom scene according to claim 1, characterized in that, In step four, the legged robot motion control model is obtained through the following preliminary steps: based on the spatial layout information of the local real scene, a three-dimensional simulation scene is constructed in the simulation environment that is consistent with the real environment in terms of geometric scale and spatial distribution. Based on the constructed 3D simulation scene, a physical model and description file of a bipedal robot are built. A reinforcement learning training mechanism is constructed based on the policy gradient theorem. A clipping mechanism is introduced to constrain the policy update amplitude. An optimization value function is used to improve the accuracy of state value estimation. An entropy regularization term is added to encourage policy exploration. The neural network model is trained in a simulation environment with uneven ground. The policy model of the simulation environment is converted into a policy model suitable for real machine deployment.

10. A legged robot navigation method for objects within a custom scene according to claim 1, characterized in that, In step four, the determination of whether the task is successful or not is as follows: by constructing a termination determination mechanism based on spatial distance, the distance relationship between the robot's final position and the target position is measured. When the distance is less than a preset threshold, the task is determined to be successful; otherwise, it is determined to be a failure.

Images