Risk-aware conditional flow matching collision-free visual navigation method and system

Abstract

The application discloses a kind of condition flow matching collision-free visual navigation method and system based on risk perception, the application is converted into two-dimensional grid by monocular depth information through obstacle perception grid, extracts point density, depth interval, multi-dimensional risk characteristics of obstacle height;After fusing environment, target, obstacle grid multi-class features, generate multiple candidate trajectories in parallel through conditional flow matching;Group relative strategy optimization is used in training stage to complete offline fine-tuning, composite soft reward is designed to contain continuous collision penalty, exponential gate open incentive, rely on group internal normalized relative advantage weighted optimization network, eliminate fine-tuning architecture in inference stage, and will not increase online operation delay.The application can explicitly quantify spatial collision risk, eliminate training gradient fault problem, greatly improve the navigation effect in complex environment, effectively reduce collision behavior, and can stably maintain sufficient safety margin when driving in narrow channel.

Description

Technical Field

[0001] This invention belongs to the technical field of visual navigation, specifically relating to a collision-free visual navigation method and system based on risk perception and conditional flow matching. Background Technology

[0002] Visual navigation is a key technology for autonomous mobile robots to perform tasks in real-world environments. In conventional approaches, the embodied agent generates control actions based on received visual observations to safely reach the target location. Early embodied navigation algorithms often employed deterministic end-to-end mappings, directly translating visual input into a single action. However, this significantly limited the model's ability to capture the distribution of multimodal actions in complex environments. To address the multimodal trajectory modeling problem, recent techniques have introduced generative strategies based on diffusion models or Conditional Flow Matching (CFM). These strategies predict multiple feasible trajectories by fitting large amounts of human demonstration data, thereby improving navigation success rates.

[0003] However, existing generative visual navigation methods still have significant drawbacks. The most significant problem is that these methods lack explicit safety-related modeling and cannot effectively extract and internalize robust safety constraint features from human demonstration data. This results in mobile robots lacking awareness of explicit safety boundaries in their trajectories when deployed in dense, cluttered, or unknown complex environments, easily leading to high-frequency collisions and unsafe behaviors. The root cause of this problem is that human demonstration data often implicitly contains and intertwines diverse personal operational preferences and constantly changing physical contexts, making it difficult for simple behavior cloning strategies to decouple and extract robust and generalizable safety distributions within the latent space. Therefore, standard generative imitation strategies have high uncertainty in the distribution of future candidate trajectories when facing unfamiliar environments, easily generating "edge-grabbing" trajectories that closely hug obstacle boundaries with extremely small safety gaps, causing serious safety hazards. In addition, some existing schemes that attempt to force alignment for safety through trajectory selection often use hard indicator functions of 0 and 1 for their reward functions, leading to technical bottlenecks such as feedback signal discontinuity, excessively sparse gradients, and unstable parameter optimization during the fine-tuning stage. Summary of the Invention

[0004] The main objective of this invention is to overcome the shortcomings and deficiencies of the prior art and provide a collision-free visual navigation method and system based on risk perception and conditional flow matching. By constructing an obstacle perception grid to achieve explicit representation of multi-dimensional risks, and combining conditional flow matching to generate multiple candidate trajectories, soft safety alignment is completed by using a group relative strategy that is only enabled during the training phase. Without increasing online inference latency, this invention improves the success rate of navigation in complex environments, reduces the collision frequency, and ensures the continuous and stable gradient of the strategy optimization process.

[0005] To achieve the above objectives, the present invention adopts the following technical solution:

[0006] In a first aspect, the present invention provides a risk-aware conditional flow matching collision-free visual navigation method, comprising the following steps:

[0007] S1. Obtain multimodal features, including temporal observation features, navigation target features, and explicit obstacle topology features;

[0008] S2. The time-series observation features, navigation target features, and explicit obstacle topology features are concatenated in the channel dimension, input into a multi-layer Transformer encoder for long-range dependency aggregation, and the final control context vector is output.

[0009] S3. Using the control context vector as a conditional constraint, the standard Gaussian noise trajectory sample is smoothly transported and transformed within continuous time steps through the conditional flow matching model, and candidate trajectory control groups are output in parallel.

[0010] S4. Based on the offline group relative trajectory alignment mechanism, perform security assessment and preference alignment on each candidate trajectory in the candidate trajectory control group, so that the trajectory distribution output by the optimized network in the same environment converges to the safe mode.

[0011] As a preferred technical solution, in step S1, obtaining the multimodal features specifically involves:

[0012] Input the current frame image into the scene encoder and output the temporal observation features;

[0013] When in target-driven mode, the target image is input to the target encoder, and the navigation target features are output.

[0014] The current frame image is transformed into an estimated depth map through a monocular relative depth estimation network. The estimated depth map is then subjected to 3D backprojection and local horizontal meshing through a projection module to dynamically construct an obstacle perception mesh. The geometric statistics of the obstacle perception mesh are then dimensionality-reduced and mapped through a mesh encoder to output explicit obstacle topological features.

[0015] As a preferred technical solution, the specific implementation steps of the obstacle sensing grid are as follows:

[0016] 3D point cloud back projection: Using the camera intrinsic parameter matrix, the pixels of the monocular depth map are back projected to the robot coordinate system to generate a local 3D point cloud.

[0017] Horizontal spatial discretization involves discretizing the ground area in front of the robot into a two-dimensional horizontal grid, wherein the two-dimensional grid comprises multiple grid cells.

[0018] Raster multidimensional risk indicator extraction: geometric indicators are extracted in parallel for each raster cell: point density, minimum and maximum depth interval, and normalized maximum height.

[0019] Raster feature encoding involves concatenating the indices of individual raster cells and inputting them into a multilayer perceptron, which outputs the topological features of individual raster cells. All the raster features are concatenated to form the explicit obstacle topological features.

[0020] As a preferred technical solution, the point density is specifically as follows:

[0021] The total number of 3D point clouds falling into the grid cells is counted, and logarithmic smoothing is performed to indicate the probability that the space is occupied by an entity and to eliminate outlier noise.

[0022] The minimum and maximum depth intervals are specifically defined as follows: the normalized axial depth limits are recorded, which respectively represent the nearest obstacle reflective surface and the farthest line-of-sight reachable boundary in the current grid cell;

[0023] The normalized maximum height specifically refers to: calculating the maximum height of the point cloud within a grid relative to the ground and normalizing it to a value. An interval is used to characterize whether there are three-dimensional obstacles that are higher than the ground within a grid cell.

[0024] As a preferred technical solution, step S3 specifically includes:

[0025] The conditional flow matching model learns how to smoothly transport and transform initial noise trajectory samples with a standard Gaussian distribution into two-dimensional future waypoints in multiple vehicle coordinate systems by fitting continuous transport flow dynamics. These trajectory samples together form a candidate trajectory control group.

[0026] As a preferred technical solution, in step S4, the offline group relative trajectory alignment mechanism specifically is as follows:

[0027] Grouping and Dual Soft Safety Reward Assessment: In the fine-tuning phase, multiple candidate trajectories are repeatedly sampled by the conditional flow matching model for visual observations, target conditions, and local obstacle layouts at the same time. The grouping module divides the multiple candidate trajectories generated based on the same conditional context into a candidate trajectory control group, and candidate trajectories generated under different observation conditions are assigned to different control groups. Subsequently, using synchronized privileged geometric layout information, the candidate trajectories in each control group are scored in parallel through a composite safety reward function, which includes collision penalty rewards and open space incentive assessments.

[0028] Intra-group relative advantage and survival of the fittest update: During the reward settlement phase, the relative advantages within the group are calculated. The total reward of each trajectory is calculated and normalized within the independent control group to determine the relative advantage of each candidate trajectory relative to the group average performance.

[0029] Preference alignment backpropagation: The relative advantage is transformed into the final preference optimization weights through the Softmax operator, which are then used to apply weighted penalties or incentives to the flow loss function of conditional flow matching.

[0030] As a preferred technical solution, the collision penalty reward is calculated by iterating through the candidate trajectory to calculate the minimum Euclidean distance from each path point to a nearby obstacle. When this distance is less than a preset safety threshold, a quadratic continuous penalty is triggered.

[0031] As a preferred technical solution, the open space incentive assessment is obtained by calculating the cosine similarity between the direction vector of the candidate trajectory and the most spacious passable central axis vector extracted from the topological features of the explicit obstacle, and the collision penalty reward is used as a continuous exponential gating to attenuate and suppress the open space incentive assessment.

[0032] As a preferred technical solution, in step S4, for a specific multimodal visual observation context vector, the conditional flow matching head does not output a single action, but instead samples and extracts a control group containing multiple heterogeneous waypoint candidate trajectories in parallel.

[0033] Secondly, the present invention provides a risk-aware conditional flow matching collision-free visual navigation system, applied to the risk-aware conditional flow matching collision-free visual navigation method, including an image acquisition module, a multimodal feature fusion module, a conditional flow matching trajectory generation module, and a trajectory output and control module.

[0034] The image acquisition module is used to acquire multimodal features, including temporal observation features, navigation target features, and explicit obstacle topological features;

[0035] The multimodal feature fusion module is used to cascade the temporal observation features, navigation target features, and explicit obstacle topology features in the channel dimension, input them into a multi-layer Transformer encoder for long-range dependency aggregation, and output the final control context vector.

[0036] The conditional flow matching trajectory generation module is used to use the control context vector as a conditional constraint term, and through the conditional flow matching model, smoothly transport and transform the standard Gaussian noise trajectory sample in continuous time steps, and output candidate trajectory control groups in parallel.

[0037] The trajectory output and control module is used to perform safety assessment and preference alignment on each candidate trajectory in the candidate trajectory control group based on the offline group relative trajectory alignment mechanism, so that the trajectory distribution output by the optimized network converges to the safety mode under the same environment.

[0038] Compared with the prior art, the present invention has the following advantages and beneficial effects:

[0039] Significantly improves navigation success rate in dense and cluttered environments. This invention employs a technical solution that fuses obstacle perception grids with conditional flow matching navigation strategies. It converts monocular depth information into spatial risk representations and introduces them into the trajectory generation process. This solves the problems of existing visual navigation methods lacking explicit obstacle risk modeling and easily generating high-risk trajectories in dense static obstacles and irregular outdoor terrain. As a result, the target navigation success rate of embodied robots is increased from 30% of the traditional baseline method to 70%.

[0040] Significantly reduces collision frequency. This invention employs an offline group relative trajectory alignment mechanism, combined with collision penalty rewards and open space incentives to optimize candidate trajectories within a group. This solves the problem that existing imitation learning navigation strategies only fit artificially demonstrated trajectories and have difficulty actively distinguishing between safe trajectories and trajectories close to obstacles. As a result, the average number of collisions per run for the robot in target navigation tasks is reduced from 2.5 to 1.1. In targetless free exploration tasks, the collision frequency is reduced by approximately 70%, improving obstacle avoidance generalization ability in unknown environments.

[0041] This invention effectively improves trajectory safety margin. Through explicit spatial risk modeling and open space reward constraints, the trajectory generation process prioritizes passable areas with larger obstacle spacing among multiple candidate trajectories with reasonable behavior. This solves the problem that existing generative navigation strategies based on imitation learning lack explicit safety preferences and have difficulty stably distinguishing trajectories with different safety margins. As a result, the robot maintains a lateral safety margin of approximately 0.24m in narrow corridor scenarios, which is significantly higher than the baseline method's 0.03m, and prioritizes the more open and safer passable area when multiple feasible passable areas exist. Attached Figure Description

[0042] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart of a risk-aware conditional flow matching collision-free visual navigation method according to an embodiment of the present invention;

[0044] Figure 2 This is a block diagram of a collision-free visual navigation system based on risk perception and conditional flow matching, according to an embodiment of the present invention. Detailed Implementation

[0045] To enable those skilled in the art to better understand the present application, the technical solutions in the embodiments of the present application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are merely some embodiments of the present application, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present application without creative effort are within the scope of protection of the present application.

[0046] In this application, the reference to "embodiment" means that a specific feature, structure, or characteristic described in connection with an embodiment may be included in at least one embodiment of this application. The appearance of this phrase in various places throughout the specification does not necessarily refer to the same embodiment, nor is it a mutually exclusive, independent, or alternative embodiment. It will be explicitly and implicitly understood by those skilled in the art that the embodiments described in this application can be combined with other embodiments.

[0047] Definitions of abbreviations and key terms:

[0048] RaFM-Nav (Risk-Aware Flow Matching Navigation): Risk-aware conditional flow matching navigation framework, the core architecture of this invention.

[0049] OPG (Obstacle Perception Grid): An obstacle perception grid, a perception module that converts monocular depth into a top-down occupied grid.

[0050] CFM (Conditional Flow Matching): A generative underlying network used to generate continuous motion trajectories.

[0051] GRPO (Group-Relative Policy Optimization) is a reinforcement learning algorithm that estimates advantage and optimizes policy by ranking samples within a group.

[0052] like Figure 1 As shown, this embodiment provides a collision-free visual navigation method based on risk perception and conditional flow matching. This invention is implemented based on a multimodal perception input, a basic generative navigation strategy network, and a group relative trajectory alignment fine-tuning architecture, specifically including the following steps:

[0053] S1. Obtain multimodal features, including temporal observation features, navigation target features, and explicit obstacle topology features;

[0054] Furthermore, the time-series observation features are obtained in the following manner:

[0055] like Figure 1 As shown in the lower left corner, the egocentric RGB image of the current frame is input to the scene encoder to extract temporal observation features. If in target-driven mode, the target image is input to the target encoder, and the navigation target features are output. .

[0056] like Figure 1 As shown in the upper left diagram, the current frame image is first transformed into an estimated depth map through a monocular relative depth estimation network. The estimated depth map is then subjected to 3D backprojection and local horizontal meshing by a projection module to dynamically construct an obstacle perception mesh. Finally, the mesh encoder performs dimensionality reduction mapping on the geometric statistics of the mesh to output explicit obstacle topological features. .

[0057] Furthermore, in order to compensate for the deficiency of implicit visual features in perceiving spatial physical boundaries in dense communities, such as Figure 1 As shown in the upper left corner, the specific implementation steps of the obstacle sensing grid are as follows:

[0058] 3D point cloud back projection: using known camera intrinsic matrix The pixel coordinates in the monocular relative depth map are inversely projected onto the three-dimensional vehicle coordinate system to generate a local 3D point cloud.

[0059] Horizontal spatial discretization and rasterization: The local ground physical area 4m horizontally and 8m vertically in front of the vehicle is discretized into a grid with a size of [missing information]. A fixed-resolution two-dimensional horizontal local mesh.

[0060] Multidimensional geometric statistics quantization description: for each independent grid cell after discretization The system extracts the following geometric indicators in parallel, which can directly map the physical collision risk:

[0061] (1) Point density Count the total number of 3D point clouds falling within this grid and perform logarithmic smoothing. The processing is used to indicate the probability that the space is occupied by an entity and to eliminate outlier noise.

[0062] (2) Minimum and maximum depth intervals : Records the normalized axial depth limits, representing the nearest obstacle reflective surface and the farthest line-of-sight reachable boundary in the current grid, respectively.

[0063] (3) Normalized maximum height : Statistically calculate the maximum height of point clouds relative to the ground within a raster and normalize it to . The interval, this value directly reflects whether there are three-dimensional obstacles that are higher than the ground within the grid.

[0064] Spatial label generation: The multidimensional statistics including occupancy, boundary depth, and absolute height are concatenated and fed into a grid encoder composed of a two-layer, multilayer perceptron, outputting a one-dimensional high-dimensional spatial topological feature vector. :

[0065]

[0066] The set of all grid markers together constitutes the top-down camera-centric view space prior matrix. .

[0067] S2. The time-series observation features, navigation target features, and explicit obstacle topology features are concatenated in the channel dimension, input into a multi-layer Transformer encoder for long-range dependency aggregation, and the final control context vector is output.

[0068] See appendix Figure 1 The flow path below generates parallel features that are concatenated and stitched together along the channel dimension, and then fed into a four-layer Transformer encoder for long-range dependency aggregation, outputting the final control context vector. :

[0069]

[0070] in, Represents the characteristics of time series observations. Represents the topological features of explicit obstacles.

[0071] S3. Using the control context vector as a conditional constraint, a conditional flow matching network is used to smoothly transport and transform the standard Gaussian noise trajectory samples within continuous time steps, and parallel outputs candidate trajectory control groups; specifically:

[0072] The context vector As a conditional constraint, it is fed into the generative action backbone network—the conditional flow matching head—in real time. The conditional flow matching head learns how to process the initial noise trajectory samples with a standard Gaussian distribution by fitting the continuous transport flow dynamics. In continuous time steps Internal smooth transport and transformation into containing The motion trajectories of two-dimensional future waypoints in the vehicle coordinate system together form a candidate trajectory control group.

[0073] S4. Based on the offline group relative trajectory alignment mechanism, perform security assessment and preference alignment on each candidate trajectory in the candidate trajectory control group, so that the trajectory distribution output by the optimized network in the same environment converges to the safe mode.

[0074] Understandably, in the initial stage without safe alignment, the original trajectory distribution output by a policy network trained purely through behavior cloning exhibits high variability. For example... Figure 1 As shown, the generated alternative waypoints (such as...) Figure 1 The dispersed paths shown by trajectory 1, trajectory 2, trajectory 3, ..., trajectory n often produce "edge-grabbing" or collision paths that closely follow or even pass through obstacles (such as stools or simulated banana props in the model).

[0075] To correct the aforementioned unsafe behaviors, the present invention establishes the following... Figure 1 The offline preference alignment optimization pipeline is shown in the upper right diagram. It's important to note that this pipeline is only activated during the training fine-tuning phase and is completely removed during actual inference deployment, thus adding no online computation latency. Its detailed process and alignment mechanism are as follows:

[0076] Parallel sampling of multiple candidate trajectories within a group: for a specific multimodal visual observation context The conditional flow matching head does not output a single action, but rather samples in parallel to extract a list containing... Control group for candidate trajectories of heterogeneous waypoints.

[0077] Intra-group partitioning and dual soft security reward evaluation: All candidate paths within a control group are fed into the group partitioning module and scored in parallel within the virtual environment using synchronized privileged geometric layout information and a composite security reward function. The scoring includes two core evaluation paths:

[0078] (a) Collision penalty reward: If trajectory 4 < trajectory 1, then iterate through and calculate the minimum Euclidean distance from each waypoint in the current candidate trajectory to the nearest known obstacle. If the distance is less than the preset safety threshold... This triggers a quadratic consecutive penalty: ;

[0079] This design avoids the problem of the derivative of traditional indicator functions returning to zero when the distance does not change abruptly, and can provide a smooth, non-sparse gradient flow.

[0080] (b) Open Space Reward Assessment: If trajectory 3 < trajectory 1, extract the widest, unobstructed, accessible central axis vector indicated in the OPG grid. Calculate the generated candidate trajectory direction vector Its cosine similarity. To prevent the robot from blindly aiming at open ground while ignoring nearby collisions, collision penalties are treated as continuous exponential gating for decay suppression:

[0081]

[0082] Intra-group relative advantage elimination update (Reward Phase): During the reward settlement phase, the relative advantages within the group are calculated. Total reward for each trajectory Normalization was performed within this independent control group, and the relative advantage of each candidate trajectory relative to the group's average performance was calculated: ;

[0083] in and These are the mean and standard deviation of the scores for the 10 sampling trajectories within the current group, respectively.

[0084] Preference Aligned Backpropagation: This transforms the relative preference into the final preference optimization weights using the Softmax operator. Flow loss function for conditional flow matching Apply weighted penalties or incentives.

[0085] like Figure 1 As shown in the optimized trajectory diagram in the lower right corner, after fine-tuning the relative preference alignment within this group, the weights of the Transformer coding layer, OPG coding layer, and conditional flow matching head within the network have been corrected. When facing obstacles in the same environment, the output generative trajectory distribution can significantly achieve overall cohesion and contraction, and actively move towards the desired safe mode of moving away from obstacles on the left and being completely centered on the wide and open safe passage on the right.

[0086] It should be noted that, for the sake of simplicity, the aforementioned method embodiments are all described as a series of actions. However, those skilled in the art should understand that the present invention is not limited to the described order of actions, because according to the present invention, some steps can be performed in other orders or simultaneously.

[0087] Based on the same idea as the risk-aware conditional flow matching collision-free visual navigation method in the above embodiments, the present invention also provides a risk-aware conditional flow matching collision-free visual navigation system, which can be used to execute the above-described risk-aware conditional flow matching collision-free visual navigation method. For ease of explanation, the structural schematic diagram of the risk-aware conditional flow matching collision-free visual navigation system embodiment only shows the parts related to the embodiments of the present invention. Those skilled in the art will understand that the illustrated structure does not constitute a limitation on the device, and it may include more or fewer components than shown, or combine certain components, or have different component arrangements.

[0088] S5: To verify the technical effect of this embodiment, a quadruped robot was used to conduct a comparative experiment in a real environment.

[0089] The experiment was deployed on a Unitree Go2 quadruped robot. The robot was controlled via velocity commands, acquiring RGB images at each moment and predicting 8 future 2D waypoints in the vehicle's coordinate system. The predicted waypoints were then converted into control commands. Experimental scenarios included dense static indoor obstacles, irregular outdoor terrain, and near-field dynamic obstacle insertion. Each method was tested 10 times in each scenario, for a total of 30 runs. The target navigation success rate, average number of collisions per run, and inference latency were statistically analyzed. The comparative experimental results are shown in Table 1.

[0090] Table 1

[0091] method Target navigation success rate (unit: %) Target navigation collision count (unit: times) Number of collisions during aimless exploration (unit: times) Inference delay (unit: ms) NoMaD 20.0 2.8 3.6 62.1 FlowNav 30.0 2.5 3.5 63.8 FlowNav+ obstacle perception grid module 30.0 2.1 2.5 65.0 FlowNav+ Group Relative Trajectory Alignment Module 40.0 1.6 1.8 63.8 Complete method of the present invention 70.0 1.1 1.0 65.0

[0092] The experimental results above show that, compared to the FlowNav baseline method, the complete method of this invention improves the target navigation success rate from 30.0% to 70.0%, reduces the average number of collisions per run in the target navigation task from 2.5 to 1.1, and reduces the average number of collisions per run in the targetless free exploration task from 3.5 to 1.0. Meanwhile, the average inference latency of the complete method of this invention is 65.0 ms, which is similar to that of the baseline method, indicating that the offline relative trajectory alignment mechanism improves safety without significantly increasing the online inference computation burden.

[0093] Please see Figure 2 In another embodiment of this application, a risk-aware conditional flow matching collision-free visual navigation system is provided, which includes an image acquisition module, a multimodal feature fusion module, a conditional flow matching trajectory generation module, and a trajectory output and control module.

[0094] The image acquisition module is used to acquire multimodal features, including temporal observation features, navigation target features, and explicit obstacle topological features;

[0095] The multimodal feature fusion module is used to cascade the temporal observation features, navigation target features, and explicit obstacle topology features in the channel dimension, input them into a multi-layer Transformer encoder for long-range dependency aggregation, and output the final control context vector.

[0096] The conditional flow matching trajectory generation module is used to use the control context vector as a conditional constraint term, and through the conditional flow matching model, smoothly transport and transform the standard Gaussian noise trajectory sample in continuous time steps, and output candidate trajectory control groups in parallel.

[0097] The trajectory output and control module is used to perform safety assessment and preference alignment on each candidate trajectory in the candidate trajectory control group based on the offline group relative trajectory alignment mechanism, so that the trajectory distribution output by the optimized network converges to the safety mode under the same environment.

[0098] Furthermore, such as Figure 2 The system of the present invention further includes a depth estimation module, an obstacle perception grid construction module, a scene encoder module, a target encoder module, and a grid feature encoding module. The implementation methods of these modules are consistent with the corresponding technical solutions of the risk perception-based conditional flow matching collision-free visual navigation method in the present invention, and will not be described in detail here.

[0099] It should be noted that the risk-aware conditional flow matching collision-free visual navigation system of the present invention corresponds one-to-one with the risk-aware conditional flow matching collision-free visual navigation method of the present invention. The technical features and beneficial effects described in the embodiments of the risk-aware conditional flow matching collision-free visual navigation method are applicable to the embodiments of the risk-aware conditional flow matching collision-free visual navigation. For details, please refer to the description in the embodiments of the method of the present invention, which will not be repeated here.

[0100] Furthermore, in the above embodiments of the risk-aware conditional flow matching collision-free visual navigation system, the logical division of each program module is merely an example. In actual applications, the above functions can be assigned to different program modules as needed, for example, for the sake of corresponding hardware configuration requirements or software implementation convenience. That is, the internal structure of the risk-aware conditional flow matching collision-free visual navigation system can be divided into different program modules to complete all or part of the functions described above.

[0101] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0102] The above embodiments are preferred embodiments of the present invention, but the embodiments of the present invention are not limited to the above embodiments. Any changes, modifications, substitutions, combinations, or simplifications made without departing from the spirit and principle of the present invention shall be considered equivalent substitutions and shall be included within the protection scope of the present invention.

Claims

1. A collision-free visual navigation method based on risk perception and conditional flow matching, characterized in that, Includes the following steps: S1. Obtain multimodal features, including temporal observation features, navigation target features, and explicit obstacle topology features; S2. The time-series observation features, navigation target features, and explicit obstacle topology features are concatenated in the channel dimension, input into a multi-layer Transformer encoder for long-range dependency aggregation, and the final control context vector is output. S3. Using the control context vector as a conditional constraint, the standard Gaussian noise trajectory sample is smoothly transported and transformed within continuous time steps through the conditional flow matching model, and candidate trajectory control groups are output in parallel. S4. Based on the offline group relative trajectory alignment mechanism, perform security assessment and preference alignment on each candidate trajectory in the candidate trajectory control group, so that the trajectory distribution output by the optimized network in the same environment converges to the safe mode.

2. The risk-aware conditional flow matching collision-free visual navigation method according to claim 1, characterized in that, In step S1, obtaining multimodal features specifically involves: Input the current frame image into the scene encoder and output the temporal observation features; When in target-driven mode, the target image is input to the target encoder, and the navigation target features are output. The current frame image is transformed into an estimated depth map through a monocular relative depth estimation network. The estimated depth map is then subjected to 3D backprojection and local horizontal meshing through a projection module to dynamically construct an obstacle perception mesh. The geometric statistics of the obstacle perception mesh are then dimensionality-reduced and mapped through a mesh encoder to output explicit obstacle topological features.

3. The risk-aware conditional flow matching collision-free visual navigation method according to claim 2, characterized in that, The specific implementation steps of the obstacle sensing grid are as follows: 3D point cloud back projection: Using the camera intrinsic parameter matrix, the pixels of the monocular depth map are back projected to the robot coordinate system to generate a local 3D point cloud. Horizontal spatial discretization involves discretizing the ground area in front of the robot into a two-dimensional horizontal grid, wherein the two-dimensional grid comprises multiple grid cells. Raster multidimensional risk indicator extraction: geometric indicators are extracted in parallel for each raster cell: point density, minimum and maximum depth interval, and normalized maximum height. Raster feature encoding involves concatenating the indices of individual raster cells and inputting them into a multilayer perceptron, which outputs the topological features of individual raster cells. All the raster features are concatenated to form the explicit obstacle topological features.

4. The risk-aware conditional flow matching collision-free visual navigation method according to claim 3, characterized in that, The point density is specifically: The total number of 3D point clouds falling into the grid cells is counted, and logarithmic smoothing is performed to indicate the probability that the space is occupied by an entity and to eliminate outlier noise. The minimum and maximum depth intervals are specifically defined as follows: the normalized axial depth limits are recorded, which respectively represent the nearest obstacle reflective surface and the farthest line-of-sight reachable boundary in the current grid cell; The normalized maximum height specifically refers to: calculating the maximum height of the point cloud within a grid relative to the ground and normalizing it to a value. An interval is used to characterize whether there are three-dimensional obstacles that are higher than the ground within a grid cell.

5. The risk-aware conditional flow matching collision-free visual navigation method according to claim 1, characterized in that, Step S3 is as follows: The conditional flow matching model learns how to smoothly transport and transform initial noise trajectory samples with a standard Gaussian distribution into two-dimensional future waypoints in multiple vehicle coordinate systems by fitting continuous transport flow dynamics. These trajectory samples together form a candidate trajectory control group.

6. The risk-aware conditional flow matching collision-free visual navigation method according to claim 1, characterized in that, In step S4, the offline group relative trajectory alignment mechanism is specifically as follows: Grouping and Dual Soft Safety Reward Assessment: In the fine-tuning phase, multiple candidate trajectories are repeatedly sampled by the conditional flow matching model for visual observations, target conditions, and local obstacle layouts at the same time. The grouping module divides the multiple candidate trajectories generated based on the same conditional context into a candidate trajectory control group, and candidate trajectories generated under different observation conditions are assigned to different control groups. Subsequently, using synchronized privileged geometric layout information, the candidate trajectories in each control group are scored in parallel through a composite safety reward function, which includes collision penalty rewards and open space incentive assessments. Intra-group relative advantage and survival of the fittest update: During the reward settlement phase, the relative advantages within the group are calculated. The total reward of each trajectory is calculated and normalized within the independent control group to determine the relative advantage of each candidate trajectory relative to the group average performance. Preference alignment backpropagation: The relative advantage is transformed into the final preference optimization weights through the Softmax operator, which are then used to apply weighted penalties or incentives to the flow loss function of conditional flow matching.

7. The risk-aware conditional flow matching collision-free visual navigation method according to claim 6, characterized in that, The collision penalty reward is calculated by iterating through the candidate trajectory to calculate the minimum Euclidean distance from each point to a nearby obstacle. When this distance is less than a preset safety threshold, a quadratic continuous penalty is triggered.

8. The risk-aware conditional flow matching collision-free visual navigation method according to claim 6, characterized in that, The open space incentive assessment is obtained by calculating the cosine similarity between the direction vector of the candidate trajectory and the most spacious passable central axis vector extracted from the topological features of the explicit obstacle, and the collision penalty reward is used as a continuous exponential gating to attenuate and suppress the open space incentive assessment.

9. The risk-aware conditional flow matching collision-free visual navigation method according to claim 1, characterized in that, In step S4, for a specific multimodal visual observation context vector, the conditional flow matching head does not output a single action, but instead samples in parallel to extract a control group containing multiple heterogeneous waypoint candidate trajectories.

10. A collision-free visual navigation system based on risk perception and conditional flow matching, characterized in that, The risk-aware conditional flow matching collision-free visual navigation method applied to any one of claims 1-8 includes an image acquisition module, a multimodal feature fusion module, a conditional flow matching trajectory generation module, and a trajectory output and control module; The image acquisition module is used to acquire multimodal features, including temporal observation features, navigation target features, and explicit obstacle topological features; The multimodal feature fusion module is used to cascade the temporal observation features, navigation target features, and explicit obstacle topology features in the channel dimension, input them into a multi-layer Transformer encoder for long-range dependency aggregation, and output the final control context vector. The conditional flow matching trajectory generation module is used to use the control context vector as a conditional constraint term, and through the conditional flow matching model, smoothly transport and transform the standard Gaussian noise trajectory sample in continuous time steps, and output candidate trajectory control groups in parallel. The trajectory output and control module is used to perform safety assessment and preference alignment on each candidate trajectory in the candidate trajectory control group based on the offline group relative trajectory alignment mechanism, so that the trajectory distribution output by the optimized network converges to the safety mode under the same environment.

Images