A method and system for distributed spatiotemporal conflict prediction and cooperative avoidance of heterogeneous embodied agents

Abstract

This invention discloses a distributed spatiotemporal conflict prediction and cooperative avoidance method and system for heterogeneous embodied intelligent agents. The method includes: acquiring perception data and entity type labels containing at least ground mobile robots and unmanned aerial vehicles (UAVs); predicting future trajectory information and calculating conflict risk metrics; when the risk exceeds a threshold, generating safety parameters based on the entity type labels, especially generating safety constraint parameters for UAVs that include at least a downwash airflow exclusion zone and / or a vertical take-off and landing exclusion zone; and distributing these parameters to the intelligent agent for its local path planner to avoid the conflict. The system includes a heterogeneous perception access unit, a trajectory prediction unit, a risk measurement unit, a safety parameter generation unit, and a distribution unit within a group control center. This invention achieves unified conflict prediction and refined safety management for cross-dimensional heterogeneous entities, improving the efficiency of ground robots while ensuring the safety of UAV operations, and can integrate external airspace rules, possessing communication fault tolerance capabilities.

Description

Technical Field

[0001] This invention relates to the field of multi-agent cooperative control and safe collision avoidance technology, and in particular to a distributed spatiotemporal conflict prediction and cooperative avoidance method and system for heterogeneous embodied intelligent agents. Background Technology

[0002] With the rapid development of smart logistics, intelligent inspection, and the low-altitude economy, the demand for simultaneously deploying ground mobile robots (such as Automated Guided Vehicles, Autonomous Mobile Robots, and AMRs) and unmanned aerial vehicles (such as multi-rotor drones) for collaborative operations in scenarios such as warehouses, factories, and industrial parks is becoming increasingly urgent. Such systems typically involve a central control center that uniformly accesses sensor data from each individual intelligent agent and coordinates their movement trajectories to avoid collisions or human-machine collisions, ensuring operational safety and efficiency.

[0003] Existing multi-robot collision avoidance technologies mainly fall into the following categories: First, group control methods based on preset routes and centralized mandatory assignment. The group control center pre-plans fixed routes or time slices for each robot, and directly issues stop or rerouting commands when a potential conflict is detected. Second, distributed collision avoidance methods based on artificial potential fields or control barrier functions. By introducing repulsive / attractive potential fields or safety barrier constraints into the robot's motion equations, the robot can autonomously avoid collisions locally. Third, trajectory planning methods that have emerged in recent years combine the probability distribution of future trajectories predicted by generative models (such as diffusion models) with control barrier functions or potential field methods.

[0004] For example, patent document CN121954039A discloses a congestion-driven spatiotemporal trajectory planning method for heterogeneous multi-agent systems. This method includes: acquiring physical parameters and task objective data of heterogeneous agents and establishing a dynamic model; defining the spatiotemporal trajectory occupancy set of the agents and establishing collision avoidance constraints; constructing a Gaussian-weighted spatiotemporal grid and establishing a spatiotemporal congestion prediction model; constructing a scene-adaptive weight adjustment mechanism based on congestion data to dynamically optimize the objective function; establishing half-space constraint relationships between agents to form a congestion-driven dynamic spatial allocation closed loop; and generating and executing the optimal trajectory using a nonlinear programming solver based on a distributed asynchronous architecture. This scheme has a certain effect on solving the coordinated collision avoidance of homogeneous or heterogeneous robot clusters on the ground.

[0005] However, the aforementioned existing technical solutions all suffer from significant functional deficiencies and processing efficiency shortcomings, particularly in scenarios involving mixed ground and aerial operations. Firstly, the method disclosed in CN121954039A is primarily designed for ground-based mobile robots. Its spatiotemporal grid only divides the two-dimensional workspace into fixed square grids, failing to characterize and process the motion state of unmanned aerial vehicles (UAVs) in three-dimensional free space. Furthermore, its heterogeneous intelligent agent models (dual integrator type, single-wheeled type, bicycle type) are essentially two-dimensional planar motion models, lacking the ability to handle three-dimensional motion and predict cross-dimensional conflicts. This results in the control center being unable to obtain a complete perception of the global space in mixed operational spaces with UAVs, thus leading to functional deficiencies.

[0006] Secondly, existing technologies, when converting predicted congestion information into collision avoidance constraints, employ a uniform half-space constraint relationship, determining normal vectors and boundary parameters based on relative position and motion trends. This "one-size-fits-all" approach fails to consider the unique physical effects of different entity types. For example, for unmanned aerial vehicles (UAVs), the downwash generated by their rotors creates a significant dynamic influence area in the space below; during vertical takeoff and landing, they momentarily occupy a vertical columnar airspace; and the aircraft's maneuverability is strictly limited by maximum tilt angle, climb / descent rate, and other factors. These physical effects are completely absent from existing half-space constraints, resulting in safety constraints that are either insufficiently margined for UAVs or overly conservative for ground robots, thus reducing overall operational efficiency.

[0007] Furthermore, the method in CN121954039A employs a fully distributed asynchronous architecture, where each agent independently computes and negotiates with its neighbors to adjust constraints. While this architecture is scalable, it has significant shortcomings when dealing with externally enforced rules (such as no-fly zones in low-altitude airspace management systems and pre-defined restricted areas within parks). Its internal negotiation mechanism cannot provide priority guarantees for externally non-negotiable constraints. When internal avoidance requirements conflict with externally enforced rules, it lacks clear and enforceable hierarchical processing rules, making it difficult to meet the requirements for legal and safe operation within regulated airspace.

[0008] Finally, this distributed scheme heavily relies on continuous communication between agents to broadcast trajectories and constraint updates, thereby maintaining constraint consistency within the system. However, in real-world deployment environments, communication latency, packet loss, and even outages are unavoidable. Existing solutions do not provide any degradation and fault tolerance mechanisms in the event of communication anomalies. Once communication quality deteriorates, agents may lose effective avoidance parameter support, posing a serious security risk.

[0009] Therefore, there is an urgent need for a method and system that can incorporate heterogeneous embodied intelligent agents such as ground mobile robots and unmanned aerial vehicles into a unified conflict prediction framework, generate differentiated safety parameters for different entity types, and integrate external mandatory rules to provide communication fault tolerance guarantees. Summary of the Invention

[0010] To address the above problems, this invention proposes a distributed spatiotemporal conflict prediction and collaborative avoidance method and system for heterogeneous embodied intelligent agents. By constructing a cross-dimensional unified spatiotemporal representation, generating safety parameters differentiated according to entity type, integrating internal and external constraints in a hierarchical manner, and implementing a communication degradation and fault tolerance mechanism, it achieves an optimized balance between safety and efficiency in complex operational scenarios involving ground robots and unmanned aerial vehicles, significantly improving the system's overall perception capability, decision robustness, and rule compliance.

[0011] To achieve the above objectives, the following technical solution is proposed: A distributed spatiotemporal conflict prediction and cooperative avoidance method for heterogeneous embodied intelligent agents includes: Acquire perception data from multiple heterogeneous embodied intelligent agents, including at least ground mobile robots and unmanned aerial vehicles. The perception data includes pose, velocity, task target point, entity type label and capability parameters of each embodied intelligent agent. Based on the perceived data, predict the trajectory information of each individual intelligent agent within a future time window; Based on the trajectory information, calculate the conflict risk measure between each embodied intelligent agent; When the conflict risk metric exceeds a preset threshold, safety parameters are generated according to the entity type label of the embodied intelligent agent involved in the conflict, based on differentiated rules. When the entity type label indicates an unmanned aerial vehicle, differentiated safety constraint parameters are generated, including at least downwash airflow exclusion zone parameters and / or vertical take-off and landing exclusion zone parameters. The generated safety parameters are sent to the corresponding embodied intelligent agent so that the local path planner of the embodied intelligent agent can make avoidance decisions.

[0012] By employing the above technical solutions, and by introducing entity type label-driven differentiated parameter generation logic in the conflict resolution process—specifically, constructing downwash airflow exclusion zone and vertical take-off and landing exclusion zone parameters specifically for UAVs—safety constraints can accurately match the physical characteristics and motion capabilities of each entity. Compared to existing technologies that impose uniform safety boundaries on all entities, this solution effectively avoids excessive restrictions on ground robots. It also addresses the safety shortcomings of UAVs regarding the influence range of rotor downwash airflow and the instantaneous airspace occupation during take-off and landing. Thus, while ensuring overall operational safety, it maximizes the mobility efficiency of ground robots and ensures sufficient safety space for UAVs during critical phases such as take-off and landing.

[0013] Preferably, before predicting the trajectory information, the method further includes: A unified spatiotemporal lattice is constructed, in which a 2.5-dimensional hierarchical grid substructure is constructed for ground activity areas, and a 3-dimensional voxel substructure or octree substructure is constructed for air activity areas. The 2.5-dimensional hierarchical grid structure and the 3-dimensional voxel substructure or octree substructure are aligned and merged in a unified world coordinate system to form a composite data structure that can be uniformly queried. The nodes of the unified spatiotemporal lattice carry entity type labels and capability parameters of the embodied intelligent agents falling into it.

[0014] Through the above technical solutions, by establishing a unified spatiotemporal lattice that integrates quasi-two-dimensional representations of the ground and true three-dimensional representations of the air, the group control center has, for the first time, the ability to simultaneously process heterogeneous ground and air entities within the same data structure. This composite data structure supports unified indexing and neighborhood queries across dimensions, enabling the rapid acquisition of the labels and capabilities of all surrounding entities (regardless of their motion dimension) at any spatiotemporal node. This provides an efficient and accurate data foundation for subsequent cross-dimensional conflict prediction and refined safety parameter generation, overcoming the limitation of traditional two-dimensional grid methods in representing the distribution of entities in high-altitude areas.

[0015] Preferably, the prediction of the trajectory information of each embodied agent within a future time window specifically involves: based on a generative model, using the historical state, capability parameters, and task objectives of each embodied agent as conditions, predicting the probability trajectory distribution of each embodied agent within a future time window; and the calculation of the conflict risk measure between each embodied agent specifically involves: calculating the overlap integral of the probability trajectory distributions of two embodied agents in the space-time domain as the conflict risk measure.

[0016] By employing a generative model to output a probabilistic trajectory distribution and using the probabilistic overlap integral as a conflict metric, this method can quantify the uncertainty of future trajectories, thus providing early warnings and implementing preventative avoidance measures when conflicts are still at a low probability stage. Compared to schemes based on deterministic trajectory occupancy and crowding scalars, probabilistic metrics are inherently robust to sensor noise, environmental disturbances, and the randomness of agent behavior. This significantly reduces the risk of collisions and false positives caused by prediction bias, improving the accuracy and recall of conflict prediction.

[0017] As a preferred option, it also includes: Receive constraint data from external airspace management systems or park restricted area rules, and mark them as non-negotiable constraints; The non-negotiable constraints are fused with the internally generated security parameters according to a weighted rule, wherein the non-negotiable constraints are assigned a high weight coefficient, so that the non-negotiable constraints dominate in the externally enforced region.

[0018] By employing the above technical solution, and by setting external mandatory constraints as non-negotiable and assigning them dominant weight, this method establishes a clear hierarchical processing mechanism between internal collaborative efficiency and external regulatory compliance. When an agent enters a no-fly zone designated by airspace management or a pre-set isolation zone within a park, the fused safety parameters will force it to detour or leave, without violating regulatory rules due to internal collaborative needs. This mechanism enables this solution to seamlessly integrate with low-altitude traffic management systems, providing a legal compliance basis for industrial application within regulated airspace and expanding the system's deployment scope.

[0019] As a preferred option, it also includes: When the communication latency between the embodied agent and the group control center exceeds a preset threshold or the number of consecutive packet losses exceeds a preset number, the embodied agent is controlled to automatically switch to local conservative mode, using the most recently received security parameters and amplifying its security boundary by a preset coefficient until communication returns to normal.

[0020] By employing the above technical solutions, and by proactively downgrading to a local conservative mode and amplifying the safety boundary when communication link quality deteriorates, this method provides the system with crucial fail-safe capabilities. Even in extreme cases where the group control center is temporarily unreachable, each individual intelligent agent can still maintain basic obstacle avoidance functions based on the most recently valid differentiated safety parameters, avoiding collisions caused by missing or outdated parameters, and significantly improving the robustness and reliability of the system in real-world complex electromagnetic environments.

[0021] Preferably, when the entity type label indicates an unmanned aerial vehicle, the generated safety parameters specifically include: The parameters of the conical downwash airflow repulsion zone are generated based on the rotor speed and payload of the unmanned aerial vehicle. The radius of the downwash airflow repulsion zone increases with the increase of rotor speed and payload, and decreases with the increase of the height difference with the ground. Based on whether the unmanned aerial vehicle is currently in the takeoff or landing phase, a vertical columnar exclusive zone parameter is generated. The radius and height range of the exclusive zone dynamically shrink or expand as the takeoff and landing process progresses. A higher-order control barrier function constraint is introduced, with the maximum tilt angle and maximum climb / descent rate of the unmanned aerial vehicle as higher-order derivative terms of the safety constraint.

[0022] By explicitly parameterizing the mapping relationship between rotor speed, load, and downwash airflow spatial range, and dynamically adjusting the takeoff and landing exclusion zone according to the flight phase, the safety constraints of this solution can follow the changes in the UAV's operating status in real time. During hovering with a heavy load, the conical downwash airflow exclusion zone automatically expands to protect the area below; after takeoff and climb, the exclusion zone gradually shrinks to release reusable airspace. Simultaneously, the higher-order control barrier function constraint uses the maneuver envelope as a higher-order term of the safety constraint, ensuring that the generated avoidance trajectory always remains within the aircraft's dynamic capabilities, eliminating the risk of loss of control due to generating avoidance commands that exceed physical limits.

[0023] Preferably, the generation of security parameters according to the differentiation rules further includes: When the entity type label indicates a ground mobile robot, a momentum-corrected safe distance parameter is generated based on its current speed and maximum deceleration. When the entity type label indicates a human worker, a preset conservative safety factor is applied on top of the basic safety distance, and the conservative safety factor is greater than 1.

[0024] Through the above technical solutions, and by applying differentiated safety modeling based on the motion capabilities and behavioral characteristics of ground entities, this approach further optimizes the balance between safety and efficiency in ground operations. High-speed, heavy-duty AGVs achieve a longer safety clearance matching their braking distance, preventing sudden braking and material slippage; low-speed, light-duty AMRs are not forced to maintain excessive space margins. For human workers whose behavior is highly unpredictable, the additional conservative coefficient compensates for the predictive model's inability to model random behavior, significantly reducing the risk of human-machine collisions and reflecting the system's human-centered safety design philosophy.

[0025] A distributed spatiotemporal conflict prediction and cooperative avoidance system for heterogeneous embodied intelligent agents includes a group control center, wherein the group control center comprises the following mutually coupled components: The heterogeneous sensing access unit is configured to receive sensing data from multiple heterogeneous embodied intelligent agents, including at least ground mobile robots and unmanned aerial vehicles. The sensing data includes pose, velocity, task target point, entity type label and capability parameters of each embodied intelligent agent. The trajectory prediction unit is configured to predict the trajectory information of each individual intelligent agent within a future time window based on the perceived data. The risk measurement unit is configured to calculate the conflict risk measurement between each embodied intelligent agent based on the trajectory information. The safety parameter generation unit is configured to generate safety parameters according to differentiated rules based on the entity type label of the embodied intelligent agent involved in the conflict when the conflict risk measurement exceeds a preset threshold. When the entity type label indicates an unmanned aerial vehicle, differentiated safety constraint parameters are generated, including at least downwash airflow exclusion zone parameters and / or vertical take-off and landing exclusion zone parameters. The distribution unit is configured to distribute the generated safety parameters to the corresponding embodied intelligent agent so that the local path planner of the embodied intelligent agent can make avoidance decisions.

[0026] Through the above technical solutions, the system can support all the above-mentioned methodological features at the architectural level because a dedicated security parameter generation unit responsible for differentiated security logic is integrated into the group control center. The signal flow between each unit is clear, forming a complete processing pipeline from data access, prediction, risk assessment to parameter generation and distribution, ensuring efficient collaboration between centralized global perception and edge-side autonomous decision-making.

[0027] Preferably, the security parameter generation unit is further configured to: generate a potential field gain coefficient or a control barrier function parameter as the security parameter; or generate a spatiotemporal corridor or time slot reservation information as the security parameter.

[0028] Through the above technical solutions, since the safety parameter generation unit supports multiple optional parameter expression forms, this system can flexibly adapt to the input interfaces of local path planners for different types of intelligent agents. For planners using the artificial potential field method, gain coefficients are directly provided; for optimization-based planners, control barrier function constraints are provided; and for traffic management systems with time slot reservation capabilities, spatiotemporal corridor information can be provided. This design improves the system's compatibility with heterogeneous end-side controllers.

[0029] Preferably, a unified spatiotemporal lattice construction unit is also included, which is configured to: construct a 2.5-dimensional hierarchical grid substructure for the ground activity area, construct a 3-dimensional voxel substructure or octree substructure for the air activity area, and align and merge the two in a unified world coordinate system to form a composite data structure that can be uniformly queried; or, adopt an implicit occupancy representation based on point cloud for the air activity area.

[0030] Through the above technical solutions, the unified spatiotemporal lattice building block provides two optional spatial region modeling methods: discrete voxels and continuous implicit representation. The system can flexibly choose according to the scene scale and accuracy requirements. Octrees are used in large-scale, low-density spatial regions to save memory, while implicit occupancy representation is switched to in dense and complex regions to obtain higher boundary accuracy, thus achieving an optimal balance between computational overhead and spatial resolution.

[0031] Therefore, the present invention has at least the following beneficial effects: Synergistic optimization of safety and operational efficiency: By generating differentiated safety parameters based on entity type, especially by accurately modeling the downwash airflow and take-off and landing exclusive zones of unmanned aerial vehicles, the efficiency loss or insufficient safety margin caused by the "one-size-fits-all" safety boundary in existing technologies is avoided. While ensuring the safety of UAV operations, the maximum mobility of ground robots is released.

[0032] Integration of rule compliance and collaboration: External airspace management or mandatory park rules are marked as non-negotiable constraints and given dominant weights, which are then integrated with internal collaborative constraints. This enables the system to achieve optimal collaboration while complying with regulations, providing technical support for commercial operations in regulated low-altitude airspace.

[0033] Enhanced system robustness: Through a local conservative mode degradation mechanism in the event of communication failure, even if the group control center is lost, each agent can still maintain basic avoidance functions based on previously received valid parameters, which significantly improves the fault tolerance capability in real deployment environments.

[0034] Cross-dimensional unified perception and decision-making: By constructing a unified spatiotemporal lattice that integrates 2.5-dimensional ground grids and 3-dimensional voxels in the air, the gap in collaborative perception caused by the different motion dimensions of ground robots and unmanned aerial vehicles is solved. This enables the group control center to uniformly manage the overall operating space that includes the three-dimensional airspace, filling the functional gaps of existing technologies in this regard. Attached Figure Description

[0035] Figure 1 This is an architecture diagram of a heterogeneous embodied intelligent agent distributed cooperative avoidance system based on a unified spatiotemporal lattice, provided for an embodiment of the present invention.

[0036] Figure 2 A schematic diagram of a unified spatiotemporal lattice data structure provided in an embodiment of the present invention.

[0037] Figure 3 This is a flowchart illustrating a method for generating security parameters based on entity type differentiation, as provided in an embodiment of the present invention.

[0038] Figure 4 This is a top-down view of a typical scenario of a ground robot and an unmanned aerial vehicle cooperating to avoid obstacles, as provided in an embodiment of the present invention.

[0039] Figure 5 A side view of the vertical safety zone of an unmanned aerial vehicle provided in an embodiment of the present invention. Detailed Implementation

[0040] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are merely illustrative of the invention and are not intended to limit the invention.

[0041] This embodiment provides a distributed spatiotemporal conflict prediction and cooperative avoidance system and method for heterogeneous embodied intelligent agents. The overall architecture of the system is as follows: Figure 1 As shown, the system includes a swarm control center and a heterogeneous embodied intelligent agent swarm connected to the swarm control center via a communication network. The heterogeneous embodied intelligent agent swarm contains at least two types of intelligent agents with different motion dimensions, such as ground mobile robots and unmanned aerial vehicles. Additionally, it may include humanoid or legged robots. The swarm control center also communicates with external airspace management systems or park restricted area rule servers to receive external mandatory constraint data.

[0042] The group control center, acting as the global coordinator, can consist of one or more high-performance servers. In a specific hardware embodiment, the group control center includes at least one central processing unit (CPU), main memory, non-volatile memory, and a network communication interface. These hardware components are coupled via a system bus. The non-volatile memory stores the operating system and program code used to implement the functions of each processing unit of this invention. When the program code is loaded into the main memory by the CPU and executed, the server becomes a functional entity comprising a heterogeneous sensing access unit, a unified spatiotemporal lattice construction unit, a probability trajectory prediction unit, a spatiotemporal intersection risk measurement unit, a security parameter generation unit, an external constraint fusion unit, and an edge-side decision-making unit. The signal flow of each unit is as follows: The heterogeneous sensing access unit receives sensing data packets uploaded by each agent through the network communication interface, parses them, and sends them to the unified spatiotemporal lattice construction unit and the probability trajectory prediction unit respectively; the unified spatiotemporal lattice data output by the lattice construction unit and the probability trajectory distribution output by the probability trajectory prediction unit are input to the risk measurement unit; the conflict risk measurement value calculated by the risk measurement unit triggers the safety parameter generation unit to generate internal negotiated safety parameters; the external constraint fusion unit receives external mandatory constraints and marks them as non-negotiable constraints; subsequently, the internal safety parameters output by the safety parameter generation unit and the non-negotiable constraints output by the external constraint fusion unit are input to the fusion module to generate the final safety parameters; finally, the distribution unit sends these final safety parameters to the corresponding agents through the network interface.

[0043] Each intelligent agent can be an AGV, AMR, multi-rotor drone, or bipedal robot. Each agent possesses its own computing power (such as an embedded processor), sensor suite (LiDAR, visual odometry, inertial measurement unit, GPS, etc.), and wireless communication module. Each agent runs a local path planner, such as a planner based on model predictive control or dynamic windowing, which receives safety parameters from the group control center as constraints and, combined with its own real-time sensor data, autonomously solves for a local avoidance trajectory that satisfies its own dynamic constraints.

[0044] Below, in conjunction with Figure 3 The method flowchart details the specific implementation steps and signal processing of the method in this embodiment.

[0045] Step 1: Heterogeneous Sensing Access: The heterogeneous sensing access unit of the swarm control center continuously receives sensing data uploaded by each agent through the communication network. For each agent, the uploaded data message contains at least the following fields: timestamp, pose vector (position coordinates x, y, z, attitude angles such as roll, pitch, yaw), velocity vector (linear velocity and angular velocity), target point coordinates, and a capability parameter vector pre-registered and stored locally in the agent or in the swarm control center's database. This capability parameter vector includes geometric envelope dimensions, maximum speed, maximum acceleration / deceleration, stopping distance, and turning radius. Specifically, for unmanned aerial vehicles (UAVs), the capability parameter vector additionally includes the number of rotors, rated speed range, maximum payload, maximum climb / descent rate, and maximum tilt angle. Each agent's data message also carries an entity type label field; for example, 0x01 represents a ground mobile robot, 0x02 represents a multi-rotor UAV, and 0x03 represents a humanoid robot. The swarm control center distinguishes entity categories based on this label.

[0046] Step 2: Unified Spatiotemporal Lattice Construction: The unified spatiotemporal lattice construction unit dynamically maintains a unified spatiotemporal lattice containing the time dimension based on the real-time positions reported by all agents. (Reference) Figure 2The lattice is constructed as follows: For the ground-based activity area (z-coordinate from the ground to below a certain height threshold), a 2.5-dimensional hierarchical lattice substructure is constructed. This substructure is essentially a two-dimensional grid map, with each grid node having an attached "floor" or height attribute to distinguish different platforms or elevated areas. For the aerial activity area, a three-dimensional voxel substructure is constructed, that is, the three-dimensional spatial domain is discretized using cubes (voxels) of fixed size. To save storage and traversal overhead, this embodiment uses an octree structure to manage aerial voxels. The root node of the octree corresponds to the entire aerial monitoring range, and then it is recursively divided into eight child nodes until the preset minimum voxel size is met. In addition, the resolution of the voxels is not fixed, but adaptively adjusted with height and local entity density. For example, a resolution of 0.5 meters is used in low-altitude dense entity areas, and reduced to 2 meters in high-altitude sparse areas.

[0047] The 2.5D and 3D substructures are aligned and merged in a unified world coordinate system. Specifically, a unified 2D horizontal index array is established based on the horizontal projection coordinates (x, y) of the 2.5D substructure. For any 3D voxel or octagonal leaf node, its projection range on the horizontal plane is calculated, and its pointer or index is mapped to the unified index list corresponding to that horizontal coordinate. This forms a composite data structure: using the (x, y) coordinate, all relevant ground grid nodes and aerial voxel nodes across all height ranges from lowest to highest can be quickly queried. Each lattice node (ground grid or aerial voxel) maintains a dynamic list storing the entity type labels and capability parameter vectors of all agents currently falling within its spatial range. This information is updated with each perception data update cycle.

[0048] In another implementation, modeling of airborne activity regions can employ a point cloud-based implicit occupancy representation, such as using a lightweight neural occupancy field to continuously model the spatial domain. The spatiotemporal lattice query approach is correspondingly replaced by a continuous position query of this implicit occupancy field function to obtain the probability of occupancy at that location, instead of discrete voxel indexing. This approach can achieve higher spatial accuracy when representing irregular obstacle boundaries.

[0049] Step 3: Probabilistic Trajectory Prediction: The probabilistic trajectory prediction unit predicts the trajectory probability distribution within a future time window N based on the historical state sequence of each agent in a unified spatiotemporal lattice, its capability parameters, and the task target point. This embodiment uses a diffusion model as the generative model, but it can also be replaced by a Transformer-based probabilistic trajectory generation model or a conditional variational autoencoder. The diffusion model is a model that generates data distribution by progressively denoising. Its inputs are the state sequence of the agent over the past H time steps, the environmental context (information of surrounding lattice nodes), and the target position; the output is the position probability distribution p(x,t) over the next N time steps.

[0050] The length of the prediction time window N is adaptively adjusted based on the agent's current state and communication status, using the following formula: N=N0+k1*(v / v_ref)+k2*delta_t_comm+k3*(1 / f_sample); Where N0 is the base prediction duration, set to 5 seconds; v is the current entity speed; v_ref is the reference speed, set to 1.5 m / s; delta_t_comm is the current communication delay between the group control center and the entity, in seconds; f_sample is the sensor sampling frequency; k1, k2, and k3 are pre-calibrated weighting coefficients, in this embodiment k1=1.0, k2=0.5, and k3=0.2. For example, for a ground robot with a current speed of 3 m / s, a communication delay of 0.1 seconds, and a sampling frequency of 10 Hz, its prediction duration N=5+1.0(3 / 1.5)+0.50.1+0.2*(1 / 10)=5+2+0.05+0.02=7.07 seconds. Thus, the robot's future trajectory will be predicted 7.07 seconds later. The faster the speed, the greater the delay, and the lower the sampling frequency, the longer the prediction window, providing sufficient lead time.

[0051] Step 4, Spatiotemporal Intersection Risk Measurement: The spatiotemporal intersection risk measurement unit receives the probability trajectory distributions of all agents. For any two agents i and j, based on their probability trajectory distributions p_i(x,t) and p_j(x,t), the spatiotemporal intersection risk measurement value R_ij is calculated. The risk measurement value is defined as the overlapping integral of the two probability density functions in space and time: R_ij = ∬p_i(x,t)·p_j(x,t)dxdt; This integral can be calculated numerically, for example, by discretizing the prediction time and spatial regions and transforming the integral into a summation. The larger the value of R_ij, the higher the probability that the future trajectories of the two agents will occupy the same spatiotemporal region, and the greater the risk of conflict.

[0052] Set a risk threshold R_th, for example, 0.15 (this threshold is an empirical value and needs to be calibrated in conjunction with the actual system). When the R_ij of any pair of agents exceeds R_th, the subsequent safety parameter generation step is triggered. In another embodiment, the conflict risk measure can also be the Mahalanobis distance between the mean of the trajectory probability distributions of the two agents. When the Mahalanobis distance is less than the preset safety threshold, the subsequent safety parameter generation step is triggered.

[0053] Step 5: Generate safety parameters differentiated by entity type: This is the core technical feature of the present invention. The safety parameter generation unit first calculates a general risk-potential field mapping based on the risk metric value R_ij. First, the repulsive gain coefficient k_repulse=g(R_ij) is generated by the monotonically increasing function g(). For example, a linear function k_repulse=k_max*(R_ij / R_th) can be used, where k_max is the maximum gain.

[0054] Then, based on the type labels of the entities involved in the conflict, the security parameters required by the local planner of each embodied agent are generated according to the following differentiation rules.

[0055] When the entity type is an unmanned aerial vehicle: (1) Generate downwash airflow repulsion zone parameters. The repulsion zone is a conical space whose radius r_downwash increases with rotor speed omega and load m_load, and decreases with the height difference delta_h above the ground. The calculation formula is: r_downwash=r_base*(omega / omega_rated)^alpha*(1+beta*m_load / m_max)*exp(-gamma*delta_h); Where r_base is the base radius of influence, taken as twice the drone's wheelbase; omega_rated is the rated rotational speed; m_max is the maximum payload; and alpha, beta, and gamma are adjustment coefficients. For example, for a drone with a rotor diameter of 0.5 meters (r_base is 1 meter), hovering at a height of 1 meter above the ground, with a rotor speed of 80% of the rated speed and a payload of 50% of the maximum payload, taking alpha=1.5, beta=0.8, and gamma=0.5, the calculated radius of influence of the downwash airflow is approximately 1*(0.8^1.5)*(1+0.8*0.5)*exp(-0.5*1)≈1*0.716*1.4*0.607≈0.61 meters. This repulsion region parameter will be sent as a local repulsive potential field gain to the drone and the agents below it that may be affected.

[0056] (2) Generate vertical cylindrical exclusive zone parameters. Based on whether the UAV is currently in takeoff or landing, generate a cylindrical exclusive zone extending upwards from the ground or landing platform. The radius r_launch and height h_launch of this zone are dynamically adjusted during takeoff and landing. During takeoff and landing, r_launch is set to 1.5 times the UAV rotor diameter plus a safety buffer, for example, 1.2 meters; the height h_launch is set to the UAV's current target altitude plus a buffer. Once the UAV climbs to its cruising altitude and stabilizes, this exclusive zone gradually shrinks until it is eliminated. This parameter is issued as a hard constraint.

[0057] (3) Introduce higher-order control barrier function constraints. To ensure that the generated avoidance trajectory does not exceed the maneuverability of the UAV, its maximum tilt angle phi_max, maximum climb rate v_z_max, etc., are used as higher-order derivative terms of the control barrier function. For example, a second-order control barrier function h(x) is constructed, whose first and second derivatives involve tilt angular velocity and climb acceleration. The constraint conditions ensure that the second derivative of h(x) always meets the safety conditions under control input, thereby avoiding issuing overly abrupt turning or ascent / descending commands.

[0058] When the entity type is a ground mobile robot: The momentum-corrected safe distance parameter d_safe is generated based on its current velocity v and maximum deceleration a_brake. The calculation formula is: d_safe=d0+v*t_brake+v^2 / (2*a_brake); Where d0 is the basic static safety distance, taken as 1.5 times the diagonal length of the robot's geometric envelope; t_brake is the system response and brake activation delay time, taken as 0.2 seconds. For example, for an AGV with a speed v = 2 m / s, a maximum deceleration a_brake = 1.5 m / s², and d0 = 0.8 m, then d_safe = 0.8 + 2 * 0.2 + 2^2 / (2 * 1.5) = 0.8 + 0.4 + 4 / 3 ≈ 0.8 + 0.4 + 1.33 = 2.53 m. This distance is converted into the repulsion radius parameter of the local potential field.

[0059] When the entity type label indicates a human worker: A preset conservative safety factor, lambda_human, with a value of 1.8, is applied to the basic safety distance d0_human. Therefore, the radius of the safety zone around the human worker is expanded to lambda_human * d0_human. This compensates for the high randomness of human behavior.

[0060] In another implementation, the generated safety parameters can take the form of a spatiotemporal corridor instead of potential field gain coefficients or CBF parameters. The group control center reserves a time window of exclusive passage within the conflict zone for embodied agents at risk of conflict, thus defining a spatiotemporal corridor. This spatiotemporal corridor information is distributed to other agents, whose local path planners treat it as a dynamic obstacle constraint to avoid. This approach facilitates integration with external airspace management systems based on time slot reservation mechanisms.

[0061] Step Six: External Constraint Fusion: The external constraint fusion unit receives constraint data from external airspace management systems or preset restricted areas within the park. This data typically defines no-fly or no-access zones in the form of polygons or voxel sets. This unit marks these external constraints as non-negotiable constraint terms U_hard. Simultaneously, it receives internally negotiable safety parameters output by the safety parameter generation unit; these parameters are typically represented as potential field function terms U_soft. Weighted rule fusion is employed. U_total = U_soft + M * U_hard; Here, M is a preset high-weight coefficient with a value of 100. This ensures that within the externally forced region, the total potential field U_total is primarily dominated by U_hard, generating a strong repulsive force that compels the agent to avoid it. Outside the external region, U_hard is 0, and the internal negotiated parameter U_soft plays a crucial role. For numerical stability, U_hard can be a continuous barrier function, smoothly transitioning from 0 to a high value near the boundary to avoid step jumps that could cause the solver to fail to converge.

[0062] Step 7: Autonomous Local Avoidance at the Terminal Side: The terminal decision-making unit of the group control center encapsulates the fused final safety parameters (including repulsion gain coefficient, safety distance, repulsion zone radius, exclusive zone range, control barrier function constraint matrix, etc.) into data packets and sends them to the corresponding embodied agents via the network. Each agent's local path planner (such as a planner based on the dynamic window method) directly loads these parameters into its objective function or constraint set. For example, in the objective function of the dynamic window method, the weight of the cost term for distance to obstacles is replaced by the received repulsion gain coefficient; during velocity sampling, the introduced safety distance d_safe serves as a hard constraint for maintaining the minimum distance with other agents. Under the premise of satisfying its own acceleration / deceleration, turning radius, and other dynamic constraints, the planner uses rolling optimization to solve for the optimal local avoidance speed command, thereby achieving autonomous avoidance. The group control center does not need to issue specific path points or forced shutdown commands, preserving the flexibility of terminal execution.

[0063] Step 8, Communication Fault Tolerance Degradation: Each agent continuously monitors the communication quality with the group control center while receiving security parameters. For example, it calculates the average round-trip time every 100 milliseconds and counts the number of packet losses in the last 10 seconds. When the average latency exceeds a preset threshold (e.g., 500 milliseconds, this is an example value and needs to be calibrated according to the system's real-time requirements) or the number of consecutive packet losses exceeds a preset number (e.g., 3 times), the agent's communication management module determines that the communication is abnormal and triggers a state switch to enter a local conservative mode. In conservative mode, the agent ignores subsequent unstable parameter updates and instead uses the last successfully received and verified security parameters. At the same time, its security boundaries (such as security distance, exclusion zone radius, etc.) are amplified by a preset coefficient, such as 1.5 times, to compensate for the risk of parameters becoming outdated. When the communication quality recovers and remains stable for one cycle (e.g., 10 seconds), it automatically switches back to normal mode and resynchronizes the latest security parameters. This mechanism ensures basic operational safety under extreme conditions.

[0064] Figure 4 This demonstration shows a scenario in a warehouse environment where a ground mobile robot is traveling along its planned path (pointed line) while an unmanned aerial vehicle (UAV) is landing. After detecting a potential conflict between their future trajectories, the swarm control center generates and distributes differentiated safety parameters based on this method. The UAV receives its vertical columnar exclusion zone and downwash repulsion zone parameters, while the ground robot receives adjusted safety distance and potential field parameters. Ultimately, the ground robot's local path planner autonomously solves for a smooth, avoidance trajectory (solid arrow), while the UAV lands safely.

[0065] Figure 5 The side view illustrates the safety zone during the takeoff and landing phase of the unmanned aerial vehicle (UAV). The downwash repulsion zone is cone-shaped, with its radius r_downwash decreasing with altitude. The vertical, columnar exclusive zone, with radius r_launch, extends upwards from the takeoff and landing point, providing the UAV with a dedicated spatial passage. Ground-based mobile robots navigate around this safety zone based on these parameters.

[0066] As can be seen from the above embodiments, the present invention, through a series of carefully designed signal processing flows and unit modules, completely realizes unified, refined, compliant, and robust cooperative obstacle avoidance control for heterogeneous embodied intelligent agents, including those on the ground and in the air. The above descriptions are merely preferred embodiments of the present invention and are not intended to limit the invention. Any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A distributed spatiotemporal conflict prediction and cooperative avoidance method for heterogeneous embodied intelligent agents, characterized in that, include: Acquire perception data from multiple heterogeneous embodied intelligent agents, including at least ground mobile robots and unmanned aerial vehicles. The perception data includes pose, velocity, task target point, entity type label and capability parameters of each embodied intelligent agent. Based on the perceived data, predict the trajectory information of each individual intelligent agent within a future time window; Based on the trajectory information, calculate the conflict risk measure between each embodied intelligent agent; When the conflict risk metric exceeds a preset threshold, safety parameters are generated according to the entity type label of the embodied intelligent agent involved in the conflict, based on differentiated rules. When the entity type label indicates an unmanned aerial vehicle, differentiated safety constraint parameters are generated, including at least downwash airflow exclusion zone parameters and / or vertical take-off and landing exclusion zone parameters. The generated safety parameters are sent to the corresponding embodied intelligent agent so that the local path planner of the embodied intelligent agent can make avoidance decisions.

2. The distributed spatiotemporal conflict prediction and cooperative avoidance method for heterogeneous embodied intelligent agents according to claim 1, characterized in that, Before predicting the trajectory information, the following is also included: A unified spatiotemporal lattice is constructed, in which a 2.5-dimensional hierarchical grid substructure is constructed for ground activity areas, and a 3-dimensional voxel substructure or octree substructure is constructed for air activity areas. The 2.5-dimensional hierarchical grid structure and the 3-dimensional voxel substructure or octree substructure are aligned and merged in a unified world coordinate system to form a composite data structure that can be uniformly queried. The nodes of the unified spatiotemporal lattice carry entity type labels and capability parameters of the embodied intelligent agents falling into it.

3. The distributed spatiotemporal conflict prediction and cooperative avoidance method for heterogeneous embodied intelligent agents according to claim 1, characterized in that, The prediction of the trajectory information of each embodied agent within a future time window specifically involves: based on a generative model, using the historical state, capability parameters, and task objectives of each embodied agent as conditions, predicting the probability trajectory distribution of each embodied agent within a future time window; and the calculation of the conflict risk measure between each embodied agent specifically involves: calculating the overlap integral of the probability trajectory distributions of two embodied agents in the space-time domain as the conflict risk measure.

4. The distributed spatiotemporal conflict prediction and cooperative avoidance method for heterogeneous embodied intelligent agents according to claim 1, characterized in that, Also includes: Receive constraint data from external airspace management systems or park restricted area rules, and mark them as non-negotiable constraints; The non-negotiable constraints are fused with the internally generated security parameters according to a weighted rule, wherein the non-negotiable constraints are assigned a high weight coefficient, so that the non-negotiable constraints dominate in the externally enforced region.

5. The distributed spatiotemporal conflict prediction and cooperative avoidance method for heterogeneous embodied intelligent agents according to claim 1, characterized in that, Also includes: When the communication latency between the embodied agent and the group control center exceeds a preset threshold or the number of consecutive packet losses exceeds a preset number, the embodied agent is controlled to automatically switch to local conservative mode, using the most recently received security parameters and amplifying its security boundary by a preset coefficient until communication returns to normal.

6. The distributed spatiotemporal conflict prediction and cooperative avoidance method for heterogeneous embodied intelligent agents according to claim 1, characterized in that, When the entity type label indicates an unmanned aerial vehicle, the generated safety parameters specifically include: The parameters of the conical downwash airflow repulsion zone are generated based on the rotor speed and payload of the unmanned aerial vehicle. The radius of the downwash airflow repulsion zone increases with the increase of rotor speed and payload, and decreases with the increase of the height difference with the ground. Based on whether the unmanned aerial vehicle is currently in the takeoff or landing phase, a vertical columnar exclusive zone parameter is generated. The radius and height range of the exclusive zone dynamically shrink or expand as the takeoff and landing process progresses. A higher-order control barrier function constraint is introduced, with the maximum tilt angle and maximum climb / descent rate of the unmanned aerial vehicle as higher-order derivative terms of the safety constraint.

7. The distributed spatiotemporal conflict prediction and cooperative avoidance method for heterogeneous embodied intelligent agents according to claim 1, characterized in that, The generation of security parameters according to differential rules also includes: When the entity type label indicates a ground mobile robot, a momentum-corrected safe distance parameter is generated based on its current speed and maximum deceleration. When the entity type label indicates a human worker, a preset conservative safety factor is applied on top of the basic safety distance, and the conservative safety factor is greater than 1.

8. A distributed spatiotemporal conflict prediction and cooperative avoidance system for heterogeneous embodied intelligent agents, applicable to the distributed spatiotemporal conflict prediction and cooperative avoidance method for heterogeneous embodied intelligent agents as described in any one of claims 1-7, characterized in that, Includes a group control center, which comprises the following interconnected components: The heterogeneous sensing access unit is configured to receive sensing data from multiple heterogeneous embodied intelligent agents, including at least ground mobile robots and unmanned aerial vehicles. The sensing data includes pose, velocity, task target point, entity type label and capability parameters of each embodied intelligent agent. The trajectory prediction unit is configured to predict the trajectory information of each individual intelligent agent within a future time window based on the perceived data. The risk measurement unit is configured to calculate the conflict risk measurement between each embodied intelligent agent based on the trajectory information. The safety parameter generation unit is configured to generate safety parameters according to differentiated rules based on the entity type label of the embodied intelligent agent involved in the conflict when the conflict risk measurement exceeds a preset threshold. When the entity type label indicates an unmanned aerial vehicle, differentiated safety constraint parameters are generated, including at least downwash airflow exclusion zone parameters and / or vertical take-off and landing exclusion zone parameters. The distribution unit is configured to distribute the generated safety parameters to the corresponding embodied intelligent agent so that the local path planner of the embodied intelligent agent can make avoidance decisions.

9. A distributed spatiotemporal conflict prediction and cooperative avoidance system for heterogeneous embodied intelligent agents according to claim 8, characterized in that, The security parameter generation unit is further configured to: generate a potential field gain coefficient or a control barrier function parameter as the security parameter; or generate a spatiotemporal corridor or time slot reservation information as the security parameter.

10. A distributed spatiotemporal conflict prediction and cooperative avoidance system for heterogeneous embodied intelligent agents according to claim 8, characterized in that, It also includes a unified spatiotemporal lattice construction unit, which is configured to: construct a 2.5-dimensional hierarchical grid substructure for ground activity areas, construct a 3-dimensional voxel substructure or octree substructure for air activity areas, and align and merge the two in a unified world coordinate system to form a composite data structure that can be uniformly queried; or, adopt implicit occupancy representation based on point clouds for air activity areas.

Images