Multi-robot cooperative control system and method based on unified field dynamic coupling

Abstract

The invention discloses a multi-robot cooperative control system and method based on unified field dynamic coupling. According to the system, a continuous space-time unified field fusing an environment and a task is constructed through a field modeling module; the field dynamics optimization module predicts, evolves and optimizes field coupling parameters of each robot based on a field model; each robot obtains a local field gradient through a distributed interface and is driven by a gradient tracking controller to move. According to the method, cooperative control is converted into a continuous evolution and local tracking process of a field, low-delay decision making, high adaptability and strong extensible cooperation of a multi-robot system are achieved, and the performance bottleneck of a traditional centralized, negotiated and potential field method in a dynamic complex scene is effectively solved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and robot control, specifically to a distributed collaborative perception, decision-making and control framework for multi-robot systems, and more particularly to a collaborative control system and method based on a unified representation and dynamic coupling of continuous spatiotemporal physical fields. Background Technology

[0002] The application of multi-robot systems in complex scenarios faces three major challenges: coordination delays, poor environmental adaptability, and scalability bottlenecks. Existing technologies mainly suffer from the following paradigm limitations: 1. Centralized planning paradigm: It relies on a central server for global calculation and scheduling, which has the risk of single point of failure. The communication and computing load increases linearly with the number of robots, the replanning delay is significant, and it is difficult to cope with highly dynamic environments.

[0003] 2. Distributed negotiation paradigm: Robots negotiate tasks and resolve conflicts through frequent communication. The communication overhead increases in combination, the convergence speed is slow, and the system performance drops sharply or fails when communication is restricted or interrupted.

[0004] 3. Reaction-based potential field paradigm: Typical examples include the artificial potential field method, where the robot relies on a pre-defined, static attractive / repulsive potential field for local obstacle avoidance. Its fundamental flaw lies in the disconnect between the field model and the task / robot state. The potential field typically only encodes simple geometric obstacle avoidance rules, failing to represent complex task objectives and higher-order constraints; it lacks the ability to predict environmental dynamics and group interactions, easily leading to local optima or oscillations; and the simple superposition of individual behaviors makes it difficult to generate efficient global collaborative patterns.

[0005] The common problem with the above paradigms is that they treat environmental perception, task understanding, individual decision-making, and group collaboration as separate functional modules. Information is transmitted and transformed discretely and sequentially between modules, leading to semantic gaps and accumulated delays. The root cause lies in the lack of a unified and continuous mathematical model that can describe the physical properties of the environment, task semantics, robot states, and their dynamic interactions. Summary of the Invention

[0006] Purpose of the invention This invention aims to overcome the aforementioned shortcomings and provide a multi-robot cooperative control system and method based on unified field dynamic coupling. Its core lies in proposing a fundamental paradigm shift: constructing a continuous spatiotemporal "field" that integrates the environment, task, and robot state, transforming cooperative control from a discrete "planning-communication-execution" mode to a continuous "field evolution-local following" mode, thereby achieving low-latency, highly adaptive, and strongly scalable self-organizing cooperation.

[0007] Technical solution The core idea of ​​this invention can be summarized as follows: environment is field, task is field source, robot is field regulator, and control is field gradient.

[0008] In this invention, the unified field Φ(r,t) can logically be composed of information fusion from multiple sources or levels, including but not limited to: a geometric field reflecting the static structure of the environment, a task field encoding task objectives and semantics, and an interactive field or risk field characterizing dynamic obstacles or the influence of other robots.

[0009] A multi-robot collaborative control system based on unified field dynamic coupling, characterized in that it includes: The field modeling module is configured to integrate multi-source perception data and high-level task instructions to construct and update in real time a spatiotemporal unified field Φ(r, t) defined on the robot's workspace. This field encodes at least the environmental accessibility attributes, task target information, and the implicit cooperative relationships between robots. The field dynamics optimization module, connected to the field modeling module, is configured to: perform short-time domain prediction of the unified field Φ(r,t) based on a preset field dynamics evolution model; and dynamically adjust the coupling parameters between each robot and the unified field through an optimization algorithm so that the predicted field evolution trajectory optimally approximates the task objective. A distributed field interaction interface, deployed on each robot body, is configured to: subscribe to real-time field information and gradient ∇Φ of its neighborhood from the field modeling module; and receive coupling parameter update instructions for the robot from the field dynamics optimization module. The gradient tracking controller, integrated into each robot body, is configured to generate low-level motion control commands based on the local field gradient information provided by the distributed field interaction interface and the robot's own dynamic model, thereby driving the robot to move along the field gradient direction.

[0010] The field dynamics evolution model is as follows: ∂Φ(r, t) / ∂t = D∇²Φ(r, t) + f_task(Φ, r, t) + Σ_i g_i(Φ, s_i, u_i(t-Δt), θ_i) in, r represents the spatial coordinates, and D represents the diffusion coefficient, which characterizes the propagation and smoothing of field information. f_task is the task source function, used to inject the task target; s_i is the real-time state vector of the i-th robot (including position, velocity, and load); u_i(t-Δt) is the control input generated in the previous control cycle; g_i is the dynamic field coupling function of the i-th robot. Its output depends on the robot's real-time state s_i, the control input u_i(t-Δt) of the previous cycle, and the adjustable coupling parameter θ_i, which is used to express the robot's local influence and interaction on the field.

[0011] This design avoids the algebraic loop problem caused by the control input u_i depending on the current field gradient.

[0012] The design of the dynamic field coupling function g_i requires establishing a mapping relationship between the robot state s_i and the local field influences. Typical implementations include, but are not limited to: The product form is: g_i = k(θ_i) · h(s_i) · ψ(u_i(t-Δt)), where h(s_i) can reflect physical characteristics such as load and speed (e.g., h(s_i)=m_i ×‖v_i‖), and ψ(·) is the direction-dependent kernel function (e.g., Gaussian type).

[0013] Neural network mapping: g_i = NN(θ_i, s_i, u_i(t-Δt); W), which learns network parameters W online to adapt to complex scenarios.

[0014] Superposition of field perturbations: g_i = Σ_k α_k(θ_i) · φ_k(r - r_i, s_i), where φ_k is a preset basis function (such as a dipole field or a vortex field).

[0015] Those skilled in the art can choose or innovate the implementation form according to application requirements.

[0016] A multi-robot cooperative control method based on unified field dynamic coupling is characterized by the following steps: S1: Unified field construction and initialization: Integrate perception and task information to generate an initial unified field Φ_0 and initialize the field dynamics model parameters; S2: Field Prediction and Cooperative Optimization: Based on the field state Φ(t_k) at the current time t_k, within a prediction time domain, to simplify the calculation and ensure the solvability of the problem, only the effects of the task source term f_task and the diffusion term D∇²Φ are considered during prediction, and the influence of the robot dynamic coupling term Σ_i g_i is temporarily ignored. The field coupling parameters {θ_i} of each robot are used as optimization variables, and the field evolution trajectory that minimizes the task cost function is solved in a rolling manner. The optimal coupling parameters {θ_i(t_{k+1})} at the next time t_{k+1} are extracted. S3: Distributed field distribution and parameter distribution: Distribute the compressed representation of the current field Φ(t_k) and gradient information to each robot, and at the same time distribute the optimized coupling parameters {θ_i(t_{k+1})} to the corresponding robot; S4: Local Field Reaction and Gradient Tracking: Within the time interval [t_k, t_{k+1}), each robot's gradient tracking controller uses the latest issued parameter θ_i(t_{k+1}) in each control cycle. In each control cycle, the robot first acts on the local field through its dynamic field coupling function g_i based on the received coupling parameter θ_i(t_{k+1}) and the control input u_i(t-Δt) from the previous cycle. Then, based on the updated local field gradient ∇Φ, it calculates the control input u_i(t) for the current cycle and executes the motion. S5: Field state update and closed loop: Based on the robot's actual motion and new sensing data, update the unified field to Φ(t_{k+1}), return to step S2, and form a closed loop of perception-field evolution-control.

[0017] Beneficial effects 1. Extremely low latency decision-making: Control commands are derived from local field gradient calculations, which are simple to calculate and avoid complex global planning or iterative negotiation. End-to-end latency can be as low as milliseconds.

[0018] 2. Embedded adaptive capability: The unified field is sensitive to environmental changes and can be updated smoothly. By adjusting the coupling parameter θ_i and the field model parameters online, the system can quickly adapt to unknown environments and dynamic tasks.

[0019] 3. Strong Scalability: The core computational complexity of the system (field prediction and optimization) is mainly related to the field resolution. By adopting distributed optimization algorithms (such as the Alternating Direction Multiplier Method (ADMM), the global optimization problem is decomposed into sub-problems associated with each robot and solved in parallel. This effectively controls the growth of the core computational time of the system, which is far lower than the linear growth of the number of robots.

[0020] 4. Emergent Collaboration and Safety Assurance: By designing a coupling function g_i, simple local gradient tracking can spontaneously generate complex and orderly group behaviors (such as self-organized formations and dynamic traffic flows). Simultaneously, the risk field can be naturally embedded into the unified field Φ, making obstacle avoidance an intrinsic property of the control law.

[0021] 5. Communication robustness: The system is tolerant of communication delays and interruptions. The robot can operate autonomously in the short term relying on local field gradients, and the central field model gradually integrates global information through diffusion terms to restore consistency.

[0022] To clearly demonstrate the comprehensive advantages of this invention over existing technological paradigms, a comparison is made from key performance dimensions as shown in the table below: Table 1: Comparison and Analysis of Key Performance Dimensions between this Invention and Existing Technological Paradigms Comparison Dimensions Typical limitations of existing technologies Technical effects of the invention The core mechanism of the invention on which it is based Real-time decision-making Centralized planning or distributed negotiation introduces significant latency (often reaching hundreds of milliseconds). Achieve millisecond-level end-to-end control latency. The control law is simplified to local field gradient calculation, avoiding complex global planning and multiple rounds of negotiation. System scalability The central computing load or inter-node communication overhead increases linearly or in combination with the number of robots. The core computing load is weakly correlated with the number of robots, supporting ultra-large-scale cluster collaboration. Field information broadcasting (O(1) communication overhead) and distributed parallel optimization (such as ADMM algorithm) are adopted. Environmental and task adaptability Models based on fixed rules or trained offline have poor generalization ability in dynamic and unknown environments. It possesses online learning and adaptive capabilities, enabling it to quickly adapt to unknown environments and dynamic tasks. Field model parameters (such as diffusion coefficient and fusion weight) can be optimized online based on real-time running data. Collaborative Intelligence and Security Simple rules are unlikely to lead to complex collaborations; security often relies on external monitoring or post-hoc correction. It achieves self-organization and emergent collaboration, and embeds proactive safety guarantees into the control law. Designed using a dynamic coupling function g_i; the risk field is naturally integrated into the unified field, and equipped with a safety circuit breaker mechanism based on real-time sensors. Communication robustness It is highly dependent on the central node or network connectivity, and communication interruption can easily lead to system-level failure. It is tolerant of communication delays and interruptions, and supports graceful degradation and secure autonomous operation. The robot can operate autonomously in the short term by relying on local field gradients; the field model continuously integrates the global state through diffusion terms. Note: The "Typical Limitations of the Prior Art" mentioned in this table correspond to the inherent defects of centralized, distributed negotiation, and reactive potential field paradigms analyzed in the background section of this specification. The effectiveness of this invention has been verified through simulation tests in the specific embodiments. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the overall system architecture of the present invention.

[0024] Figure 2 This is a schematic diagram of the unified field Φ of the present invention (showing the integration of the geometric layer, semantic layer, and dynamic layer).

[0025] Figure 3 This is a flowchart of the prediction and optimization process based on the field dynamics model of this invention.

[0026] Figure 4 This is a schematic diagram illustrating the principle of local gradient tracking and field reaction in the robot of this invention.

[0027] Figure 5 This is a schematic diagram of the scenario and field evolution of Embodiment 1 of the present invention (cooperative scheduling of warehouse AGVs).

[0028] Figure 6 This is a diagram showing the field entropy change and robot trajectory in Embodiment 2 of the present invention (UAV swarm coverage). Detailed Implementation

[0029] The present invention will be further described below with reference to the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of the invention.

[0030] It should be noted first that the control flow of this system is based on discrete-time stepping. The optimization period T_opt (e.g., 100ms) of the field dynamics optimization module is decoupled from the control period T_ctrl (e.g., 20ms) of the robot's underlying gradient tracking controller. The optimization module asynchronously calculates and updates the coupling parameters, while the tracking controller performs real-time gradient calculation and control within each T_ctrl cycle using the latest field information and coupling parameters. The optimized coupling parameters θ_i(t_{k+1}) remain constant within the time interval [t_k, t_{k+1}) and are repeatedly used by the gradient tracking controller in various control cycles within this interval.

[0031] Example 1: Collaborative Scheduling of Multiple AGVs in Warehouse Logistics like Figure 5 As shown, in a large warehouse scenario, dozens of AGVs perform shelf handling tasks. The simulation verification in this embodiment is based on ROS (Robot Operating System) and the Gazebo simulation platform, and a standard warehouse model with 100 AGVs and an area of ​​2000 square meters is constructed.

[0032] Unified field construction: Φ(r,t) = w_c C(r) + w_t T(r,t) - w_d Σ_i R_i(r, s_i).

[0033] C(r): Static access cost field constructed based on laser SLAM map, with low value at the passageway and high value at the shelf area.

[0034] T(r,t): Dynamic task field, which becomes a "potential trap" when generated at the location of the shelf to be picked up or placed, and its depth is related to the urgency of the task.

[0035] R_i: The repulsion field generated by the i-th AGV, whose strength is positively correlated with the real-time load status of the AGV, and whose spatial distribution is directional: the repulsion intensity generated in the direction of the current AGV's velocity vector is greater than that in the opposite direction, thus forming a directional "motion trend field", which effectively avoids rear-end collisions and guides lateral avoidance.

[0036] Dynamic Coupling and Optimization: The field dynamics optimization module adopts a model predictive control (MPC) framework, periodically using the coupling parameters θ_i of each AGV as optimization variables at intervals of T_opt (e.g., 100ms) to solve for the optimal field evolution in the next few seconds. The optimization objective is to minimize the total task completion time and the degree of traffic congestion.

[0037] Control and Effects: Each AGV's gradient tracking controller drives it to move along the -∇Φ direction every T_ctrl cycle (e.g., 20ms). When calculating the dynamic field coupling function R_i, the AGV velocity vector calculated in the previous control cycle is used to ensure the solvability of the field dynamics equations. In this mode, AGVs automatically flow towards regions with high task potential; at narrow passage intersections, AGVs with heavier loads naturally gain priority due to their repulsive field strength; the system as a whole exhibits a self-organized, smooth traffic flow.

[0038] Performance Verification: Benchmark tests show that in a warehousing scenario with 100 AGVs, compared to the traditional market auction algorithm: the average task completion delay decreased from 180±32 seconds to 72±15 seconds (a 60% reduction); the collision frequency within the channel decreased from 5.1 times / minute to 0.8 times / minute (an 84% reduction); and the peak CPU utilization of the central node decreased from 89% to 67% (a 25% reduction). The field dynamics optimization module uses the Alternating Direction Multiplier Method (ADMM) for distributed solution. Actual test data shows that when the number of collaborative AGVs increases from 10 to 100: the computation time for a single solution by the central optimization module increases by 280% (compared to a 900% increase in the traditional centralized MPC solution); the increase in downlink communication bandwidth from the center to individual machines caused by each additional AGV remains stable within 0.3KB / s.

[0039] Example 2: Dynamic Area Coverage by UAV Swarm like Figure 6 As shown, the drone swarm needs to continuously monitor an irregular area.

[0040] Unified field construction: Define a coverage field C(r, t), whose value represents the urgency of spatial location r needing to be covered. The initial coverage of the area is uniform, and the coverage value decreases after UAV detection, and slowly recovers over time (simulating information aging).

[0041] Field dynamics model: ∂C / ∂t = D_c∇²C - Σ_i S_i(r, s_i). Where S_i is the "coverage reduction" function for UAV i, where the coverage value of the area flown by the UAV decreases significantly. The diffusion term D_c∇²C simulates the outward propagation of low-coverage information from already covered areas.

[0042] Control: Each drone flies along the coverage field gradient (from low coverage to high coverage) to achieve autonomous swarm diffusion and exploration of uncovered areas.

[0043] Results: This scheme achieves autonomous coverage without partitioning or fixed routes, and its coverage efficiency and balance are significantly better than traditional partitioned coverage or random walk strategies.

[0044] Example 3: System Robustness Implementation and Security Enhancement This embodiment describes the enhanced characteristics of the system under non-ideal conditions.

[0045] Communication degradation and security circuit breakers: such as Figure 4 As shown, when the robot cannot receive central field updates due to communication interruption, its gradient tracking controller initiates a degraded mode. The controller performs local calculations based on the last received field information and integrates a temporary local risk field generated in real time by the onboard LiDAR and visual sensors. Crucially, the controller incorporates a built-in safety circuit breaker: when the controller detects an irreconcilable conflict between the suggested motion direction calculated based on the field gradient and the obstacle information perceived in real time by the onboard sensors, it determines it as a 'field-perception mismatch'. Specific conflict detection uses vector angle calculation: the controller calculates in real time the angle θ between the local field gradient direction vector ∇Φ and the sensor obstacle avoidance direction vector V_obs = arccos( [∇Φ · V_obs] / (|∇Φ| |V_obs|) ). When θ > 30° or the estimated time to collision (TTC) calculated based on sensor data is less than 0.5 seconds, a safety circuit breaker is triggered. At this time, the controller immediately switches to the real-time obstacle avoidance mode based on sensor data to ensure absolute safety. After the conflict is resolved or communication is restored, the system automatically switches back to the field gradient tracking mode. In simulation tests that included 20% random packet loss, the system still maintained a task completion rate of over 90%.

[0046] Online Parameter Learning and Adaptation: The system includes a field model parameter learning unit that runs as a background process. This unit updates the parameters by minimizing the loss function L. A specific example of the loss function is: L = 0.7 × average task completion time + 0.3 × ||Φ_pred - Φ_actual||², where Φ_pred is the predicted value of the field model, and Φ_actual is the actual field state. This unit uses the Adam optimizer for gradient descent to fine-tune key parameters in the field dynamics model online, such as the diffusion coefficient D and the fusion weights of each field layer (w_c, w_t, w_d). The parameter learning unit continuously analyzes system runtime data to optimize the field model, paying particular attention to the correlation between safety events (such as the frequency of circuit breaker triggering and the amplitude of obstacle avoidance actions) and field prediction errors. The updated parameters are provided to the field dynamics optimization module for prediction optimization in the next cycle, thus forming a complete adaptive closed loop of "execution-evaluation-learning-optimization".

[0047] Field information compression: In step S3, to reduce communication bandwidth, the field information distributed to the robot is lossy compressed. The compression algorithm must ensure that the decompressed field information satisfies the following: the positional error of local extrema (such as the center of the task potential well) is less than 5 cm, the direction error of the field gradient on the robot's planned path is less than 10 degrees, and the topological structure of the repulsive field contour lines around key obstacles remains unchanged.

[0048] This invention introduces a unified field Φ(r,t) as the unique continuous representation of the environment, task, and multi-robot interaction, utilizes the inherent laws of field dynamics for prediction and optimization, and simplifies control to tracking the local field gradient. It fundamentally breaks through the inherent bottleneck of traditional cooperative control architecture and provides an efficient, robust, and scalable solution for robot collaboration in large-scale, highly dynamic, and highly uncertain scenarios.

Claims

1. A multi-robot collaborative control system based on unified field dynamic coupling, characterized in that, include: The field modeling module is configured to construct and update in real time a spatiotemporal unified field Φ(r, t) that contains information about the coding environment and the task objective; The field dynamics optimization module is connected to the field modeling module and is configured to predict the unified field based on a preset field dynamics model, and optimize the field evolution trajectory by dynamically adjusting the field coupling parameters of each robot. A distributed field interaction interface is deployed on each robot to acquire local field information and gradients and receive coupling parameter commands. A gradient tracking controller, integrated into each robot, is used to drive robot motion based on the local field gradient.

2. The system according to claim 1, characterized in that, The field dynamics model is as follows: ∂Φ(r, t) / ∂t = D∇²Φ(r, t) + f_task(Φ, r, t) + Σ_i g_i(Φ, s_i, u_i(t-Δt), θ_i) Where r is the spatial coordinate, s_i is the real-time state vector of the i-th robot, u_i(t-Δt) is the control input generated in the previous control cycle, g_i is the dynamic field coupling function of the i-th robot, and θ_i is a coupling parameter that can be optimized online. The coupling parameter θ_i is at least used to adjust the reaction strength or spatial distribution characteristics of the robot to the unified field, as characterized by the function g_i.

3. The system according to claim 2, characterized in that, The output intensity or spatial distribution of the dynamic field coupling function g_i is associated with the real-time physical state s_i of the robot.

4. The system according to claim 3, characterized in that, The field generated by the dynamic field coupling function g_i is asymmetrically distributed in space, with the field strength in the direction of the robot's current motion velocity vector being greater than that in the opposite direction.

5. The system according to claim 2, characterized in that, The field dynamics optimization module adopts a model predictive control framework, with the goal of minimizing the task cost function for the future field sequence, and continuously optimizes the coupling parameter {θ_i}; and the optimized coupling parameter θ_i(t_{k+1}) remains constant within the time interval [t_k, t_{k+1}), which is used by the gradient tracking controller within this interval.

6. The system according to claim 1, characterized in that, The spatiotemporal unified field constructed by the field modeling module is a multi-layer fusion field, which includes at least: a passage cost layer based on geometric perception, a task-related layer based on semantic understanding, and a spatiotemporal risk layer that encodes dynamic obstacle prediction.

7. The system according to claim 2, characterized in that, It also includes a field parameter learning unit, configured to adjust the parameters in the field dynamics model online based on system operating data.

8. The system according to claim 1, characterized in that, The gradient tracking controller is equipped with a safety circuit breaker mechanism: when the deviation between the motion direction calculated based on the local field gradient and the obstacle avoidance direction calculated based on the airborne real-time sensor data exceeds a preset safety threshold, it switches to the obstacle avoidance mode based on sensor data. The preset safety thresholds include: the angle between the direction vectors of the two objects is greater than 30 degrees, or the estimated collision time calculated based on the relative speed and position is less than 0.5 seconds.

9. A multi-robot cooperative control method based on unified field dynamic coupling, characterized in that, include: S1: Construct and initialize a unified field that integrates environment and task information; S2: Field Prediction and Cooperative Optimization: Based on the field state Φ(t_k) at the current time t_k, only the effects of the task source term and the diffusion term are considered in the prediction time domain, ignoring the influence of the robot's dynamic coupling term. The predicted trajectory is made to tend towards the task target by optimizing the robot's field coupling parameter {θ_i}, and the optimized parameters {θ_i(t_{k+1})} at the next time t_{k+1} are obtained. S3: Distribute field information and optimized coupling parameters {θ_i(t_{k+1})} to each robot; S4: Local field reaction and gradient tracking: In the time interval [t_k, t_{k+1}), each robot uses the parameter θ_i(t_{k+1}) in each control cycle. First, based on this parameter and the control input of the previous cycle, it influences the local field through the field coupling function, and then tracks the updated local field gradient to move. S5: Update the field based on motion and perception results, and execute in a closed loop.

10. The method according to claim 9, characterized in that, The optimization in step S2 is performed using a distributed or parallel computing architecture.

11. The method according to claim 9, characterized in that, In step S3, the field information distributed to the robot is lossy compressed. The compression algorithm used must ensure that the decompressed field information satisfies the following: the position error of the local extreme point is less than 5 cm, the direction error of the field gradient on the robot's planned path is less than 10 degrees, and the topological structure of the repulsive field contour lines around the key obstacles remains unchanged.

12. A robot swarm, characterized in that, It comprises multiple robots, each equipped with a distributed field interaction interface and gradient tracking controller as described in any one of claims 1-8.

Images