Distributed consensus-driven embodied intelligent multi-machine conflict resolution method

Abstract

The application relates to a distributed consensus-driven embodied intelligent multi-machine conflict resolution method, in particular to the technical field of cooperative decision and control, which predicts network situation through a hidden Markov model, so that a robot can perceive link quality in advance; a minimum braking safety distance formula is used to calculate physical conflict urgency and the physical conflict urgency is used as a consensus priority scalar to perform pruning, so that invalid negotiation is avoided; a long-term and short-term memory and a self-attention combined prediction model is used to realize speculative pre-consensus, prediction is used to replace real-time interaction when communication is limited, and delay is greatly reduced; a Markov decision process and a deep Q network are used to adaptively select an optimal consensus strategy, and a future time window length, a positioning error margin and a prediction confidence determination threshold are reversely modulated to form a closed-loop optimization; finally, agreement can be quickly reached in a clock drift and high packet loss environment, and the risks of deadlock and collision are eliminated, so that the system can continuously and stably operate.

Description

Technical Field

[0001] This invention relates to the field of collaborative decision-making and control technology, and more specifically, to a distributed consensus-driven method for resolving conflicts among embodied intelligent multi-machine systems. Background Technology

[0002] With the large-scale deployment of industrial scenarios such as smart warehousing, unmanned mining areas, and automated ports, hundreds of embodied intelligent robots form a distributed collaborative operation system via wireless networks. These robots are typically equipped with edge computing units and rely on WiFi, 5G, or Mesh networks for real-time communication and coordination. In actual operating environments, dense metal shelves, moving equipment, and complex building structures severely obstruct wireless signals, leading to frequent communication link interruptions, end-to-end transmission latency fluctuations exceeding 120 milliseconds, and packet loss rates exceeding 10%. Meanwhile, in indoor or underground environments where GPS is denied, each robot relies on a local crystal oscillator to maintain its clock, resulting in significant clock drift after prolonged operation, causing inaccurate message sorting based on timestamps. Under these adverse conditions, multiple robots need to make real-time decisions on critical matters such as right-of-way allocation, task assignment, and resource utilization based on local observation information and asynchronously arriving messages. When multiple robots simultaneously make conflicting decisions regarding the same physical space or the same task based on expired state information, the system faces severe coordination dilemmas.

[0003] Existing distributed consensus algorithms such as Paxos, Raft, and PBFT assume basic network reliability, controllable communication latency, and synchronized node clocks in their initial design, ensuring strong consistency in ideal data center environments. However, in wide-area deployment scenarios of embodied intelligent multi-machine systems, the multiple communication rounds of these algorithms result in decision delays far exceeding the sub-second response window required for robot motion control. While vector clocks and Lamport logical clocks can partially alleviate clock synchronization issues, in scenarios with dynamically changing network topologies, the scale of metadata they maintain grows quadratically with the number of nodes, rapidly exhausting the limited storage and computing resources of edge devices in large-scale clusters. Natural language negotiation schemes based on large language models, while offering advantages in flexibility and expressiveness, are prone to semantic ambiguity in asynchronous communication and incomplete information environments, leading to different robots interpreting the same negotiation result in opposite ways, thus causing physical collisions. In summary, existing technologies, under harsh conditions of limited communication and clock drift, cannot reach consensus on key decisions such as right-of-way, task assignment, and resource allocation within 500 milliseconds. Therefore, a lightweight, novel asynchronous consensus protocol oriented towards embodied physical constraints is urgently needed. Summary of the Invention

[0004] This invention addresses the technical problems existing in the prior art by providing a distributed consensus-driven method for resolving conflicts among embodied intelligent multi-machine systems, thereby solving the problems mentioned in the background section.

[0005] The technical solution of this invention to solve the above-mentioned technical problems is as follows: specifically, it includes the following steps:

[0006] Step S1: Each robot constructs a hidden Markov model based on historical communication logs to predict the latency and packet loss rate of each link within a preset future time window, and outputs a network situation vector.

[0007] Step S2: Each robot calculates the physical conflict urgency with neighboring robots based on the network situation vector and its own kinematic parameters, including current speed, maximum braking deceleration and positioning error margin, using the minimum braking safety distance formula. This physical conflict urgency is used as a consensus priority scalar. Physical constraint pruning is performed on the received conflict proposals, and the pruning decision cache is output. The conflict proposal is a right-of-way occupancy request initiated by a neighboring robot.

[0008] Step S3: Each robot uses a combined prediction model of long short-term memory and self-attention to predict the future state of its neighboring robots. Based on the preset prediction confidence threshold and the urgency of physical conflict, it decides whether to replace real-time communication with speculative pre-consensus and outputs the prediction confidence, speculative execution hit rate and prediction error. The prediction confidence is automatically calculated by the combined prediction model based on historical prediction deviations.

[0009] Step S4: Each robot takes the network situation vector, prediction confidence, inferred execution hit rate, prediction error, and physical conflict urgency as state input Markov decision process, selects the optimal consensus strategy through deep Q network, and modulates the future time window length in step S1, the positioning error margin in the minimum braking safety distance formula in step S2, and the prediction confidence judgment threshold in step S3 according to the selected optimal consensus strategy to achieve adaptive switching of consensus strategy.

[0010] In a preferred embodiment, the specific operations for each robot to construct the hidden Markov model in step S1 are as follows:

[0011] Define a set of hidden states, and use the generalized Pareto distribution of the associated time delay and the log-normal distribution of the packet loss rate as the observation probability density function for each hidden state;

[0012] Each robot maintains a state transition count matrix. After each hidden state inference is completed, the state transition count matrix is ​​updated according to the maximum a posteriori hidden state sequence obtained by the Viterbi algorithm, and the estimated value of the transition matrix is ​​calculated by the Dirichlet a posteriori expectation formula.

[0013] Among them, the generalized Pareto distribution is used to characterize the heavy-tailed feature of network latency, the log-normal distribution is used to describe the skewed distribution of packet loss rate in the interval between zero and one, and the update of the transition matrix adopts the Dirichlet-multinomial conjugate model to realize sequential incremental learning without storing complete historical data.

[0014] In a preferred embodiment, in step S1, the specific operation of each robot predicting the latency and packet loss rate of each link within the future time window is as follows:

[0015] At the beginning of each preset decision cycle, the posterior distribution of the hidden state at the current time is calculated using the forward algorithm;

[0016] Subsequently, the matrix exponential approximation method based on the transition matrix calculates the distribution of the hidden state at the end of the future time window based on the posterior distribution of the current hidden state, where the matrix exponent is calculated quickly using Pad approximation.

[0017] Then, based on the hidden state distribution at the end of the future time window, the mean predicted delay and the mean predicted packet loss rate of each link are calculated respectively, and the Shannon entropy of the hidden state distribution at the end of the future time window is calculated as a measure of prediction uncertainty.

[0018] Finally, the average predicted latency, average predicted packet loss rate, and Shannon entropy of all neighboring links are concatenated into a vector to form the network situation vector.

[0019] In a preferred embodiment, the specific operation of determining the minimum braking safety distance between each robot and its neighboring robots in step S2 is as follows:

[0020] First, the mean predicted delay and mean predicted packet loss rate of each link are extracted from the network situation vector. Then, the positioning error margin is calculated using the exponential saturation function. The exponential saturation function takes the mean predicted delay as the independent variable and the increment of the positioning error margin as the dependent variable, and adds a linear term of the mean predicted packet loss rate to obtain the final positioning error margin.

[0021] Then, the minimum braking safety distance is calculated by summing the square of its current speed divided by twice its maximum braking deceleration, the square of the neighbor's current speed divided by twice the neighbor's maximum braking deceleration, and the positioning error margin.

[0022] In a preferred embodiment, the specific operation of each robot generating the physical conflict urgency and consensus priority scalar in step S2 is as follows:

[0023] First, obtain the current Euclidean distance between yourself and each neighboring robot, and compare it with the minimum braking safety distance. If the current Euclidean distance is less than the minimum braking safety distance, the physical conflict indication value is set to 1; otherwise, it is set to 0.

[0024] Then, the physical conflict indication value corresponding to each neighbor is multiplied by one, plus the product of a preset uncertainty amplification factor and the Shannon entropy of that link in the network situation vector, and then summed over all neighbors to obtain the physical conflict urgency.

[0025] Meanwhile, the consensus priority scalar is obtained by multiplying the preset task priority, the inverse of the preset path cost, and the inverse of the average predicted latency calculated from the average predicted latency of all neighbor links in the network situation vector by preset weight coefficients and summing them.

[0026] Finally, the robot decides on the received conflicting proposals based on the consensus priority scalar. If the consensus priority scalar of this robot is higher than the consensus priority scalar of the proposal initiator, the proposal is rejected; otherwise, it is allowed to pass, and the decision is stored in the pruning decision cache.

[0027] In a preferred embodiment, the specific operation of each robot constructing the neighbor historical state sequence in step S3 is as follows:

[0028] Each robot maintains a fixed-length sliding window to store the state packets received from each neighboring robot in the most recent decision cycles. The state packets include position coordinates, velocity vector, task priority, and consensus priority scalar. This information is then concatenated into a feature vector to obtain a sequence of feature vectors.

[0029] Subsequently, a long short-term memory network is used to temporally encode the feature vector sequence of each neighbor to obtain the hidden state of each neighbor;

[0030] Then, a multi-head self-attention mechanism is introduced to interact and fuse the hidden states of all neighbors, and output the fused feature vector of each neighbor.

[0031] In a preferred embodiment, the specific operation of each robot predicting the future state of its neighbors and calculating the prediction confidence in step S3 is as follows:

[0032] The hidden state of each neighbor is concatenated with the fused feature vector and input into the mixing density network. The network outputs the mixing coefficients, mean, and covariance diagonal elements of multiple Gaussian components to obtain the probability distribution of the future state. The weighted sum of the means of each Gaussian component is used as the predicted future state.

[0033] Meanwhile, prediction confidence is calculated based on the exponentially weighted moving average of historical prediction bias and the negative exponent of the differential entropy of the probability distribution of future states. The historical prediction bias is measured by Huber loss, and the negative exponent of the differential entropy serves as a proxy for the degree of distribution concentration.

[0034] In a preferred embodiment, the specific operation of each robot deciding whether to enable speculative pre-consensus in step S3 is as follows:

[0035] The global prediction confidence, obtained by averaging the prediction confidence of all neighbors, is compared with a preset prediction confidence threshold. At the same time, the physical conflict urgency is compared with a preset physical conflict urgency threshold.

[0036] If the global prediction confidence is not lower than the judgment threshold and the physical conflict urgency is lower than the physical conflict urgency threshold, then speculative pre-consensus is enabled, and the predicted future state is used to replace real-time communication to participate in conflict resolution; otherwise, real-time communication mode is maintained.

[0037] At the end of each decision cycle, the prediction execution hit rate and prediction error are updated. The prediction execution hit rate is the proportion of successful predictions in the past decision cycles to the total number of uses. The prediction error is measured by the symmetric mean absolute percentage error.

[0038] In a preferred embodiment, the specific operations for each robot to construct the state space and reward function of the Markov decision process in step S4 are as follows:

[0039] First, the network situation vector output in step S1 is aggregated and dimensionality reduced. The average value of the predicted delay of all neighbor links, the average value of the predicted packet loss rate, the average value of Shannon entropy, the maximum value of the predicted delay, and the maximum value of the predicted packet loss rate are extracted, for a total of five aggregated statistics. These are then combined with the prediction confidence, inference execution hit rate, prediction error, and physical conflict urgency output in step S3 to form a nine-dimensional state vector.

[0040] Meanwhile, a reward function is designed, in which the communication overhead penalty term is assigned a normalized value according to the selected optimal consensus strategy, the delay penalty term is assigned a normalized value according to the selected optimal consensus strategy, the collision risk term maps the physical conflict urgency to the zero-to-one interval through the logistic function, and the task delay term fuses the average prediction delay and prediction error to the zero-to-one interval through the hyperbolic tangent function. The four penalties are multiplied by dynamic weights respectively, and the sum is taken as the negative value as the reward value.

[0041] Each robot inputs its nine-dimensional state vector into a deep Q-network, outputs the Q-values ​​of four actions, and uses a greedy strategy with decreasing exploration rate to select the action that maximizes the Q-value as the optimal consensus strategy.

[0042] In a preferred embodiment, the specific operations of each robot performing inverse modulation and adaptive weight adjustment in step S4 are as follows:

[0043] Based on the selected optimal consensus strategy and the current state, fuzzy control rules are used to adjust the future time window length in step S1, the positioning error margin in step S2, and the prediction confidence threshold in step S3, respectively. The adjustment of the future time window length is based on the selected optimal consensus strategy and the average prediction delay, the adjustment of the positioning error margin is based on the urgency of physical conflicts, and the adjustment of the prediction confidence threshold is based on the inferred execution hit rate and the prediction error.

[0044] Meanwhile, a sliding window is maintained to statistically analyze the actual average values ​​of recent communication overhead penalties, delay penalties, collision risk items, and task delay items. The four weight coefficients in the reward function are dynamically updated exponentially based on the deviation from the preset target value, and then normalized so that their sum is one.

[0045] The beneficial effects of this invention are as follows: It predicts network conditions using a Hidden Markov Model, enabling the robot to perceive link quality in advance; it calculates the urgency of physical conflicts using the minimum braking safety distance formula and uses it as a consensus priority scalar for pruning, avoiding ineffective negotiation; it achieves speculative pre-consensus by leveraging a prediction model combining Long Short-Term Memory and Self-Attention, replacing real-time interaction with prediction when communication is limited, significantly reducing latency; then, the Markov Decision Process and Deep Q-Network adaptively select the optimal consensus strategy, and inversely modulate the future time window length, positioning error margin, and prediction confidence threshold to form a closed-loop optimization; finally, it can quickly reach consensus even in clock drift and high packet loss environments, eliminating deadlock and collision risks and ensuring the continuous and stable operation of the system. Attached Figure Description

[0046] Figure 1 This is a flowchart of the method of the present invention. Detailed Implementation

[0047] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0048] In the description of this application, the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Thus, a feature defined as "first" or "second" may explicitly or implicitly include one or more of the stated features. In the description of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0049] In the description of this application, the term "for example" is used to mean "used as an example, illustration, or description." Any embodiment described as "for example" in this application is not necessarily to be construed as being more preferred or advantageous than other embodiments. The following description is provided to enable any person skilled in the art to make and use the invention. Details are set forth in the following description for purposes of explanation. It should be understood that those skilled in the art will recognize that the invention can be made without using these specific details. In other instances, well-known structures and processes will not be described in detail to avoid obscuring the description of the invention with unnecessary detail. Therefore, the invention is not intended to be limited to the embodiments shown, but is consistent with the broadest scope of the principles and features disclosed in this application.

[0050] Example 1: This example provides the following... Figure 1 The method for resolving conflicts among embodied intelligence multi-machines driven by distributed consensus includes the following steps:

[0051] Step S1: Each robot constructs a hidden Markov model based on historical communication logs to predict the latency and packet loss rate of each link within a preset future time window, and outputs a network situation vector.

[0052] Step S2: Each robot calculates the physical conflict urgency with neighboring robots based on the network situation vector and its own kinematic parameters, including current speed, maximum braking deceleration and positioning error margin, using the minimum braking safety distance formula. This physical conflict urgency is used as a consensus priority scalar. Physical constraint pruning is performed on the received conflict proposals, and the pruning decision cache is output. The conflict proposal is a right-of-way or resource occupation request initiated by a neighboring robot.

[0053] Step S3: Each robot uses a combined prediction model of long short-term memory and self-attention to predict the future state of its neighboring robots. Based on the preset prediction confidence threshold and the urgency of physical conflict, it decides whether to replace real-time communication with speculative pre-consensus and outputs the prediction confidence, speculative execution hit rate and prediction error. The prediction confidence is automatically calculated by the combined prediction model based on historical prediction deviations.

[0054] Step S4: Each robot takes the network situation vector, prediction confidence, inferred execution hit rate, prediction error, and physical conflict urgency as state input Markov decision process, selects the optimal consensus strategy through deep Q network, and modulates the future time window length in step S1, the positioning error margin in the minimum braking safety distance formula in step S2, and the prediction confidence judgment threshold in step S3 according to the selected optimal consensus strategy to achieve adaptive switching of consensus strategy.

[0055] In this embodiment, it is specifically necessary to explain the specific operations of each robot in constructing the Hidden Markov Model in step S1:

[0056] A set of hidden states is defined, with the generalized Pareto distribution of latency and the log-normal distribution of packet loss rate as the observation probability density function for each hidden state. The set of hidden states contains three hidden states, corresponding to three modes: excellent, moderate, and poor network quality, respectively. The shape parameter of the generalized Pareto distribution is initially set to 0.3, the scale parameter is initially set to 5 milliseconds, and the position parameter is initially set to 10 milliseconds. The log-mean of the log-normal distribution is initially set to -2, and the log-scale parameter is initially set to 0.5. These initial parameters are pre-trained using maximum likelihood estimation from 1 million historical communication logs collected offline and fine-tuned during operation through online Bayesian updates.

[0057] Each robot maintains a state transition counting matrix. After each hidden state inference, the state transition counting matrix is ​​updated according to the maximum a posteriori hidden state sequence obtained by the Viterbi algorithm, and the estimated value of the transition matrix is ​​calculated using the Dirichlet a posteriori expectation formula. The specific calculation process of the Viterbi algorithm is as follows: First, the path probabilities are initialized by setting the forward probabilities of each hidden state at the first time step to the initial distribution multiplied by the observation probability; then, the maximum probability path of each hidden state at each time step is calculated recursively, and the backtracking pointer is recorded; finally, the hidden state with the highest probability is selected from the last time step, and the complete optimal hidden state sequence is obtained by backtracking. The size of the transition counting matrix is ​​3x3, and the initial value is all 0. The Dirichlet prior hyperparameter is uniformly set to 1, i.e., uniform prior.

[0058] The estimated value of the transition matrix is ​​calculated as follows: each transition probability is equal to the corresponding count plus 1 divided by the sum of all counts in that row plus 3. This update method does not require storing all historical data; it only needs to maintain the counting matrix to achieve sequential incremental learning.

[0059] Among them, the generalized Pareto distribution is used to characterize the heavy-tailed feature of network latency, the log-normal distribution is used to describe the skewed distribution of packet loss rate in the interval of zero to one, and the update of the transition matrix adopts the Dirichlet-multinomial conjugate model to realize sequential incremental learning without storing complete historical data. The advantage of using the Dirichlet-multinomial conjugate model is that it can update the transition probability online in real time without retraining the entire model, which is particularly suitable for scenarios where edge robots have limited computing power and the network state changes slowly over time.

[0060] In step S1, the specific operations for each robot to predict the latency and packet loss rate of each link within the future time window are as follows:

[0061] At the start of each preset decision cycle, the forward algorithm is used to calculate the posterior distribution of the hidden state at the current time. The preset decision cycle is 200 milliseconds, synchronized with the robot's motion control cycle. The specific calculation process of the forward algorithm is as follows: First, the posterior distribution of the hidden state at time 1 is initialized by multiplying the initial distribution by the observation probability and normalizing it. Second, for each subsequent time, the prior distribution at the current time is predicted based on the posterior distribution of the previous time and the transition matrix, then multiplied by the current observation probability to obtain the unnormalized posterior, and finally normalized to obtain the posterior distribution of the hidden state at the current time. The time complexity of this algorithm is proportional to the square of the number of hidden states, and the computational cost is extremely low when the number of hidden states is 3.

[0062] Subsequently, a matrix exponential approximation method based on the transition matrix is ​​used to calculate the hidden state distribution at the end of the future time window based on the posterior distribution of the current hidden state. The matrix exponent is calculated quickly using the Pad approximation. The initial length of the future time window is set to 1000 milliseconds, which can be modulated in reverse by step S4. The specific calculation process of the matrix exponential approximation is as follows: First, the logarithmic matrix of the transition matrix is ​​calculated using the diagonal Pad approximation in the Pad approximation, with the order set to 2, i.e., using rational functions where both the numerator and denominator are quadratic polynomials to approximate the exponential function. Then, the logarithmic matrix is ​​multiplied by the ratio of the future time window length to the decision period, i.e., 5 times, and the exponent is calculated to obtain the 5th power of the transition matrix. Finally, the posterior distribution of the current hidden state is multiplied by the power of this transition matrix to obtain the hidden state distribution at the end of the future time window. Compared with Taylor series, the Pad approximation converges faster and has better numerical stability, making it suitable for real-time calculation on edge devices.

[0063] Then, based on the distribution of hidden states at the end of the future time window, the mean predicted delay and the mean predicted packet loss rate of each link are calculated, and the Shannon entropy of the distribution of hidden states at the end of the future time window is calculated as a measure of prediction uncertainty. The calculation process of the mean predicted delay is as follows: for each hidden state, the mean of its generalized Pareto distribution is taken, that is, the position parameter plus the scale parameter divided by 1 minus the shape parameter; then the mean is multiplied by the probability of the hidden state at a future time, and the sum is obtained for all hidden states to obtain the mean predicted delay.

[0064] The process of calculating the predicted average packet loss rate is as follows: For each hidden state, take the mean of its log-normal distribution, which is the logarithmic mean with the natural constant e as the base and the exponent as the logarithm, plus half the square of the logarithmic scale parameter; then multiply the mean by the probability of the hidden state at a future time, and sum over all hidden states to obtain the predicted average packet loss rate.

[0065] The Shannon entropy is calculated as follows: take the natural logarithm of the probability of each hidden state, multiply it by the probability itself, take the negative value, and finally sum over all hidden states; the value of Shannon entropy ranges from 0 to 1.099, and the larger the entropy value, the higher the uncertainty of the prediction; when the Shannon entropy is greater than 0.8, the prediction result is considered unreliable, and step S4 will tend to choose a more conservative consensus strategy.

[0066] Finally, the average predicted latency, average predicted packet loss rate, and Shannon entropy of all neighbor links are concatenated into a vector to form the network situation vector. Assuming that a robot has 5 neighbors, the dimension of the network situation vector is 15, meaning that each neighbor contributes 3 values. The concatenation order is arranged in ascending order according to the neighbor robot numbers to ensure that the vector structure within each robot is consistent, making it easy to directly index and use in subsequent steps.

[0067] In this embodiment, it is specifically necessary to explain the operation of the minimum braking safety distance between each robot and its neighboring robots in step S2 as follows:

[0068] First, the average predicted delay and average predicted packet loss rate for each link are extracted from the network situation vector. Then, the positioning error margin is calculated using an exponential saturation function. This function uses the average predicted delay as the independent variable and the increment of the positioning error margin as the dependent variable, while also adding a linear term of the average predicted packet loss rate to obtain the final positioning error margin. The specific form of the exponential saturation function is: the positioning error margin equals the basic positioning error margin plus the delay impact amplitude coefficient multiplied by one, minus the exponential function value of the delay sensitivity coefficient (base e, negative exponent) multiplied by the average predicted delay, plus the packet loss impact coefficient multiplied by the average predicted packet loss rate. The basic positioning error margin is set to 0.1 meters, the delay impact coefficient is set to 0.3 meters, the delay sensitivity coefficient is set to the negative first power of 0.02 milliseconds, and the packet loss impact coefficient is set to 0.5 meters. This function can simulate the nonlinear decay effect of communication delay on positioning accuracy: when the delay is small, the error margin grows rapidly; when the delay exceeds a certain threshold, the growth tends to saturate, avoiding infinite amplification; the linear term predicting the mean packet loss rate reflects the additional uncertainty caused by packet loss; this formula does not use ordinary statistics such as standard deviation or variance, but achieves a nonlinear mapping with clear physical meaning through an exponential saturation function.

[0069] Then, the minimum braking safety distance is calculated by summing the square of the current speed divided by twice the current speed of the robot, the square of the current speed of the neighbor divided by twice the maximum braking deceleration of the neighbor, and the positioning error margin. The formula for calculating the minimum braking safety distance is: the square of the current speed of the robot divided by twice the current speed of the robot, plus the square of the current speed of the neighbor divided by twice the maximum braking deceleration of the neighbor, plus the positioning error margin. The units of the current speed of the robot and the current speed of the neighbor are meters per second, which are obtained in real time by the odometer or inertial measurement unit. The units of the maximum braking deceleration of the robot and the maximum braking deceleration of the neighbor are meters per second squared, which are determined by the robot hardware specifications and are usually between 2 and 5 meters per second squared. In this embodiment, 3 meters per second squared is uniformly used. The positioning error margin is calculated by the aforementioned exponential saturation function and is in meters. This distance ensures that no collision will occur when both parties brake simultaneously at maximum deceleration, and at the same time, the dynamic positioning error margin compensates for the impact of communication quality on positioning accuracy.

[0070] In step S2, the specific operations for each robot to generate physical conflict urgency and consensus priority scalars are as follows:

[0071] First, the current Euclidean distance between itself and each neighboring robot is obtained and compared with the minimum braking safety distance. If the current Euclidean distance is less than the minimum braking safety distance, the physical conflict indication value is set to 1; otherwise, it is set to 0. The current Euclidean distance is measured in real time by LiDAR or a visual positioning system, and the unit is meters. The comparison process is executed once in each decision cycle, which is 200 milliseconds. The physical conflict indication value is a Boolean variable, where 1 indicates that there is a collision risk and 0 indicates that it is safe. This indication value is the basis for subsequent calculation of the urgency of physical conflict.

[0072] Then, the physical conflict indication value corresponding to each neighbor is multiplied by 1, plus a preset uncertainty amplification factor, and the product of the Shannon entropy of that link in the network situation vector. This product is then summed over all neighbors to obtain the physical conflict urgency. The formula for calculating the physical conflict urgency is: for all neighbors, multiply the physical conflict indication value by 1, plus a preset uncertainty amplification factor, and then sum the products. The preset uncertainty amplification factor is set to 2.0, and the Shannon entropy ranges from 0 to 1.099, calculated in step S1. This formula means that when the Shannon entropy is large, even if the physical conflict indication value is 0 (current distance is greater than the safe distance), due to high prediction uncertainty, the neighbor still contributes a certain urgency weight, thus enabling the system to adopt a more conservative strategy when network quality is poor. The physical conflict urgency is a non-negative real number; the larger the value, the higher the overall collision risk faced by the robot.

[0073] Simultaneously, the consensus priority scalar is obtained by multiplying the preset task priority, the reciprocal of the preset path cost, and the reciprocal of the average predicted latency calculated from the average predicted latency of all neighbor links in the network situation vector by preset weight coefficients, and then summing them. The calculation formula for the consensus priority scalar is: preset task priority multiplied by the first weight coefficient, plus the reciprocal of the preset path cost multiplied by the second weight coefficient, plus the reciprocal of the average predicted latency multiplied by the third weight coefficient. The preset task priority is a positive integer assigned by the upper-layer task management system; a larger value indicates a more urgent task. In this embodiment, the value ranges from 1 to 1. 0; The preset path cost is the estimated remaining length of the robot's current path, in meters, provided by the path planning module; the average predicted latency is the arithmetic mean of the predicted latency of all neighboring links, in milliseconds; the first weight coefficient is set to 0.5, the second weight coefficient is set to 0.3, and the third weight coefficient is set to 0.2, with the sum of the three being 1; to prevent division by zero, a very small positive number 0.001 is added to the reciprocal calculation of the path cost and the average predicted latency; this scalar comprehensively considers the importance of the task, the urgency of the task, and the communication quality, so that robots with poor network conditions or urgent tasks have a higher bargaining power;

[0074] Finally, the received conflict proposals are adjudicated based on the consensus priority scalar. If the consensus priority scalar of this robot is higher than that of the proposal initiator, the proposal is rejected; otherwise, it is allowed to pass, and the judgment result is stored in the pruning judgment cache. The conflict proposal is initiated by the neighboring robot and includes information such as the initiator's identifier, the requested resource or right-of-way identifier, and the proposed occupation time period. This robot compares its own consensus priority scalar with that of the proposal initiator, the latter being carried in the proposal message or obtained from the local cache. If the robot's priority is higher, the proposal is rejected, that is, the neighboring robot is not allowed to occupy the requested resource or right-of-way; otherwise, it is allowed to pass. The judgment result is stored in the pruning judgment cache in the form of a record. Each record includes the proposal identifier, the initiator's identifier, the judgment result (pass or reject), and the judgment timestamp. This cache will be used in step S3 to determine whether speculative pre-consensus needs to be initiated: if a neighbor's proposal is frequently rejected, it indicates that the conflict in the area is intense, and it is necessary to strengthen the prediction of the neighbor's behavior.

[0075] In this embodiment, it is specifically necessary to explain the specific operation of each robot constructing the neighbor historical state sequence in step S3:

[0076] Each robot maintains a fixed-length sliding window, storing state packets received from each neighboring robot within the most recent decision cycles. Each state packet includes position coordinates, velocity vector, task priority, and a consensus priority scalar. This information is concatenated into a feature vector sequence. The sliding window length is set to 20 decision cycles, each lasting 200 milliseconds, thus covering the historical data of the past 4 seconds. The position coordinates in the state packet are two-dimensional planar coordinates in meters; the velocity vector is a two-dimensional vector in meters per second; the task priority is a positive integer ranging from 1 to 10, assigned by the upper-level task management system; and the consensus priority scalar is the scalar value calculated in step S2. These four... After concatenating the information, the feature vector has a dimension of 6, namely two-dimensional position, two-dimensional velocity, one-dimensional task priority, and one-dimensional consensus priority scalar; the sliding window size maintained by each robot is 20 multiplied by 6, that is, 120 values; the hidden layer dimension of the Long Short-Term Memory network is set to 64, and the number of network layers is 1. A single-layer unidirectional Long Short-Term Memory network is used to reduce the amount of computation; the number of heads in the multi-head self-attention mechanism is set to 4, and the key dimension of each head is 16, that is, 64 divided by 4; the learnable projection matrix of the self-attention mechanism is obtained through offline pre-training. The pre-training dataset is 1 million frames of multi-robot interaction trajectories collected in the simulation environment. The Adam optimizer is used during training, the learning rate is set to 0.001, and the number of training rounds is 50.

[0077] Subsequently, a Long Short-Term Memory (LSTM) network is used to temporally encode the feature vector sequence of each neighbor to obtain the hidden state of each neighbor. The input of the LSM network is the feature vector sequence of 20 time steps, and the output is the hidden state of the last time step, with a dimension of 64. This hidden state encodes the movement pattern and decision-making rules of the neighboring robot over the past 4 seconds. The activation functions of the forget gate, input gate, and output gate of the LSM network are all Sigmoid functions, and the activation function of the candidate memory unit is the hyperbolic tangent function. All neighbors share the same set of LSM network parameters to reduce the number of parameters and promote generalization.

[0078] Then, a multi-head self-attention mechanism is introduced to interact and fuse the hidden states of all neighbors, outputting a fused feature vector for each neighbor. The hidden states of all neighbors are stacked into a matrix with the number of rows equal to the number of neighbors and 64 columns. The multi-head self-attention mechanism projects this matrix onto four attention heads, each independently calculating the dot product attention of the query, key, and value. The outputs of the four heads are then concatenated and mapped back to 64 dimensions through a linear layer. Each row in the output matrix of the self-attention mechanism corresponds to a fused feature vector of a neighbor. This vector contains not only the neighbor's own motion information but also the potential influence of other neighbors on the neighbor, such as the avoidance behavior of other robots causing the neighbor to change its trajectory. The fused feature vector has a dimension of 64.

[0079] In step S3, the specific operations for each robot to predict the future state of its neighbors and calculate the prediction confidence are as follows:

[0080] The hidden state of each neighbor is concatenated with the fused feature vector and input into the hybrid density network. The network outputs the mixing coefficients, mean, and diagonal elements of the covariance of multiple Gaussian components to obtain the probability distribution of the future state. The weighted sum of the means of each Gaussian component is used as the predicted future state. The hidden state dimension is 64, and the fused feature vector dimension is 64, resulting in a 128-dimensional input vector. The hybrid density network contains two fully connected hidden layers: the first layer has 256 neurons, and the second layer has 128 neurons. The activation functions for both are linear rectified functions. The number of neurons in the output layer is calculated based on the number of Gaussian components and the feature vector dimension: the mixing coefficients output 5 values, the mean of each Gaussian component outputs 6 values ​​(same as the feature vector dimension), and the diagonal elements of the covariance of each Gaussian component output 6 values. Therefore, the total number of neurons in the output layer is 5 + 5 × 6 + 5 × 6, or 65. The mixing coefficients are normalized using the Softmax function to ensure a sum of 1; the diagonal elements of the covariance are ensured to be positive using an exponential function.

[0081] The training method for the hybrid density network is as follows: a large amount of real neighbor state transition data is collected in a simulation environment, and the Adam optimizer is used for training with negative log-likelihood as the loss function and a learning rate of 0.0005 for 100 training rounds. The predicted future state is the weighted sum of the means of each Gaussian component, that is, the mean of each component is multiplied by the mixing coefficient of that component and then summed to obtain a 6-dimensional vector containing the predicted position, velocity, task priority and consensus priority scalars.

[0082] Simultaneously, prediction confidence is calculated based on the exponentially weighted moving average of historical prediction bias and the negative exponent of the differential entropy of the probability distribution of future states. The historical prediction bias is measured using Huber loss, and the negative exponent of the differential entropy serves as a proxy for the degree of distribution concentration. The initial value of the exponentially weighted moving average of historical prediction bias is 0. At the end of each decision cycle, the Huber loss between the actual state at the current moment and the predicted state of the previous cycle is calculated. The threshold for Huber loss is 0.5; squared loss is used when the absolute value of the error is less than 0.5, otherwise linear loss is used. Then, the historical bias is updated exponentially with a decay factor of 0.9: the new bias equals 0.9 multiplied by the bias at the previous moment plus 0.1 multiplied by the current Huber loss.

[0083] The calculation process of differential entropy is as follows: First, calculate the negative logarithm of the mixing coefficient of each Gaussian component, multiply it by the coefficient itself, and then sum them to obtain the first term; then, for each Gaussian component, calculate the sum of the logarithms of its variances in each dimension, multiply it by the mixing coefficient of that component, sum it, divide it by 2, and add a constant term to obtain the second term; add the two terms to obtain the differential entropy; the negative exponent of the differential entropy is the value of the differential entropy with the natural constant e as the base and the exponent being negative; finally, the prediction confidence is equal to 1 divided by 1 plus the historical bias, and then multiplied by the negative exponent of the differential entropy; this value is between 0 and 1, and the smaller the historical bias and the more concentrated the distribution, the higher the confidence; take the average of the prediction confidence of all neighbors to obtain the global prediction confidence;

[0084] In step S3, the specific operation for each robot to decide whether to enable speculative pre-consensus is as follows:

[0085] The global prediction confidence, obtained by averaging the prediction confidence of all neighbors, is compared with a preset prediction confidence threshold. Simultaneously, the physical conflict urgency is compared with a preset physical conflict urgency threshold. The preset prediction confidence threshold is initially set to 0.75, which can be modulated in reverse by step S4, with a modulation range of 0.5 to 0.95. The preset physical conflict urgency threshold is initially set to 3.0 and remains fixed. The global prediction confidence is obtained by calculating the arithmetic mean of the prediction confidence of all neighbors, with a value range of 0 to 1. The physical conflict urgency comes from step S2 and is a non-negative real number.

[0086] If the global prediction confidence is not lower than the decision threshold and the physical conflict urgency is lower than the physical conflict urgency threshold, speculative pre-consensus is enabled, using the predicted future state to replace real-time communication for conflict resolution; otherwise, real-time communication mode is maintained. In speculative pre-consensus mode, the robot no longer broadcasts standard consensus messages, but directly uses the predicted neighbor future state in step S3 as input to perform physical constraint pruning and conflict proposal adjudication in step S2. This mode can save about 70% of communication overhead and reduce the decision latency from 800 milliseconds in real-time communication mode to less than 300 milliseconds. When the global prediction confidence is lower than 0.75 or the physical conflict urgency reaches 3.0 or above, the robot resumes real-time communication mode to ensure safety.

[0087] At the end of each decision cycle, the speculative execution hit rate and prediction error are updated. The speculative execution hit rate is the proportion of successful speculations over the past several decision cycles to the total number of uses. The prediction error is measured using the symmetric mean absolute percentage error. The statistical window length for the speculative execution hit rate is set to 50 decision cycles, or 10 seconds. A successful speculation is defined as: in a cycle where speculative pre-consensus is enabled, the Euclidean distance between the predicted future state and the actual received state is less than the tolerance error threshold of 0.2. The hit rate is equal to the number of successful speculations within the statistical window divided by the total number of uses. If speculative pre-consensus is not enabled within the statistical window, the hit rate remains the previous value.

[0088] The prediction error is measured using a symmetric mean absolute percentage error (MAS). The calculation process is as follows: For each neighbor, calculate the Euclidean distance between the actual state and the predicted state, divide it by the sum of the norms of the actual state and the predicted state, multiply by 2, and then take the average over all neighbors. To prevent division by zero, a very small positive number 0.000001 is added to the denominator. The prediction error ranges from 0 to 2, with smaller values ​​indicating more accurate predictions. The prediction execution hit rate and prediction error will be used as inputs to the state space of the Markov decision process in step S4 to evaluate the reliability of the current prediction system.

[0089] In this embodiment, it is specifically necessary to explain the specific operations of each robot in constructing the state space and reward function of the Markov decision process in step S4:

[0090] First, the network situation vector output in step S1 is aggregated and its dimensionality reduced. Five aggregated statistics are extracted: the average of the predicted latency of all neighboring links, the average of the predicted packet loss rate, the average of the Shannon entropy, the maximum of the predicted latency, and the maximum of the predicted packet loss rate. These are then combined with the prediction confidence, speculative execution hit rate, and prediction error output in step S3, and the physical conflict urgency output in step S2, to form a nine-dimensional state vector. The specific calculation process for aggregation and dimensionality reduction is as follows: For each robot, first, the average predicted latency of all its neighboring links is obtained... The average predicted latency is obtained by calculating the arithmetic mean of the values ​​in the value list; similarly, the average predicted packet loss rate is obtained by calculating the arithmetic mean of the predicted packet loss rates of all neighbor links; the average Shannon entropy is obtained by calculating the arithmetic mean of the Shannon entropy of all neighbor links; the maximum predicted latency is obtained by finding the maximum value in the value list of predicted latency; the maximum predicted packet loss rate is obtained by finding the maximum value in the value list of predicted packet loss rates; these five statistics, together with the prediction confidence, speculative execution hit rate, prediction error, and physical conflict urgency, are concatenated into a nine-dimensional vector in a fixed order.

[0091] The main network structure of the deep Q-network is as follows: 9 neurons in the input layer, 128 neurons in the first hidden layer using the Swish activation function, 64 neurons in the second hidden layer using the Swish activation function, and 4 neurons in the output layer corresponding to the Q-values ​​of the four actions. The target network structure is exactly the same as the main network, and the parameters are copied from the main network every 50 decision cycles. Training uses an experience replay mechanism with an experience pool capacity of 10,000 experiences. Each time, 32 experiences are randomly selected from the pool to calculate the loss. The loss function is Huber loss, with a threshold set to 1.0. The optimizer is Adam, with a learning rate set to 0.0003. The exploration rate is initially 1.0, decays by multiplying by 0.99 every decision cycle, and decreases to a minimum of 0.05.

[0092] Simultaneously, a reward function is designed, where the communication overhead penalty term is assigned a normalized value based on the selected optimal consensus strategy, the delay penalty term is assigned a normalized value based on the selected optimal consensus strategy, the collision risk term maps the physical conflict urgency to the zero-to-one interval using a logistic function, and the task delay term fuses the average prediction latency and prediction error to the zero-to-one interval using a hyperbolic tangent function. The four penalties are multiplied by dynamic weights, and the sum of these four values ​​is negative to obtain the reward value. The normalized value of the communication overhead penalty term is set as follows: 0.2 for the fast consensus strategy, and 0.2 for the robust consensus strategy. Consensus strategies are defined as follows: 1.0 for consensus strategy, 0.0 for speculative consensus strategy, and 0.3 for hybrid consensus strategy. The normalized values ​​for the delay penalty are set as follows: 0.3 for fast consensus strategy, 1.0 for robust consensus strategy, 0.1 for speculative consensus strategy, and 0.4 for hybrid consensus strategy. The logistic function takes the following form: collision risk equals 1 divided by 1 plus the physical conflict urgency base e and negative exponent minus 3.0. When the physical conflict urgency reaches 3.0, the collision risk is 0.5; after exceeding 3.0, the risk rises sharply and approaches 1.0.

[0093] The specific form of the hyperbolic tangent function is: the task delay equals the hyperbolic tangent function applied to the average prediction delay divided by 200 plus the prediction error; the average prediction delay is in milliseconds, and the prediction error ranges from 0 to 2. Therefore, the input of the hyperbolic tangent function is usually between 0 and 3, and the output is between 0 and 0.995; the initial value of the four dynamic weights is 0.25, which is dynamically updated by the adaptive weight adjustment mechanism in step S4.

[0094] Each robot inputs its nine-dimensional state vector into a deep Q-network, which outputs the Q-values ​​of four actions. A greedy strategy with decreasing exploration rate is used to select the action that maximizes the Q-value as the optimal consensus strategy. The greedy strategy with decreasing exploration rate is as follows: In each decision cycle, a random number between 0 and 1 is generated. If the random number is less than the current exploration rate, one of the four actions is randomly selected evenly as the current action; otherwise, the action with the largest Q-value output by the main network is selected. The exploration rate is multiplied by 0.99 after each decision cycle until it reaches a lower limit of 0.05. The selected action is the optimal consensus strategy for the current decision cycle, and the robot immediately executes the communication mode corresponding to this strategy.

[0095] In step S4, the specific operations of each robot performing inverse modulation and adaptive weight adjustment are as follows:

[0096] Based on the selected optimal consensus strategy and the current state, fuzzy control rules are used to adjust the future time window length in step S1, the positioning error margin in step S2, and the prediction confidence threshold in step S3. The adjustment of the future time window length is based on the selected optimal consensus strategy and the average prediction latency; the adjustment of the positioning error margin is based on the urgency of physical conflicts; and the adjustment of the prediction confidence threshold is based on the inferred execution hit rate and the prediction error. The specific logic of the fuzzy control rules is as follows: For the future time window length: if the selected strategy is fast consensus or hybrid consensus and the average prediction latency is less than 100 milliseconds, then the future time window length is increased. The time window length is reduced by 50 milliseconds, but not exceeding the upper limit of 2000 milliseconds. If the selected strategy is robust consensus or the average prediction latency is greater than 300 milliseconds, the future time window length is reduced by 50 milliseconds, but not lower than the lower limit of 200 milliseconds. If the selected strategy is speculative consensus, the future time window length remains unchanged. For positioning error margin: if the physical conflict urgency is greater than 2.0, the basic positioning error margin is increased by 0.05 meters, but not exceeding the upper limit of 0.3 meters. If the physical conflict urgency is less than 1.0 and the selected strategy is speculative consensus or fast consensus, the basic positioning error margin is reduced by 0.03 meters, but not lower than the lower limit of 0.05 meters. Otherwise, it remains unchanged.

[0097] For the prediction confidence threshold: if the prediction execution hit rate is greater than 0.9 and the prediction error is less than 0.1, the prediction confidence threshold is lowered by 0.02, but not lower than the lower limit of 0.5; if the prediction execution hit rate is less than 0.7 or the prediction error is greater than 0.3, the prediction confidence threshold is raised by 0.03, but not exceeding the upper limit of 0.95; otherwise, it remains unchanged; all modulation step sizes are small to avoid system oscillation.

[0098] Simultaneously, a sliding window is maintained to statistically analyze the actual average values ​​of recent communication overhead penalty, delay penalty, collision risk, and task delay. The four weight coefficients in the reward function are dynamically updated exponentially based on the deviation from the preset target value, and then normalized to sum to one. The length of the sliding window is set to 100 decision cycles, or 20 seconds. At the end of each decision cycle, the values ​​of the current cycle's communication overhead penalty, delay penalty, collision risk, and task delay are stored in the window, and the actual average value of each penalty within the window is calculated.

[0099] The preset target values ​​are: communication overhead target 0.3, latency target 0.4, collision risk target 0.1, and task delay target 0.2. The weight update adopts an exponential method: each weight is multiplied by the result of the preset target value minus the actual average value, with the natural constant e as the base, and then multiplied by the learning rate of 0.1. Then, the four updated weights are normalized so that their sum equals 1. This mechanism enables the system to automatically increase attention to the poorly performing targets, achieving adaptive balance of multiple targets.

[0100] It should be noted that the descriptions of each embodiment in the above embodiments have different focuses. For parts that are not described in detail in a certain embodiment, please refer to the relevant descriptions in other embodiments.

[0101] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0102] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded computer, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0103] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0104] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0105] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0106] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A distributed consensus-driven method for resolving conflicts among embodied intelligence multi-machine systems, characterized in that, Specifically, the steps include the following: Step S1: Each robot constructs a hidden Markov model based on historical communication logs to predict the latency and packet loss rate of each link within a preset future time window, and outputs a network situation vector. Step S2: Each robot calculates the physical conflict urgency with neighboring robots based on the network situation vector and its own kinematic parameters, including current speed, maximum braking deceleration and positioning error margin, using the minimum braking safety distance formula. This physical conflict urgency is used as a consensus priority scalar. Physical constraint pruning is performed on the received conflict proposals, and the pruning decision cache is output. The conflict proposal is a right-of-way occupancy request initiated by a neighboring robot. Step S3: Each robot uses a combined prediction model of long short-term memory and self-attention to predict the future state of its neighboring robots. Based on the preset prediction confidence threshold and the urgency of physical conflict, it decides whether to replace real-time communication with speculative pre-consensus and outputs the prediction confidence, speculative execution hit rate and prediction error. The prediction confidence is automatically calculated by the combined prediction model based on historical prediction deviations. Step S4: Each robot takes the network situation vector, prediction confidence, inferred execution hit rate, prediction error, and physical conflict urgency as state input Markov decision process, selects the optimal consensus strategy through deep Q network, and modulates the future time window length in step S1, the positioning error margin in the minimum braking safety distance formula in step S2, and the prediction confidence judgment threshold in step S3 according to the selected optimal consensus strategy to achieve adaptive switching of consensus strategy.

2. The distributed consensus-driven embodied intelligence multi-machine conflict resolution method according to claim 1, characterized in that: In step S1, the specific operations for each robot to construct a hidden Markov model are as follows: Define a set of hidden states, and use the generalized Pareto distribution of the associated time delay and the log-normal distribution of the packet loss rate as the observation probability density function for each hidden state; Each robot maintains a state transition count matrix. After each hidden state inference is completed, the state transition count matrix is ​​updated according to the maximum a posteriori hidden state sequence obtained by the Viterbi algorithm, and the estimated value of the transition matrix is ​​calculated by the Dirichlet a posteriori expectation formula. Among them, the generalized Pareto distribution is used to characterize the heavy-tailed feature of network latency, the log-normal distribution is used to describe the skewed distribution of packet loss rate in the interval between zero and one, and the update of the transition matrix adopts the Dirichlet-multinomial conjugate model to realize sequential incremental learning without storing complete historical data.

3. The distributed consensus-driven embodied intelligence multi-machine conflict resolution method according to claim 2, characterized in that: In step S1, the specific operations for each robot to predict the latency and packet loss rate of each link within the future time window are as follows: At the beginning of each preset decision cycle, the posterior distribution of the hidden state at the current time is calculated using the forward algorithm; Subsequently, the matrix exponential approximation method based on the transition matrix calculates the distribution of the hidden state at the end of the future time window based on the posterior distribution of the current hidden state, where the matrix exponent is calculated quickly using Pad approximation. Then, based on the hidden state distribution at the end of the future time window, the mean predicted delay and the mean predicted packet loss rate of each link are calculated respectively, and the Shannon entropy of the hidden state distribution at the end of the future time window is calculated as a measure of prediction uncertainty. Finally, the average predicted latency, average predicted packet loss rate, and Shannon entropy of all neighboring links are concatenated into a vector to form the network situation vector.

4. The distributed consensus-driven embodied intelligence multi-machine conflict resolution method according to claim 3, characterized in that: In step S2, the specific operation for determining the minimum braking safety distance between each robot and its neighboring robots is as follows: First, the mean predicted delay and mean predicted packet loss rate of each link are extracted from the network situation vector. Then, the positioning error margin is calculated using the exponential saturation function. The exponential saturation function takes the mean predicted delay as the independent variable and the increment of the positioning error margin as the dependent variable, and adds a linear term of the mean predicted packet loss rate to obtain the final positioning error margin. Then, the minimum braking safety distance is calculated by summing the square of its current speed divided by twice its maximum braking deceleration, the square of the neighbor's current speed divided by twice the neighbor's maximum braking deceleration, and the positioning error margin.

5. The distributed consensus-driven embodied intelligence multi-machine conflict resolution method according to claim 4, characterized in that: In step S2, the specific operations for each robot to generate physical conflict urgency and consensus priority scalars are as follows: First, obtain the current Euclidean distance between yourself and each neighboring robot, and compare it with the minimum braking safety distance. If the current Euclidean distance is less than the minimum braking safety distance, the physical conflict indication value is set to 1; otherwise, it is set to 0. Then, the physical conflict indication value corresponding to each neighbor is multiplied by one, plus the product of a preset uncertainty amplification factor and the Shannon entropy of that link in the network situation vector, and then summed over all neighbors to obtain the physical conflict urgency. Meanwhile, the consensus priority scalar is obtained by multiplying the preset task priority, the inverse of the preset path cost, and the inverse of the average predicted latency calculated from the average predicted latency of all neighbor links in the network situation vector by preset weight coefficients and summing them. Finally, the robot decides on the received conflicting proposals based on the consensus priority scalar. If the consensus priority scalar of this robot is higher than the consensus priority scalar of the proposal initiator, the proposal is rejected; otherwise, it is allowed to pass, and the decision is stored in the pruning decision cache.

6. The distributed consensus-driven embodied intelligence multi-machine conflict resolution method according to claim 5, characterized in that: In step S3, the specific operations for each robot to construct the neighbor's historical state sequence are as follows: Each robot maintains a fixed-length sliding window to store the state packets received from each neighboring robot in the most recent decision cycles. The state packets include position coordinates, velocity vector, task priority, and consensus priority scalar. This information is then concatenated into a feature vector to obtain a sequence of feature vectors. Subsequently, a long short-term memory network is used to temporally encode the feature vector sequence of each neighbor to obtain the hidden state of each neighbor; Then, a multi-head self-attention mechanism is introduced to interact and fuse the hidden states of all neighbors, and output the fused feature vector of each neighbor.

7. The distributed consensus-driven embodied intelligence multi-machine conflict resolution method according to claim 6, characterized in that: In step S3, the specific operations for each robot to predict the future state of its neighbors and calculate the prediction confidence are as follows: The hidden state of each neighbor is concatenated with the fused feature vector and input into the mixing density network. The network outputs the mixing coefficients, mean, and covariance diagonal elements of multiple Gaussian components to obtain the probability distribution of the future state. The weighted sum of the means of each Gaussian component is used as the predicted future state. Meanwhile, prediction confidence is calculated based on the exponentially weighted moving average of historical prediction bias and the negative exponent of the differential entropy of the probability distribution of future states. The historical prediction bias is measured by Huber loss, and the negative exponent of the differential entropy serves as a proxy for the degree of distribution concentration.

8. The distributed consensus-driven embodied intelligence multi-machine conflict resolution method according to claim 7, characterized in that: In step S3, the specific operation for each robot to decide whether to enable speculative pre-consensus is as follows: The global prediction confidence, obtained by averaging the prediction confidence of all neighbors, is compared with a preset prediction confidence threshold. At the same time, the physical conflict urgency is compared with a preset physical conflict urgency threshold. If the global prediction confidence is not lower than the judgment threshold and the physical conflict urgency is lower than the physical conflict urgency threshold, then speculative pre-consensus is enabled, and the predicted future state is used to replace real-time communication to participate in conflict resolution; otherwise, real-time communication mode is maintained. At the end of each decision cycle, the prediction execution hit rate and prediction error are updated. The prediction execution hit rate is the proportion of successful predictions in the past decision cycles to the total number of uses. The prediction error is measured by the symmetric mean absolute percentage error.

9. A distributed consensus-driven method for resolving conflicts among embodied intelligent multi-machine systems according to claim 8, characterized in that: In step S4, the specific operations for each robot to construct the state space and reward function of the Markov decision process are as follows: First, the network situation vector output in step S1 is aggregated and dimensionality reduced. The average value of the predicted delay of all neighbor links, the average value of the predicted packet loss rate, the average value of Shannon entropy, the maximum value of the predicted delay, and the maximum value of the predicted packet loss rate are extracted, for a total of five aggregated statistics. These are then combined with the prediction confidence, inference execution hit rate, prediction error, and physical conflict urgency output in step S3 to form a nine-dimensional state vector. Meanwhile, a reward function is designed, in which the communication overhead penalty term is assigned a normalized value according to the selected optimal consensus strategy, the delay penalty term is assigned a normalized value according to the selected optimal consensus strategy, the collision risk term maps the physical conflict urgency to the zero-to-one interval through the logistic function, and the task delay term fuses the average prediction delay and prediction error to the zero-to-one interval through the hyperbolic tangent function. The four penalties are multiplied by dynamic weights respectively, and the sum is taken as the negative value as the reward value. Each robot inputs its nine-dimensional state vector into a deep Q-network, outputs the Q-values ​​of four actions, and uses a greedy strategy with decreasing exploration rate to select the action that maximizes the Q-value as the optimal consensus strategy.

10. A distributed consensus-driven method for resolving conflicts among embodied intelligent multi-machine systems according to claim 9, characterized in that: In step S4, the specific operations of each robot performing inverse modulation and adaptive weight adjustment are as follows: Based on the selected optimal consensus strategy and the current state, fuzzy control rules are used to adjust the future time window length in step S1, the positioning error margin in step S2, and the prediction confidence threshold in step S3, respectively. The adjustment of the future time window length is based on the selected optimal consensus strategy and the average prediction delay, the adjustment of the positioning error margin is based on the urgency of physical conflicts, and the adjustment of the prediction confidence threshold is based on the inferred execution hit rate and the prediction error. Meanwhile, a sliding window is maintained to statistically analyze the actual average values ​​of recent communication overhead penalties, delay penalties, collision risk items, and task delay items. The four weight coefficients in the reward function are dynamically updated exponentially based on the deviation from the preset target value, and then normalized so that their sum is one.

Images