Deep learning-based autonomous decision-making method and system for automatic driving carrying robot

Abstract

The invention discloses a deep learning-based autonomous decision-making method for an automatic driving carrying robot, and particularly relates to the technical field of automatic control. According to the invention, full-process optimization of environment perception, decision planning and motion control is realized through a three-layer progressive decision architecture. The method specifically comprises the steps that a global target guiding layer generates a global path based on a semantic map; the local situation reasoning layer adopts a multi-scale space-time attention mechanism to carry out environment perception and reasoning; the real-time motion control layer generates a control instruction through on-line optimization of a safe motion envelope; the technical problems that in the prior art, simulation and reality are poor, the dynamic environment adaptability is insufficient, and the decision-making safety is poor are effectively solved, and the decision-making reliability, the environment adaptability and the operation safety of the automatic driving carrying robot in a complex scene are remarkably improved.

Description

Technical Field

[0001] This invention relates to the field of automation control technology, and in particular to an autonomous decision-making method and system for an autonomous driving transport robot based on deep learning. Background Technology

[0002] With the rapid development of artificial intelligence and robotics, autonomous transport robots are increasingly being used in logistics warehousing, smart ports, and urban delivery. These robots need to achieve fully autonomous navigation and decision-making in complex dynamic environments, and their core technology lies in the integrated application of environmental perception, decision planning, and motion control.

[0003] Currently, the autonomous decision-making methods for autonomous driving transport robots mainly fall into the following technical categories: First, there's the decision-making method based on traditional modular architecture. This method divides the autonomous driving system into independent modules such as perception, localization, planning, and control, making decisions through a serial pipeline approach. While this architecture offers good interpretability, it has significant limitations: errors between modules accumulate and propagate at each level, leading to a decline in overall system performance; fixed interfaces between modules struggle to adapt to dynamically changing environments; and the lack of a global optimization mechanism makes it prone to decision conflicts in complex scenarios. This rigid architecture exhibits poor adaptability, especially when encountering unforeseen edge cases.

[0004] Second, there is the end-to-end deep learning approach. This method directly maps sensor data into control commands using a single deep neural network, avoiding the problems caused by module partitioning. However, this approach has significant drawbacks: it lacks a clear safety mechanism, has poor interpretability of the decision-making process, is sensitive to the distribution of training data, and performs inconsistently when encountering scenarios not covered by the training set; it is difficult to incorporate domain knowledge and physical constraints, potentially leading to dangerous actions that do not conform to the laws of dynamics.

[0005] Third, there are reinforcement learning-based decision-making methods. These methods learn optimal policies through interaction between the agent and its environment, demonstrating advantages in handling sequential decision-making problems. However, existing reinforcement learning methods face significant challenges in autonomous driving applications: the simulation environment differs significantly from the real world (the Sim2Real problem), causing policies trained in simulation to perform poorly in real-world environments; training efficiency is low, requiring large amounts of interactive data; and safety guarantees are insufficient, potentially leading to dangerous behaviors during exploration.

[0006] In addition, existing methods face common technical challenges in the following aspects: In terms of environmental perception, the spatiotemporal inconsistency of multi-sensor data leads to unsatisfactory fusion results. Differences in sensor acquisition frequencies, data processing latency, and coordinate systems make it difficult to achieve accurate spatiotemporal alignment in dynamic environments. Especially in rapidly changing scenarios, this lag and misalignment of perceived information can severely impact the accuracy of decision-making.

[0007] Regarding decision security, most existing methods employ a posteriori security verification mechanism, meaning that security assessment is performed after a decision is generated. This "decision first, security later" model has inherent flaws: when an unsafe decision is discovered, there may not be sufficient time and space leeway for correction. Furthermore, there is a lack of effective mechanisms to embed security constraints into the decision generation process.

[0008] In terms of adaptability, existing systems often require redesign or extensive parameter adjustments when facing new operating environments or task requirements. They lack the ability to adapt to new scenarios through autonomous learning and evolution while maintaining the core architecture. This adaptability deficiency severely limits the deployment scope and practical value of autonomous driving transport robots.

[0009] Especially in practical applications such as logistics warehousing and port transportation, autonomous transport robots need to cope with multiple challenges, including mixed traffic of people and vehicles, dynamic obstacles, and complex weather conditions. Existing technologies have significant shortcomings in terms of safety, reliability, and adaptability. Therefore, there is an urgent need for a new autonomous decision-making method that can fully consider the special application requirements of autonomous transport robots, possess strong environmental adaptability, inherent safety guarantees, and continuous evolution capabilities. Summary of the Invention

[0010] The purpose of this invention is to overcome the shortcomings of the prior art and provide an autonomous decision-making method and system for autonomous driving transport robots based on deep learning. It aims to solve technical problems such as the gap between simulation and reality, insufficient adaptability to dynamic environments, and lack of decision-making safety in the prior art, and significantly improve the decision-making reliability, environmental adaptability, and operational safety of autonomous driving transport robots in complex scenarios.

[0011] To achieve the above objectives, this application proposes a deep learning-based autonomous decision-making method for autonomous driving transport robots, comprising the following steps: Environmental data is collected through multiple sensors, including lidar, cameras, and inertial measurement units. The environmental data is input into the autonomous decision-making model trained based on the curriculum adversarial self-evolutionary learning framework; The autonomous decision-making model performs a three-level progressive decision-making process, including: A global goal-oriented layer generates global path planning based on semantic maps and task objectives; The local contextual reasoning layer employs a multi-scale spatiotemporal attention mechanism to perceive and reason about the dynamic environment, generating local navigation strategies. The real-time motion control layer generates control commands based on an online optimization method for safe motion envelopes. The autonomous driving robot executes corresponding motion control according to the control instructions.

[0012] As a further solution, the curriculum adversarial self-evolutionary learning framework includes: During the course learning phase, autonomous decision-making models are trained in a simulation environment in order of increasing difficulty. The difficulty is dynamically adjusted based on environmental complexity, obstacle density, and task requirements. During the adversarial training phase, adversarial networks are used to generate challenging scenarios that can cause the autonomous decision-making model to fail. These challenging scenarios include suddenly appearing obstacles and aggressive driving behaviors. During the self-evolution phase, edge cases encountered in actual operation are fed back to the training data pool for incremental training of the model.

[0013] As a further solution, the course learning phase employs a difficulty evaluation function based on multi-objective optimization: in, SR t Indicates the success rate. CR t Indicates the risk coefficient. UNC t This indicates uncertainty in decision-making. α, β, γ These represent the corresponding weight coefficients; when Difficulty t Increase training difficulty when the score is below the first threshold. Difficulty t Reduce training difficulty when the threshold is exceeded.

[0014] As a further solution, the multi-scale spatiotemporal attention mechanism includes: Multi-sensor data is spatiotemporally aligned using a cross-modal spatiotemporal alignment module, which employs a hybrid architecture of deformable convolution and temporal calibration. By employing a hierarchical attention distillation mechanism to progressively compress multi-scale features, the computational complexity is reduced from O(N^2) to O(N^2). 2 The time complexity is reduced to O(NlogN); The attention intensity of different spatiotemporal features is dynamically adjusted by using a spatiotemporal attention weight matrix. The calculation formula for the spatiotemporal attention weight matrix is ​​as follows: Among them, M geometric Let T represent the geometric consistency matrix. delay This represents time delay compensation, σ represents the activation function, Q represents the query vector, K represents the key vector, and d represents the time delay compensation. k This represents the dimension of the key vector.

[0015] As a further solution, the online optimization method for secure motion envelope includes: Based on neural differential equations, learn the dynamic model of autonomous driving vehicles and predict the safe motion envelope in the future time period. The boundary of the safe motion envelope is dynamically optimized based on the real-time perceived environmental risks using an incremental convex optimization solver. The safety of the safety motion envelope is ensured through a safety verification and backoff mechanism, which triggers backoff control when a potential collision risk is detected.

[0016] As a further solution, the online optimization process of the secure motion envelope is represented as a constrained optimization problem: min_E[λ1×Risk(E)+λ2×(1 / Utility(E))]; Constraints: E Feasible_Trajectories; Where E represents the optimized safe motion envelope, Risk(E) represents the expected collision probability, Utility(E) represents the envelope volume, and λ1 and λ2 are weighting coefficients.

[0017] As a further solution, the three-layer progressive decision processing also includes a cross-level information coupling mechanism: The flow of information between decision-making levels is controlled through a cross-level attention gating mechanism, wherein the gating function is: g_i→j=σ(W_g·[h_i,h_j,c_context]+b_g); Where h_i and h_j represent the features of the source layer and the target layer, respectively, c_context represents the context information, and W_g and b_g are learnable parameters; Consistency constraint propagation transmits high-level constraints to the lower level as dual variables.

[0018] As a further solution, adaptive computing resource allocation strategies are also included: Computational resources are dynamically allocated based on the scene risk level, which is calculated based on obstacle distance, motion state, and environmental complexity. The prediction time domain is dynamically adjusted based on the motion state, and the calculation formula for the prediction time domain is as follows: T_horizon=T_min+(T_max-T_min)×(v / v_max)×(d_obstacle / d_safe); Where v represents the current speed, T_min represents the minimum prediction time domain, T_max represents the maximum prediction time domain, v_max represents the maximum allowed speed, d_obstacle represents the distance to the nearest obstacle, and d_safe represents the safe distance.

[0019] As a further solution, the adversarial network in the adversarial training phase adopts a hierarchical adversarial training mechanism, including: An environmental dynamic adversarial network is used to generate abnormal behavior patterns of dynamic obstacles. Its objective function is: max(φ_e)E[s~p_env][log(1-D_policy(π_main(s)))]; Where φ_e represents the parameters of the dynamic adversarial network, D_policy is the policy discriminator, π_main is the policy of the master decision network, s represents the environment state, p_env represents the environment state distribution, and E represents the mathematical expectation. Sensor interference adversarial networks are used to simulate multi-sensor data failure and noise interference. Their interference modes include: LiDAR point cloud loss: Randomly discard point cloud data according to the loss rate ρ; Camera image occlusion: Generate a random occlusion area based on the occlusion area A_occlude; IMU data drift: Injecting random bias noise b_imu; Scene complexity adversarial networks are used to adjust the environment structure and task constraints. They learn scene topology relationships through graph neural networks to generate challenging scene configurations. G_challenge=GNN_θ_g(G_base⊕ΔG_adv); Where G_base=(V_base,E_base) is the base scene graph, V_base={v_i},i=1toN is the set of nodes, E_base={e_ij} is the set of edges, ΔG_adv=(ΔV,ΔE) is the adversarial graph structure perturbation, ⊕ represents the graph structure fusion operation, and GNN_θ_g represents the graph neural network with parameter θ_g; The hierarchical adversarial training mechanism employs an alternating optimization strategy, updating sequentially in each training round t: Parameters of the dynamic adversarial network: φ_e^(t+1)=φ_e^(t)+η_e× _φ_eJ_e(φ_e^(t)); Sensor interference adversarial network parameters: φ_s^(t+1)=φ_s^(t)+η_s× _φ_sJ_s(φ_s^(t)); Parameters of the adversarial network for scene complexity: θ_g^(t+1)=θ_g^(t)+η_g× _θ_gJ_g(θ_g^(t)); Where η_e, η_s, and η_g are the learning rates of each adversarial network, and J_e, J_s, and J_g are the optimization objective functions of each adversarial network. Represents the gradient operator; The hierarchical adversarial training process forms a multi-round game training with the master decision network until a Nash equilibrium is reached, which satisfies: E[J_main(π_main,{φ_e,φ_s,θ_g})]≤E[J_main(π_main,{φ_e,φ_s,θ_g})]; E[J_adv(π_main,{φ_e,φ_s,θ_g})]≤E[J_adv(π_main,{φ_e,φ_s,θ_g})]; Where π_main represents the optimal master decision network strategy, {φ_e,φ_s,θ_g} represents the optimal set of parameters for the adversarial network, and J_main and J_adv represent the overall objective functions of the master decision network and the adversarial network, respectively.

[0020] On the other hand, the present invention also provides an autonomous driving transport robot system, comprising: A multi-sensor module is used to collect environmental data; A processor configured to execute a deep learning-based autonomous decision-making method for an autonomous vehicle robot as described in any of the preceding items; The control execution module is used to perform motion control according to the control instructions generated by the processor.

[0021] Compared with related technologies, the autonomous decision-making method and system for autonomous driving transport robots based on deep learning provided by this invention have the following advantages: 1. This invention utilizes a course-based adversarial self-evolutionary learning framework. The system undergoes progressive training from simple to complex in a simulation environment, followed by challenging training with a hierarchical adversarial network (environmental dynamics, sensor interference, scene complexity), enabling the decision-making model to possess strong adaptability to various extreme scenarios and edge cases. 2. This invention innovatively proposes an online optimization method for safe motion envelope, transforming the traditional posterior safety check into a priori safety constraint, thereby avoiding the generation of dangerous decisions; by learning the dynamic model through neural differential equations and combining it with an incremental convex optimization solver, the real-time accurate calculation of the safety boundary is achieved. 3. This invention effectively bridges the gap between the simulation environment and the real world by combining course learning with adversarial training. The multi-target adversarial scenario generator can automatically generate high-quality training samples, reducing the workload of manually designing test scenarios; the multi-scale spatiotemporal attention mechanism effectively solves the problem of spatiotemporal inconsistency of data from different sensors through a cross-modal spatiotemporal alignment module and a hierarchical attention distillation mechanism. 4. The three-layer progressive decision architecture of this invention achieves deep coupling and efficient collaboration among decision levels through cross-level attention gating mechanism and consistency constraint propagation; the self-evolution mechanism enables the system to continuously learn and improve by collecting edge cases online and feeding them back to the training loop. Attached Figure Description

[0022] The accompanying drawings, which are incorporated in and form part of this specification, illustrate embodiments consistent with this application and, together with the description, serve to explain the principles of this application.

[0023] To more clearly illustrate the technical solutions in the embodiments of this application or related technologies, the accompanying drawings used in the description of the embodiments or related technologies will be briefly introduced below. Obviously, those skilled in the art can obtain other drawings based on these drawings without creative effort.

[0024] Figure 1 A schematic diagram illustrating the steps of an autonomous decision-making method for an autonomous driving transport robot based on deep learning, provided by this invention; Figure 2 A schematic diagram of an autonomous decision-making system for an autonomous driving transport robot based on deep learning, provided by the present invention; The purpose, features, and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0026] Please see Figure 1This embodiment provides a deep learning-based autonomous decision-making method for autonomous driving vehicles, which includes the following steps: Environmental data is collected through multiple sensors, including lidar, cameras, and inertial measurement units. The environmental data is input into the autonomous decision-making model trained based on the curriculum adversarial self-evolutionary learning framework; The autonomous decision-making model performs a three-level progressive decision-making process, including: A global goal-oriented layer generates global path planning based on semantic maps and task objectives; The local contextual reasoning layer employs a multi-scale spatiotemporal attention mechanism to perceive and reason about the dynamic environment, generating local navigation strategies. The real-time motion control layer generates control commands based on an online optimization method for safe motion envelopes. The autonomous driving robot executes corresponding motion control according to the control instructions.

[0027] It should be noted that the core of this embodiment lies in constructing a "curriculum-based adversarial self-evolutionary learning framework" that integrates perception, decision-making, and control, and employs a "three-layer progressive decision-making architecture" to execute specific tasks. This method first perceives the environment through multi-sensor fusion, then the autonomous decision-making model, trained through progressive adversarial methods, performs three layers of processing: a global goal-oriented layer performs macro-level task decomposition and path planning; a local contextual reasoning layer uses a multi-scale spatiotemporal attention mechanism to achieve accurate perception and real-time reasoning of the dynamic environment; and finally, a real-time motion control layer generates safe control commands online through safe motion envelope optimization. Its key improvement lies in the organic integration of traditionally separate modules through end-to-end deep learning methods, and the introduction of curriculum learning and adversarial training mechanisms to enhance the model's robustness and environmental adaptability.

[0028] Furthermore, the curriculum adversarial self-evolutionary learning framework includes: During the course learning phase, autonomous decision-making models are trained in a simulation environment in order of increasing difficulty. The difficulty is dynamically adjusted based on environmental complexity, obstacle density, and task requirements. During the adversarial training phase, adversarial networks are used to generate challenging scenarios that can cause the autonomous decision-making model to fail. These challenging scenarios include suddenly appearing obstacles and aggressive driving behaviors. During the self-evolution phase, edge cases encountered in actual operation are fed back to the training data pool for incremental training of the model.

[0029] Specifically, this implementation framework first employs a progressive training strategy, from easy to difficult, in a simulation environment during the course learning phase, enabling the model to robustly learn basic decision-making capabilities. Then, it enters the adversarial training phase, utilizing adversarial networks to proactively generate challenging scenarios (such as sudden obstacles), thereby forcibly improving the model's decision-making robustness in extreme situations. Finally, through a self-evolution phase, a closed loop from theory to practice is formed. The system can automatically recycle "marginal cases" with high decision uncertainty or near-failure encountered in actual operation and use them for incremental optimization of the model, thus achieving continuous self-evolution of decision-making capabilities.

[0030] These three stages constitute a complete "training-challenge-evolution" cycle, systematically solving the problem of model adaptability from simulation environments to real-world applications. It not only significantly improves the model's performance when facing known challenges, but more importantly, it endows the system with the ability to generalize to unknown scenarios and the ability to learn throughout life, which is something that traditional one-time training models cannot achieve.

[0031] Furthermore, the course learning phase employs a difficulty evaluation function based on multi-objective optimization: in, SR t Indicates the success rate. CR t Indicates the risk coefficient. UNC t This indicates uncertainty in decision-making. α, β, γ These represent the corresponding weight coefficients; when Difficulty t Increase training difficulty when the score is below the first threshold. Difficulty t Reduce training difficulty when the threshold is exceeded.

[0032] Specifically, this embodiment quantifies the difficulty assessment mechanism during the course learning phase. By introducing a multi-objective optimized difficulty assessment function, adaptive control of the training process is achieved. The core of this embodiment lies in transforming the originally subjective difficulty judgment into a quantitative assessment system based on success rate, risk coefficient, and decision uncertainty. Through mathematical modeling, dynamic adjustment and scientific management of course difficulty are realized.

[0033] This embodiment shifts control of the training process from human experience to data-driven approaches. By automatically adjusting the training difficulty through real-time performance feedback, it avoids both model crashes caused by overly difficult training and performance bottlenecks resulting from overly easy training. This quantification mechanism ensures that the learning phase always operates within the optimal challenge range, thereby significantly improving training efficiency and the final model performance.

[0034] Furthermore, the multi-scale spatiotemporal attention mechanism includes: Multi-sensor data is spatiotemporally aligned using a cross-modal spatiotemporal alignment module, which employs a hybrid architecture of deformable convolution and temporal calibration. By employing a hierarchical attention distillation mechanism to progressively compress multi-scale features, the computational complexity is reduced from O(N^2) to O(N^2). 2 The time complexity is reduced to O(NlogN); The attention intensity of different spatiotemporal features is dynamically adjusted by using a spatiotemporal attention weight matrix. The calculation formula for the spatiotemporal attention weight matrix is ​​as follows: Among them, M geometric Let T represent the geometric consistency matrix. delay This represents time delay compensation, σ represents the activation function, Q represents the query vector, K represents the key vector, and d represents the time delay compensation. k This represents the dimension of the key vector.

[0035] Specifically, this embodiment refines the key technologies of the multi-scale spatiotemporal attention mechanism and constructs a complete cross-modal spatiotemporal data processing workflow. Its core innovation lies in solving the problem of inconsistency between multi-sensor data in time and space through a cross-modal spatiotemporal alignment module, and in significantly reducing computational complexity while maintaining perception performance by adopting a hierarchical attention distillation mechanism, thus forming a full-link optimization from data fusion to efficient computation.

[0036] The outstanding contribution of this embodiment lies in simultaneously overcoming two key challenges in autonomous driving environmental perception: on the one hand, it achieves accurate alignment of multi-source sensor data through a hybrid architecture of deformable convolution and temporal calibration; on the other hand, it reduces the computational complexity from quadratic to linear logarithmic levels through progressive feature compression, thereby meeting the stringent real-time requirements of autonomous driving systems while ensuring perception accuracy.

[0037] Furthermore, the online optimization method for the safe motion envelope includes: Based on neural differential equations, learn the dynamic model of autonomous driving vehicles and predict the safe motion envelope in the future time period. The boundary of the safe motion envelope is dynamically optimized based on the real-time perceived environmental risks using an incremental convex optimization solver. The safety of the safety motion envelope is ensured through a safety verification and backoff mechanism, which triggers backoff control when a potential collision risk is detected.

[0038] Specifically, this embodiment refines the key technologies of the online optimization method for safe motion envelopes, constructing a complete real-time safety assurance system. Its core innovation lies in accurately learning vehicle dynamics characteristics through neural differential equations to predict safety boundaries, employing an incremental convex optimization solver to achieve real-time dynamic adjustment of the envelope, and finally ensuring system safety under extreme conditions through formal verification and a backoff mechanism.

[0039] The key value of this embodiment lies in transforming the traditional static safety boundary into a dynamic and optimizable safety envelope. This overcomes the conservatism inherent in fixed constraints and ensures the safety of real-time decisions through a rigorous mathematical verification mechanism. This "prediction-optimization-verification" technical closed loop enables autonomous driving robots to intelligently balance safety and maneuverability in complex dynamic environments, fundamentally improving the system's reliability and decision-making quality.

[0040] Furthermore, the online optimization process of the safe motion envelope is expressed as a constrained optimization problem: min_E[λ1×Risk(E)+λ2×(1 / Utility(E))]; Constraints: E Feasible_Trajectories; Where E represents the optimized safe motion envelope, Risk(E) represents the expected collision probability, Utility(E) represents the envelope volume, and λ1 and λ2 are weighting coefficients.

[0041] Specifically, this embodiment formalizes the safety motion envelope optimization problem into a constrained minimization problem using mathematical methods, clarifying the optimization objective and constraints. Its core innovation lies in constructing a dual-objective optimization framework that simultaneously considers collision risk expectation and decision flexibility, achieving a dynamic trade-off between safety and traffic efficiency through weighting coefficients.

[0042] The key contribution of this embodiment lies in transforming the safety boundary problem, which originally relied on experience-based adjustments, into a computable optimization problem, providing a rigorous mathematical foundation for safety decision-making in autonomous driving systems. This formal formulation not only ensures the theoretical rigor of safety envelope optimization but also, through explicit optimization objectives and constraints, enables the system to automatically find the optimal balance between safety and efficiency in complex scenarios, significantly improving the intelligent decision-making level of autonomous driving robots.

[0043] Furthermore, the three-layer progressive decision processing also includes a cross-level information coupling mechanism: The flow of information between decision-making levels is controlled through a cross-level attention gating mechanism, wherein the gating function is: g_i→j=σ(W_g·[h_i,h_j,c_context]+b_g); Where h_i and h_j represent the features of the source layer and the target layer, respectively, c_context represents the context information, and W_g and b_g are learnable parameters; Consistency constraint propagation transmits high-level constraints to the lower level as dual variables.

[0044] Specifically, this embodiment proposes an innovative solution for cross-level information coupling to address the collaboration problem among three-layer decision-making architectures. Its core lies in achieving intelligent control of information flow through a cross-level attention gating mechanism, and accurately transmitting high-level decision intentions to the lower-level execution units in mathematical form using consistency constraint propagation. This effectively solves the technical challenges of decision disconnect and goal conflict in traditional layered architectures.

[0045] The significant advancement of this embodiment lies in establishing a closed-loop mechanism for bidirectional communication between layers. This mechanism not only ensures the local execution of global strategies but also guarantees that lower-level feedback can influence higher-level decisions in real time. This deeply coupled design enables all layers of the autonomous driving system to form an organic whole, significantly improving the consistency of decision-making and system stability in complex scenarios. It represents a key technological breakthrough in realizing intelligent decision-making from "layered fragmentation" to "collaborative integration."

[0046] Furthermore, it also includes adaptive computing resource allocation strategies: Computational resources are dynamically allocated based on the scene risk level, which is calculated based on obstacle distance, motion state, and environmental complexity. The prediction time domain is dynamically adjusted based on the motion state, and the calculation formula for the prediction time domain is as follows: T_horizon=T_min+(T_max-T_min)×(v / v_max)×(d_obstacle / d_safe); Where v represents the current speed, T_min represents the minimum prediction time domain, T_max represents the maximum prediction time domain, v_max represents the maximum allowed speed, d_obstacle represents the distance to the nearest obstacle, and d_safe represents the safe distance.

[0047] Specifically, this embodiment constructs an adaptive computing resource management mechanism that achieves intelligent allocation of computing resources through dynamic risk assessment. Its core innovation lies in establishing a direct correlation between scenario risk level and computing resource allocation, and designing a dynamic adjustment formula in the prediction time domain based on motion state, enabling the system to optimize computing efficiency while ensuring safety.

[0048] The key value of this embodiment lies in resolving the contradiction between the limited computing resources of autonomous driving systems and the demands of handling complex environments. By quantifying risk as a function of multiple factors such as speed and obstacle distance, the system can intelligently allocate more computing resources to ensure safety in high-risk scenarios and conserve resources to improve efficiency in low-risk scenarios. This flexible computing architecture significantly enhances the practicality and adaptability of autonomous vehicles under resource-constrained conditions.

[0049] Furthermore, the adversarial network in the adversarial training phase adopts a hierarchical adversarial training mechanism, including: An environmental dynamic adversarial network is used to generate abnormal behavior patterns of dynamic obstacles. Its objective function is: max(φ_e)E[s~p_env][log(1-D_policy(π_main(s)))]; Where φ_e represents the parameters of the dynamic adversarial network, D_policy is the policy discriminator, π_main is the policy of the master decision network, s represents the environment state, p_env represents the environment state distribution, and E represents the mathematical expectation. Sensor interference adversarial networks are used to simulate multi-sensor data failure and noise interference. Their interference modes include: LiDAR point cloud loss: Randomly discard point cloud data according to the loss rate ρ; Camera image occlusion: Generate a random occlusion area based on the occlusion area A_occlude; IMU data drift: Injecting random bias noise b_imu; Scene complexity adversarial networks are used to adjust the environment structure and task constraints. They learn scene topology relationships through graph neural networks to generate challenging scene configurations. G_challenge=GNN_θ_g(G_base⊕ΔG_adv); Where G_base=(V_base,E_base) is the base scene graph, V_base={v_i},i=1toN is the set of nodes, E_base={e_ij} is the set of edges, ΔG_adv=(ΔV,ΔE) is the adversarial graph structure perturbation, ⊕ represents the graph structure fusion operation, and GNN_θ_g represents the graph neural network with parameter θ_g; The hierarchical adversarial training mechanism employs an alternating optimization strategy, updating sequentially in each training round t: Parameters of the dynamic adversarial network: φ_e^(t+1)=φ_e^(t)+η_e× _φ_eJ_e(φ_e^(t)); Sensor interference adversarial network parameters: φ_s^(t+1)=φ_s^(t)+η_s× _φ_sJ_s(φ_s^(t)); Parameters of the adversarial network for scene complexity: θ_g^(t+1)=θ_g^(t)+η_g× _θ_gJ_g(θ_g^(t)); Where η_e, η_s, and η_g are the learning rates of each adversarial network, and J_e, J_s, and J_g are the optimization objective functions of each adversarial network. Represents the gradient operator; The hierarchical adversarial training process forms a multi-round game training with the master decision network until a Nash equilibrium is reached, which satisfies: E[J_main(π_main,{φ_e,φ_s,θ_g})]≤E[J_main(π_main,{φ_e,φ_s,θ_g})]; E[J_adv(π_main,{φ_e,φ_s,θ_g})]≤E[J_adv(π_main,{φ_e,φ_s,θ_g})]; Where π_main represents the optimal master decision network strategy, {φ_e,φ_s,θ_g} represents the optimal set of parameters for the adversarial network, and J_main and J_adv represent the overall objective functions of the master decision network and the adversarial network, respectively.

[0050] Specifically, this embodiment significantly deepens the adversarial training mechanism, constructing a layered adversarial training architecture. Its core innovation lies in refining the single adversarial network into three specialized adversarial networks: environmental dynamics, sensor interference, and scene complexity. These networks respectively target dynamic obstacle behavior, multi-sensor failure modes, and environmental topology, forming a comprehensive and multi-layered adversarial training system.

[0051] The key value of this embodiment lies in its systematic design of adversarial attack dimensions, which significantly enhances the adaptability of the main decision-making network to various extreme scenarios. The three adversarial networks employ an alternating optimization strategy to simulate the most challenging operating conditions that may be encountered in the real world. This enables the autonomous driving system to maintain stable and reliable decision-making performance even under multiple disturbances such as sensor anomalies and sudden environmental changes, fundamentally enhancing the system's robustness and safety.

[0052] Please see Figure 2 The present invention also provides an autonomous driving transport robot system, comprising: A multi-sensor module is used to collect environmental data; A processor configured to execute a deep learning-based autonomous decision-making method for an autonomous vehicle robot as described in any of the preceding items; The control execution module is used to perform motion control according to the control instructions generated by the processor.

[0053] The above are only some embodiments of this application and do not limit the patent scope of this application. All equivalent structural transformations made under the technical concept of this application and using the contents of the specification and drawings of this application, or direct / indirect applications in other related technical fields, are included in the patent protection scope of this application.

Claims

1. A deep learning-based autonomous decision-making method for autonomous driving transport robots, characterized in that, Includes the following steps: Environmental data is collected through multiple sensors, including lidar, cameras, and inertial measurement units. The environmental data is input into the autonomous decision-making model trained based on the curriculum adversarial self-evolutionary learning framework; The autonomous decision-making model performs a three-level progressive decision-making process, including: A global goal-oriented layer generates global path planning based on semantic maps and task objectives; The local contextual reasoning layer employs a multi-scale spatiotemporal attention mechanism to perceive and reason about the dynamic environment, generating local navigation strategies. The real-time motion control layer generates control commands based on an online optimization method for safe motion envelopes. The autonomous driving robot executes corresponding motion control according to the control instructions.

2. The autonomous decision-making method for an autonomous driving transport robot based on deep learning according to claim 1, characterized in that, The course adversarial self-evolutionary learning framework includes: During the course learning phase, autonomous decision-making models are trained in a simulation environment in order of increasing difficulty. The difficulty is dynamically adjusted based on environmental complexity, obstacle density, and task requirements. During the adversarial training phase, adversarial networks are used to generate challenging scenarios that can cause the autonomous decision-making model to fail. These challenging scenarios include suddenly appearing obstacles and aggressive driving behaviors. During the self-evolution phase, edge cases encountered in actual operation are fed back to the training data pool for incremental training of the model.

3. The autonomous decision-making method for an autonomous driving transport robot based on deep learning according to claim 1, characterized in that, The course learning phase employs a difficulty evaluation function based on multi-objective optimization: in, SR t Indicates the success rate. CR t Indicates the risk coefficient. UNC t This indicates uncertainty in decision-making. α, β, γ These represent the corresponding weight coefficients; when Difficulty t Increase training difficulty when the score is below the first threshold. Difficulty t Reduce training difficulty when the threshold is exceeded.

4. The autonomous decision-making method for an autonomous driving transport robot based on deep learning according to claim 1, characterized in that, The multi-scale spatiotemporal attention mechanism includes: Multi-sensor data is spatiotemporally aligned using a cross-modal spatiotemporal alignment module, which employs a hybrid architecture of deformable convolution and temporal calibration. By employing a hierarchical attention distillation mechanism to progressively compress multi-scale features, the computational complexity is reduced from O(N^2) to O(N^2). 2 The time complexity is reduced to O(NlogN); The attention intensity of different spatiotemporal features is dynamically adjusted by using a spatiotemporal attention weight matrix. The calculation formula for the spatiotemporal attention weight matrix is ​​as follows: Among them, M geometric Let T represent the geometric consistency matrix. delay This represents time delay compensation, σ represents the activation function, Q represents the query vector, K represents the key vector, and d represents the time delay compensation. k This represents the dimension of the key vector.

5. The autonomous decision-making method for an autonomous driving transport robot based on deep learning according to claim 1, characterized in that, The online optimization method for the safe motion envelope includes: Based on neural differential equations, learn the dynamic model of autonomous driving vehicles and predict the safe motion envelope in the future time period. The boundary of the safe motion envelope is dynamically optimized based on the real-time perceived environmental risks using an incremental convex optimization solver. The safety of the safety motion envelope is ensured through a safety verification and backoff mechanism, which triggers backoff control when a potential collision risk is detected.

6. The autonomous decision-making method for an autonomous driving transport robot based on deep learning according to claim 5, characterized in that, The online optimization process of the safe motion envelope is represented as a constrained optimization problem: min_E[λ1×Risk(E)+λ2×(1 / Utility(E))]; Constraints: E Feasible_Trajectories; Where E represents the optimized safe motion envelope, Risk(E) represents the expected collision probability, Utility(E) represents the envelope volume, and λ1 and λ2 are weighting coefficients.

7. The autonomous decision-making method for an autonomous driving transport robot based on deep learning according to claim 1, characterized in that, The three-tiered progressive decision processing also includes a cross-level information coupling mechanism: The flow of information between decision-making levels is controlled through a cross-level attention gating mechanism, wherein the gating function is: g_i→j=σ(W_g·[h_i,h_j,c_context]+b_g); Where h_i and h_j represent the features of the source layer and the target layer, respectively, c_context represents the context information, and W_g and b_g are learnable parameters; Consistency constraint propagation transmits high-level constraints to the lower level in the form of dual variables.

8. The autonomous decision-making method for an autonomous driving transport robot based on deep learning according to claim 1, characterized in that, It also includes adaptive computing resource allocation strategies: Computational resources are dynamically allocated based on the scene risk level, which is calculated based on obstacle distance, motion state, and environmental complexity. The prediction time domain is dynamically adjusted based on the motion state, and the calculation formula for the prediction time domain is as follows: T_horizon=T_min+(T_max-T_min)×(v / v_max)×(d_obstacle / d_safe); Where v represents the current speed, T_min represents the minimum prediction time domain, T_max represents the maximum prediction time domain, v_max represents the maximum allowed speed, d_obstacle represents the distance to the nearest obstacle, and d_safe represents the safe distance.

9. The autonomous decision-making method for an autonomous driving transport robot based on deep learning according to claim 2, characterized in that, Its features are, The adversarial network in the adversarial training phase adopts a hierarchical adversarial training mechanism, including: An environmental dynamic adversarial network is used to generate abnormal behavior patterns of dynamic obstacles. Its objective function is: max(φ_e)E[s~p_env][log(1-D_policy(π_main(s)))]; Where φ_e represents the parameters of the dynamic adversarial network, D_policy is the policy discriminator, π_main is the policy of the master decision network, s represents the environment state, p_env represents the environment state distribution, and E represents the mathematical expectation. Sensor interference adversarial networks are used to simulate multi-sensor data failure and noise interference. Their interference modes include: LiDAR point cloud loss: Randomly discard point cloud data according to the loss rate ρ; Camera image occlusion: Generate a random occlusion area based on the occlusion area A_occlude; IMU data drift: Injecting random bias noise b_imu; Scene complexity adversarial networks are used to adjust the environment structure and task constraints. They learn scene topology relationships through graph neural networks to generate challenging scene configurations. G_challenge=GNN_θ_g(G_base⊕ΔG_adv); Where G_base=(V_base,E_base) is the base scene graph, V_base={v_i},i=1toN is the set of nodes, E_base={e_ij} is the set of edges, ΔG_adv=(ΔV,ΔE) is the adversarial graph structure perturbation, ⊕ represents the graph structure fusion operation, and GNN_θ_g represents the graph neural network with parameter θ_g; The hierarchical adversarial training mechanism employs an alternating optimization strategy, updating sequentially in each training round t: Parameters of the dynamic adversarial network: φ_e^(t+1)=φ_e^(t)+η_e× _φ_eJ_e(φ_e^(t)); Sensor interference adversarial network parameters: φ_s^(t+1)=φ_s^(t)+η_s× _φ_sJ_s(φ_s^(t)); Parameters of the adversarial network for scene complexity: θ_g^(t+1)=θ_g^(t)+η_g× _θ_gJ_g(θ_g^(t)); Where η_e, η_s, and η_g are the learning rates of each adversarial network, and J_e, J_s, and J_g are the optimization objective functions of each adversarial network. Represents the gradient operator; The hierarchical adversarial training process forms a multi-round game training with the master decision network until a Nash equilibrium is reached, which satisfies: E[J_main(π_main,{φ_e,φ_s,θ_g})]≤E[J_main(π_main,{φ_e,φ_s,θ_g})]; E[J_adv(π_main,{φ_e,φ_s,θ_g})]≤E[J_adv(π_main,{φ_e,φ_s,θ_g})]; Where π_main represents the optimal master decision network strategy, {φ_e,φ_s,θ_g} represents the optimal set of parameters for the adversarial network, and J_main and J_adv represent the overall objective functions of the master decision network and the adversarial network, respectively.

10. An autonomous driving transport robot system, characterized in that, include: A multi-sensor module is used to collect environmental data; A processor configured to execute a deep learning-based autonomous decision-making method for an autonomous driving transport robot, as described in any one of claims 1-9; The control execution module is used to perform motion control according to the control instructions generated by the processor.

Images