An intelligent day lily picking robot control system

Abstract

The application discloses a kind of intelligentization day lily picking robot control systems, including field picking data acquisition module, picking robot motion control module, sample library construction module, integrated state evolution model design module, adaptive picking strategy learning module, picking action execution module.The application belongs to the field of intelligent control, specifically refers to a kind of intelligentization day lily picking robot control system, and the present scheme constructs picking robot motion control module, and the speed control design is differentiated, dynamic, to avoid plant damage caused by pose deviation from source;Based on adaptive row order design, sample pairing calibration is carried out;Mature-damage, pose deviation guide, equipment-plant protection is used to build immediate reward, adapt to various picking uncertainty scenarios;With predicted immediate reward and agronomic risk value as dual trigger condition, cover picking robot control decision-making scenario;Further improve the control effect of day lily picking robot.

Description

Technical Field

[0001] This invention relates to the field of intelligent control, specifically to an intelligent control system for a daylily harvesting robot. Background Technology

[0002] The control system for daylily harvesting robots is a closed-loop intelligent control system that relies on sensing, decision-making, and execution hardware and supporting algorithms to achieve environmental perception, mature flower identification and positioning, harvesting action planning, and drive control during the daylily harvesting process. However, general daylily harvesting robot control systems suffer from problems such as inconsistent motion speed regulation, which can easily cause plant damage and harvesting failures; crude reward mechanisms that fail to guide optimal harvesting behavior; and poor control reliability. Furthermore, general daylily harvesting robot control systems often fail to capture the uncertainties of the harvesting process, resulting in numerous invalid harvests, high damage rates, and operational delays due to single decision failures, leading to poor control performance. Summary of the Invention

[0003] To address the above issues and overcome the shortcomings of existing technologies, this invention provides an intelligent daylily harvesting robot control system. Addressing the problems of general daylily harvesting robot control systems, such as indifferentiated motion speed regulation leading to plant damage and harvesting failure, and a crudely designed reward mechanism that fails to guide optimal harvesting behavior, resulting in poor control reliability, this solution constructs a motion control module for the harvesting robot. Through differentiated and dynamic speed control design, it avoids plant damage caused by pose deviations at the source. Based on an adaptive row order design, it performs sample pairing calibration. The deviation is decomposed into immediate reward deviation and state transition probability deviation, accurately locating the source of deviation between simulated and real samples, thereby achieving efficient sample selection. To ensure the reliability of the daylily harvesting robot's control, this solution addresses the common problems of conventional daylily harvesting robot control systems failing to capture the uncertainties of the harvesting process, resulting in numerous invalid harvests, high damage rates, and operational delays due to single-decision failures, leading to poor control performance. This solution employs a real-time reward system based on maturity-damage, posture deviation guidance, and equipment-plant protection, achieving a refined reward design with both agronomic and hardware constraints to adapt to various harvesting uncertainty scenarios. By using predicted real-time rewards and agronomic risk values ​​as dual triggering conditions, it ensures that re-decision covers all invalid and high-risk harvesting behaviors while avoiding meaningless blind re-decision, thus covering all control decision-making scenarios for the harvesting robot and ultimately improving the control performance of the daylily harvesting robot.

[0004] The technical solution adopted by the present invention is as follows: The present invention provides an intelligent daylily harvesting robot control system, including a field harvesting data acquisition module, a harvesting robot motion control module, a sample library construction module, an integrated state evolution model design module, an adaptive harvesting strategy learning module, and a harvesting action execution module;

[0005] The field harvesting data acquisition module obtains historical robot operation data, forming an observation space;

[0006] The motion control module of the harvesting robot determines the speed threshold for each motion direction through continuous control coefficients, so as to realize independent speed adjustment of the harvesting robot in the motion direction as needed.

[0007] The sample library construction module generates simulated samples by adaptively adjusting the order of picking rows based on the observation space, thereby constructing a high-quality sample library.

[0008] The integrated state evolution model design module uses a high-quality sample library as training data and maximum expected likelihood as the training objective to fit the transition rules of state-action-next state and immediate reward in the picking scenario, and constructs a probabilistic integrated state evolution model.

[0009] The adaptive harvesting strategy learning module integrates real-time observation data with the prediction results of the integrated state evolution model, aiming to maximize long-term cumulative rewards and achieve adaptive harvesting strategy learning.

[0010] The picking action execution module executes picking actions based on real-time observation status.

[0011] Furthermore, the field harvesting data acquisition module acquires historical robot operation data and performs preprocessing to ultimately form an observation space; the preprocessing includes missing value imputation, outlier handling, and standardization.

[0012] Furthermore, the motion control module of the harvesting robot specifically includes:

[0013] Calculate the speed threshold for each direction of motion, configure the rated maximum speed for each direction of motion, and use independent continuous control coefficients as speed regulation weights.

[0014] The continuous control coefficients are dynamically updated based on the joint pose deviations in each motion direction of the robot.

[0015] Furthermore, the sample library construction module specifically includes:

[0016] Adaptive row order design dynamically adjusts the number of action steps for a single picking row;

[0017] A standardized simulated sample library is generated by generating a real-simulation dual sample structure through row order calibration.

[0018] Set sample screening thresholds to filter simulated samples;

[0019] Bias decomposition; performing bias decomposition on Q-value deviation to quantify the impact of state transition probability differences on future value;

[0020] The immediate reward design includes a maturity-damage reward, a posture deviation guidance reward, and an equipment-plant protection reward. The maturity-damage reward only recognizes the optimal harvest when the fruit is at its best maturity and without damage, with the remaining rewards and penalties decreasing in value. The posture deviation guidance reward is quantified based on the direction of the maximum deviation. The equipment-plant protection reward is designed based on the core indicators of robot hardware protection and the clustering characteristics of daylilies.

[0021] The construction of a high-quality sample library involves merging simulated samples that meet the sample discrimination threshold with real samples to obtain a high-quality sample library.

[0022] Furthermore, the integrated state evolution model design module specifically includes:

[0023] An integrated state evolution model was designed to fit the continuous transition pattern of state-action-next state + immediate reward in the daylily picking scenario.

[0024] The training objectives for the model are constructed using the maximum expected likelihood criterion.

[0025] Furthermore, the adaptive harvesting strategy learning module specifically includes:

[0026] An end-to-end policy network is constructed, taking the observed state as input and outputting the distribution of picking action instructions and continuous control coefficients for each movement direction. A long-term cumulative reward is defined to reflect the overall value of the picking action. The policy network adopts a fully connected network structure, with the input layer dimension being the dimension of the observation space features. The output layer uses a Gaussian distribution for the action instruction distribution and a Sigmoid activation output for the continuous control coefficients.

[0027] Calculate long-term cumulative rewards to reflect the overall value from a series of picking actions;

[0028] Construct the editing target function to implement strategy updates.

[0029] Furthermore, the picking action execution module specifically includes:

[0030] Harvesting strategy generation: Combining the trained integrated state evolution model and adaptive harvesting strategy, the final harvesting decision is formed. The real-time perceived environmental observation state is input into the policy network, and the optimal harvesting action command and velocity threshold for each motion direction are output. At the same time, the observation state and action command are input into the state evolution model to predict the robot pose and harvesting execution effect at the next moment.

[0031] By incorporating advanced decision-making and design, early warning of agronomical risks during daylily harvesting is introduced to minimize the damage rate during daylily harvesting.

[0032] The end effector executes the action, converting the optimal picking action command and optimal speed threshold output by the strategy network into electrical signal control values ​​for the robot drive module, ultimately enabling the harvesting of daylilies.

[0033] The beneficial effects achieved by the present invention using the above solution are as follows:

[0034] (1) In view of the problems that the control system of general daylily picking robot has no differentiation in motion speed regulation, which can easily cause plant damage and picking failure, and the reward mechanism is crudely designed and does not guide the optimal picking behavior in a targeted manner, thus resulting in poor control reliability, this solution constructs a motion control module for picking robot. Through differentiated and dynamic speed control design, plant damage caused by posture deviation is avoided from the source; based on the adaptive row order design, sample pairing calibration is performed; the deviation is decomposed into immediate reward deviation and state transition probability deviation, and the source of deviation between simulated sample and real sample is accurately located; thus, efficient sample identification is completed to ensure the final control reliability of daylily picking robot.

[0035] (2) In view of the problems that the control system of general daylily picking robot cannot capture the uncertainty of the picking process, resulting in many invalid pickings, high damage rate, and easy operation jamming due to single decision failure, thus leading to poor control effect, this solution adopts the maturity-damage, posture deviation guidance, and equipment-plant protection to construct an instant reward, realize the refined reward design of agronomic + hardware dual constraints, and adapt to various picking uncertainty scenarios; with the prediction of instant reward and agronomic risk value as dual trigger conditions, it ensures that the re-decision covers all invalid and high-risk picking behaviors, and avoids meaningless blind re-decision, covering the control decision scenarios of the picking robot; thereby improving the control effect of daylily picking robot. Attached Figure Description

[0036] Figure 1 The present invention provides an operational flowchart of an intelligent daylily harvesting robot control system.

[0037] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used together with the embodiments of the invention to explain the invention and do not constitute a limitation thereof. Detailed Implementation

[0038] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0039] In the description of this invention, it should be understood that the terms "upper", "lower", "front", "rear", "left", "right", "top", "bottom", "inner", "outer", etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings. They are only for the convenience of describing this invention and simplifying the description, and do not indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation. Therefore, they should not be construed as limitations on this invention.

[0040] Example 1, see Figure 1 The present invention provides an intelligent daylily harvesting robot control system, including a field harvesting data acquisition module, a harvesting robot motion control module, a sample library construction module, an integrated state evolution model design module, an adaptive harvesting strategy learning module, and a harvesting action execution module;

[0041] The field harvesting data acquisition module obtains historical robot operation data to form an observation space; and sends the data to the harvesting robot motion control module.

[0042] The harvesting robot motion control module determines the speed threshold for each motion direction through continuous control coefficients, enabling independent on-demand speed adjustment of the harvesting robot's motion direction; and sends the data to the sample library construction module.

[0043] The sample library construction module generates simulated samples based on the observation space by adaptively adjusting the order of the picking rows, thus constructing a high-quality sample library; and sends the data to the integrated state evolution model design module.

[0044] The integrated state evolution model design module uses a high-quality sample library as training data and maximum expected likelihood as the training objective to fit the transition rules of state-action-next state and immediate reward in the picking scenario, and constructs a probabilistic integrated state evolution model; and sends the data to the adaptive picking strategy learning module.

[0045] The adaptive harvesting strategy learning module integrates real-time observation data with the prediction results of the integrated state evolution model, aiming to maximize long-term cumulative rewards, and achieves adaptive harvesting strategy learning; and sends the data to the harvesting action execution module;

[0046] The picking action execution module executes picking actions based on real-time observation status.

[0047] Example 2, see Figure 1 This embodiment is based on the above embodiment. The field harvesting data acquisition module obtains historical robot operation data, performs preprocessing, and finally constitutes the observation space O, represented as: Preprocessing includes missing value imputation (spline interpolation), outlier handling (3σ criterion), and standardization (min-max normalization); among which, It is a characteristic vector of daylily plant status, including flower maturity, flower stem posture, and distance between flower and leaf; It is the real-time pose vector of the robot's end effector, including three-dimensional coordinates and three-axis angles; It is the velocity vector of the robot joints and end effector. These are environmental disturbance characteristics, including light intensity and plant density; the data acquisition frequency is 10Hz, and the historical harvested real sample database is denoted as [missing information]. The single sample format is ,in, The picking action command is at time t. The reward for the action at time t; and These are the observation states at time t and time t+1, respectively.

[0048] Example 3, see Figure 1 This embodiment is based on the above embodiment. The motion control module of the harvesting robot independently determines the speed threshold for the movement direction of each joint of the robot and the harvesting direction of the end effector. Combined with the real-time sensing of the plant status and posture deviation, the speed value is dynamically adjusted to achieve on-demand speed adjustment; specifically including:

[0049] Calculate the speed threshold for each direction of motion, configure the rated maximum speed for each direction of motion, and use independent continuous control coefficients as speed regulation weights, expressed as follows: ;in, It is the velocity threshold for the robot's u-th motion direction (joint movement / end-effector picking, pose direction); It is the continuous control coefficient for the u-th motion direction, with a value of 0 to 1. It is output independently and reflects the speed regulation intensity in the corresponding direction. The smaller the value, the stricter the speed limit. It is the rated maximum speed in the u-th direction of motion, which is determined by the robot's hardware performance;

[0050] The continuous control coefficients are dynamically updated based on the joint pose deviations in each motion direction of the robot, and are expressed as follows: ;in, It is the real-time deviation value corresponding to the u-th motion direction, representing the joint pose deviation; This is the maximum permissible deviation in the direction of motion, with an angle range of 0.5 to 2 degrees (normalized). If this deviation is exceeded, the robot will stop moving.

[0051] Example 4, see Figure 1This embodiment is based on the above embodiment. The sample library construction module generates simulated harvesting samples based on the observation space using a state evolution model, and accurately identifies the simulated samples through a Q-value deviation discrimination threshold. Finally, qualified simulated samples are fused with real samples to construct a high-quality sample library; specifically including:

[0052] The adaptive row order design dynamically adjusts the number of action steps for a single picking row to adapt to the accuracy requirements of different stages of picking training, avoiding the accumulation of sample bias caused by a fixed number of steps. This is expressed as: ;in, is the row order order in the e-th training round, which is the number of picking action steps contained in a single row order; a and b are the minimum and maximum training rounds for adjusting the row order order, with values ​​of 100~300 and 800~1200 respectively; x and y are the minimum and maximum picking row order orders, with values ​​of 8~15 and 40~60 respectively. The minimum order ensures the accuracy of basic training, and the maximum order adapts to the complete picking action sequence;

[0053] To ensure a precise match between the simulated and real picking row order in the state-action dimension, a real-simulation dual sample structure is generated through row order calibration. The calibrated sample is represented as follows: ;in, and These are the observed states at steps k+m and k+m+1 in the actual picking sequence, respectively. It is the picking action instruction at step k+m; This simulates the observation state at step k+m in the picking sequence; m is the row sequence calibration step size; k is the random initial state index; through pairing, the simulated samples are highly matched with real samples in action sequences and state transition paths to generate a standardized simulated sample library, denoted as . ;

[0054] By setting a sample screening threshold, high-quality simulated samples with high consistency with the value of real samples are selected, while samples with excessive deviation are removed to optimize the quality of the sample library. This can be represented as: ;in, and These are real samples and simulated samples The Q-value is represented as the long-term value of the state-action sequence. It is the Q-value deviation discrimination threshold, with a value ranging from 0.05 to 0.2;

[0055] Bias decomposition; Bias decomposition is performed on the Q-value deviation to quantify the impact of state transition probability differences on future value, improving the interpretability of the identification, expressed as: ;in, It is the immediate reward bias between real and simulated samples; and These are the immediate rewards for state-action interactions in real and simulated scenarios, respectively. It is a discount factor, ranging from 0.9 to 1, reflecting the degree of influence of future rewards on the current Q value; and These are state-action scenarios in real and simulated scenarios, respectively. , Transition to the next state The probability of; The next state The corresponding optimal picking action instruction is obtained from the action output of the pre-trained policy network.

[0056] By performing the above operations, this solution addresses the problems of general daylily harvesting robot control systems, such as inconsistent motion speed regulation, which easily leads to plant damage and harvesting failure, and a crudely designed reward mechanism that fails to guide optimal harvesting behavior, resulting in poor control reliability. This solution constructs a motion control module for the harvesting robot. Through differentiated and dynamic speed control design, it avoids plant damage caused by pose deviations at the source. Based on an adaptive row order design, it performs sample pairing calibration. Deviations are decomposed into immediate reward deviations and state transition probability deviations to accurately locate the source of deviation between simulated and real samples. This enables efficient sample identification, ensuring the final control reliability of the daylily harvesting robot.

[0057] Example 5, see Figure 1 This embodiment is based on the above embodiment, and the sample library construction module further includes:

[0058] The immediate reward design includes ripening-damage rewards, posture deviation guidance rewards, and equipment-plant protection rewards, which are weighted and summed with a total weight coefficient of 1; the ripening-damage reward... Only harvesting at optimal ripeness and without damage is considered the best outcome; all other harvests are rewarded or penalized according to decreasing value, as shown below: Maturity is identified by the robot's vision module, while damage is jointly detected by the vision and force perception modules; pose deviation is used to guide rewards. The direction of maximum deviation is quantized (the robot harvesting only needs to focus on the critical direction of the maximum deviation), and is represented as: ; It is the maximum pose deviation among all directions of robot motion; yes The maximum permissible deviation in the corresponding direction; when the deviation is 0. Highest reward; Equipment-Plant Protection Reward Designed based on the core performance indicators of robot hardware protection and the clustering characteristics of daylilies, it is expressed as follows: ; It is the clamping force of the end effector; 6N is the safe clamping force threshold for harvesting daylilies; It is the robot's maximum speed. It is the optimal speed threshold;

[0059] The construction of a high-quality sample library involves merging simulated samples that meet the sample discrimination threshold with real samples to obtain a high-quality sample library. The acquisition of Q-values ​​adopts a two-step logic of "pre-trained policy network solution - sample discrimination - formal training optimization". First, basic pre-training is performed through the adaptive picking policy learning module to obtain a Q-value calculation network that can be used for sample discrimination. Then, the module is formally trained using the discriminated high-quality sample library to achieve accurate optimization of the policy network. The pre-trained Q-value calculation network is only used for sample discrimination.

[0060] Example 6, see Figure 1 This embodiment, based on the above embodiment, integrates the state evolution model design module to construct an integrated state evolution model. Using a high-quality sample library D as training data and maximum expected likelihood as the training objective, the model accurately fits the transition pattern of observation state-action-next state in the daylily harvesting scenario. Simultaneously, by integrating sub-models, it captures cognitive and accidental uncertainties (random biases caused by environmental disturbances and plant shaking) during the harvesting process, improving the model's fitting ability and robustness to the harvesting scenario. Specifically, it includes:

[0061] An integrated state evolution model is designed to fit the continuous transition pattern of state-action-next state + immediate reward in the daylily picking scenario. A single state evolution sub-model is represented as follows: ;in, It is a single integrated sub-model, where i is the sub-model index. These are model parameters; It is the model's prediction of the next moment's observation state and immediate reward; It represents the current observation status and harvesting action instructions; It follows a normal distribution, making it suitable for predicting continuous states during the harvesting process; It is the predicted mean of the sub-model, reflecting the core prediction result of the state-reward; The prediction covariance of the sub-model reflects the uncertainty of the prediction result. The larger the covariance, the higher the uncertainty. After the predicted observation state is output, it needs to be denormalized to restore it to physical dimensions for subsequent deviation judgment and state analysis of harvesting. The number of integrated sub-models ranges from 8 to 12. The uncertainty is effectively captured by training multiple sub-models independently. All sub-models use the same network structure and training objectives, only the initialization parameters are different.

[0062] The model training objective is constructed using the maximum expectation likelihood criterion, and is expressed as follows: ; It is the likelihood loss function of the model, and the objective is to maximize the function. High-quality sample library The expected value of the samples is calculated; the model is trained using the Adam optimizer with a learning rate of 1e. -4 ~5e -4 The training batch size is set to 32~64, and training continues until the loss function converges or the maximum number of training epochs (1000~2000 epochs) is reached. An early stopping strategy is used during training to avoid overfitting.

[0063] Example 7, see Figure 1 This embodiment is based on the above embodiment. The adaptive picking strategy learning module relies on the state prediction, immediate reward prediction, and uncertainty assessment results output by the integrated state evolution model design module. With maximizing the long-term cumulative reward of picking as the core learning objective, it uses a high-quality sample library D and real-time picking interaction data. Through a policy network, it autonomously learns the mapping pattern of observed state → action command → speed threshold, achieving an adaptive picking strategy without manually preset rules. The module adopts an Actor-Critic dual-network architecture. The Actor is the policy network, responsible for outputting the action command distribution and continuous control coefficients; the Critic is the value network, responsible for fitting the state-action Q-value. In the pre-training stage, only the Critic network is trained for sample identification. In the formal training stage, the Actor and Critic networks are jointly trained to achieve end-to-end optimization of the strategy. Specifically, it includes:

[0064] Construct an end-to-end policy network that takes the observed state as input and outputs the distribution of picking action instructions. and continuous control coefficients in each direction of motion The long-term cumulative reward is defined to reflect the overall value of the picking action. The policy network adopts a fully connected network structure, with the input layer dimension being the dimension of the observation space features, 3-4 hidden layers, 128-256 neurons per layer, and ReLU activation function. The output layer uses a Gaussian distribution for the action command distribution and a Sigmoid activation function for the continuous control coefficient, ensuring that the value is in the range of 0-1. The value network takes the concatenated vector of the observation state and action command as input and outputs the Q-value of the state-action. The network structure is consistent with the policy network, with a linear output layer and no activation function.

[0065] The long-term cumulative reward is calculated as follows: ; It is the long-term cumulative reward at time t, reflecting the overall value of a series of picking actions starting from time t; The reward is the immediate reward at time t+k. A successful harvest (flowers undamaged and of matching maturity) is a positive reward, while a failed harvest (missed harvest, damaged, or harvested immature flowers) is a negative reward. The reward is 0 when there is no action. The immediate reward needs to be standardized to map the reward value to the range [-1,1] to improve the stability of the strategy learning. T is the total number of steps in the harvesting sequence, ranging from 40 to 60.

[0066] Construct the editing target function to implement the strategy update, represented as: ; ;in, It is the editing objective function, and the objective is to maximize the function to implement the strategy update; It is the ratio of the action probability before and after the strategy update; It is the advantage function, which reflects the value advantage of the current action relative to the average action. It is the difference between the Q value of the value network and the state value. It is the editing coefficient, with a value ranging from 0.1 to 0.3, which limits the range of variation of the action probability ratio; These are the network parameters before the policy update; during the formal training phase, the policy network and value network are jointly optimized. The value network uses mean squared bias loss to fit the long-term cumulative reward, while the policy network uses a clipping objective function for optimization. Both networks share the Adam optimizer, and the learning rate is set to 3e. -4 ~1e -3 The training batch size is consistent with the state evolution model. An experience replay pool is introduced during training to store real-time picking interaction data, thus combining offline and online learning.

[0067] Example 8, see Figure 1 This embodiment is based on the above embodiment, and the picking action execution module specifically includes:

[0068] Harvesting strategy generation: Combining the trained ensemble state evolution model and adaptive harvesting strategy, the final harvesting decision is formed. Real-time environmental observations are input into the policy network, which outputs the optimal harvesting action command and velocity thresholds for each motion direction. Simultaneously, the observed state and action commands are input into the state evolution model to predict the robot's pose and harvesting performance at the next moment, pre-judging the rationality of the action execution. If the prediction results indicate excessive pose deviation or damage risk, the action and speed are adjusted in real-time through the policy network. Specifically, the predicted output of the harvesting action and speed parameters is represented as follows: ;in, It is the optimal picking action command output by the strategy network, including end effector opening and closing, pose adjustment, and joint movement; These are the optimal continuous control coefficients for each direction of motion; It is the optimal speed threshold for each direction of motion; after the speed threshold is output, it needs to be limited and the minimum running speed (5%~10% of the maximum speed) needs to be set to avoid the robot getting stuck when the speed is 0.

[0069] To mitigate harvesting damage caused by delayed re-decision (where predicted pose deviations do not lead to failure but are close to the damage threshold for flower stem fragility), a re-decision design is implemented to provide early warning of agronomical risks during daylily harvesting, minimizing the damage rate. This also enhances the targeted nature of feedback corrections and reduces ineffective re-decision. Specifically, the design first uses an integrated state evolution model to predict the next state and immediate reward, represented as follows: ;in, It is the optimal observation state predicted by the ensemble dynamics model for the next moment; B is the immediate reward predicted by the kinetic ensemble model; B is the number of ensemble sub-models; based on the optimal observation state at the next time step after inverse normalization, agronomic characteristics are obtained and risk values ​​are calculated. The probability of harvesting damage is quantified as follows: ;in, It predicts agronomic characteristics; The threshold value for safe agronomical characteristics in daylily harvesting is selected based on the minimum safe distance between the flower stem and the end effector; ρ(·) is the cosine similarity; a double-trigger condition for re-decision is designed, where the condition is triggered if either condition is met, expressed as: ; This is the agronomic risk threshold, ranging from 0.7 to 0.85; meeting this condition triggers the strategy network to re-determine and adjust its approach. and Continue until the conditions are no longer met; set a maximum number of re-decision attempts (3-5 times). If the conditions still cannot be met after exceeding the number of attempts, abandon the harvesting of the current plant to avoid operational delays.

[0070] The end effector performs the action, and and Converted into electrical signal control values ​​for the robot drive module , It is the electrical signal conversion coefficient; the conversion coefficient is set separately according to the hardware characteristics of each drive joint of the robot, and the end effector is set to 0.01~0.05; after the action is completed, the complete sample of this harvest is added to the experience playback pool to realize the continuous updating of the model; all inputs need to be normalized and all outputs need to be denormalized; finally, the accurate and safe harvesting of daylilies is achieved.

[0071] By performing the above operations, this solution addresses the problems of general daylily harvesting robot control systems failing to capture the uncertainties of the harvesting process, resulting in numerous invalid harvests, high damage rates, and operational delays due to single-decision failures, leading to poor control performance. Instead, it employs a mature-damage, pose deviation-guided, and equipment-plant protection-based instant reward system to achieve a refined reward design with both agronomic and hardware constraints, adapting to various harvesting uncertainty scenarios. By using predicted instant rewards and agronomic risk values ​​as dual triggering conditions, it ensures that re-decision covers all invalid and high-risk harvesting behaviors while avoiding meaningless blind re-decision, thus covering all control decision scenarios for the harvesting robot and improving its control performance.

[0072] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention.

[0073] The present invention and its embodiments have been described above. This description is not restrictive, and the accompanying drawings are only one embodiment of the present invention; the actual structure is not limited thereto. In conclusion, if those skilled in the art are inspired by this description and design similar structures and embodiments without departing from the spirit of the invention, such designs should fall within the protection scope of the present invention.

Claims

1. An intelligent lily harvesting robot control system, characterized in that: The system includes a field harvesting data acquisition module, a harvesting robot motion control module, a sample library construction module, an integrated state evolution model design module, an adaptive harvesting strategy learning module, and a harvesting action execution module; The field harvesting data acquisition module obtains historical robot operation data, forming an observation space; The motion control module of the harvesting robot determines the speed threshold for each motion direction through continuous control coefficients, so as to realize independent speed adjustment of the harvesting robot in the motion direction as needed. The sample library construction module generates simulated samples by adaptively adjusting the order of picking rows based on the observation space, thereby constructing a high-quality sample library. The integrated state evolution model design module uses a high-quality sample library as training data and maximum expected likelihood as the training objective to fit the transition rules of state-action-next state and immediate reward in the picking scenario, and constructs a probabilistic integrated state evolution model. The adaptive harvesting strategy learning module integrates real-time observation data with the prediction results of the integrated state evolution model, aiming to maximize long-term cumulative rewards and achieve adaptive harvesting strategy learning. The picking action execution module executes picking actions based on real-time observation status; The motion control module for the harvesting robot specifically includes: Calculate the speed threshold for each direction of motion, configure the rated maximum speed for each direction of motion, and use independent continuous control coefficients as speed regulation weights. The continuous control coefficients are dynamically updated based on the joint pose deviations in each motion direction of the robot. The sample library construction module specifically includes: Adaptive row order design dynamically adjusts the number of action steps for a single picking row; A standardized simulated sample library is generated by generating a real-simulation dual sample structure through row order calibration. Set sample screening thresholds to filter simulated samples; Bias decomposition; The bias decomposition of Q-value deviation quantifies the impact of state transition probability differences on future value, expressed as: ;in, It is the immediate reward bias between real and simulated samples; and These are the immediate rewards for state-action interactions in real and simulated scenarios, respectively. It is a discount factor, ranging from 0.9 to 1, reflecting the degree of influence of future rewards on the current Q value; and These are state-action scenarios in real and simulated scenarios, respectively. , Transition to the next state The probability of; The next state The corresponding optimal picking action instruction is obtained from the action output of the pre-trained policy network; and These are real samples and simulated samples The Q value; the bias is decomposed into immediate reward bias and state transition probability bias, so as to accurately locate the source of the bias between simulated samples and real samples; Instant reward design; High-quality sample library construction involves merging simulated samples that meet the sample discrimination threshold with real samples to obtain a high-quality sample library. The instant reward design includes a maturity-damage reward, a posture deviation guidance reward, and a device-plant protection reward. The maturity-damage reward only recognizes optimal maturity plus no damage as the best harvest, with the rest being rewarded or penalized according to decreasing value. The posture deviation guidance reward is quantified based on the direction of the maximum deviation. The device-plant protection reward is designed based on the core indicators of robot hardware protection and the clustering characteristics of daylilies.

2. The intelligent daylily harvesting robot control system according to claim 1, characterized in that: The integrated state evolution model design module specifically includes: An integrated state evolution model was designed to fit the continuous transition pattern of state-action-next state + immediate reward in the daylily picking scenario. The training objectives for the model are constructed using the maximum expected likelihood criterion.

3. The intelligent daylily harvesting robot control system according to claim 2, characterized in that: The adaptive harvesting strategy learning module specifically includes: An end-to-end policy network is constructed, taking the observed state as input and outputting the distribution of picking action instructions and continuous control coefficients for each movement direction. A long-term cumulative reward is defined to reflect the overall value of the picking action. The policy network adopts a fully connected network structure, with the input layer dimension being the dimension of the observation space features. The output layer uses a Gaussian distribution for the action instruction distribution and a Sigmoid activation output for the continuous control coefficients. Calculate long-term cumulative rewards to reflect the overall value from a series of picking actions; Construct the editing target function to implement strategy updates.

4. The intelligent daylily harvesting robot control system according to claim 3, characterized in that: The harvesting action execution module specifically includes: Harvesting strategy generation: Combining the trained integrated state evolution model and adaptive harvesting strategy, the final harvesting decision is formed. The real-time perceived environmental observation state is input into the policy network, and the optimal harvesting action command and velocity threshold for each motion direction are output. At the same time, the observation state and action command are input into the state evolution model to predict the robot pose and harvesting execution effect at the next moment. By incorporating advanced decision-making and design, early warning of agronomical risks during daylily harvesting is introduced to minimize the damage rate during daylily harvesting. The end effector executes the action, converting the optimal picking action command and optimal speed threshold output by the strategy network into electrical signal control values ​​for the robot drive module, ultimately enabling the harvesting of daylilies.

5. The intelligent daylily harvesting robot control system according to claim 4, characterized in that: The field harvesting data acquisition module obtains historical robot operation data and performs preprocessing to ultimately form an observation space; the preprocessing includes missing value imputation, outlier handling, and standardization.

Images