Golf caddie robot behavior decision-making method based on language interaction

Abstract

The invention relates to the technical field of artificial intelligence and robots, in particular to a golf caddie robot behavior decision-making method based on language interaction, which is applied to a golf caddie robot with a mobile chassis, a loading / ball bag mechanism and a voice interaction and sensor, and comprises the following steps: voice acquisition and anti-noise preprocessing; performing golf semantic understanding and intention parameterization; estimating a context state; performing behavior decision based on a behavior state machine; performing safe interlocking and action execution; the binding record and the strategy are updated; through a built-in domain semantic dictionary, synonymous mapping and a context disambiguation mechanism, a natural language instruction is analyzed into a structured intention and parameters, the misunderstanding rate is greatly reduced, and the interaction efficiency is improved.

Description

Technical Field

[0001] This invention relates to the fields of artificial intelligence and robotics, specifically to a method for decision-making regarding the behavior of a golf caddy robot based on language interaction. Background Technology

[0002] With the expanding application of robotics technology in sports scenarios, golf caddy robots / caddie carts / following robots have gradually appeared on driving ranges and golf courses. Their common functions include carrying golf bags, remote control or automatic following, navigation and obstacle avoidance, and simple voice announcements. Meanwhile, voice interaction technology and large-scale model dialogue are relatively common in the consumer electronics field. However, golf courses, as unique outdoor environments, are characterized by high noise, multiple users, and multi-stage tasks, requiring robots to provide companionship services like a caddie. This necessitates that robots possess the following capabilities: understanding specific instructions in the context of golf (such as "next hole," "take the 7-iron," "stand further away," etc.), recognizing the hitting stage and scene rules (tee box / fairway / green / bunker, etc.), and making appropriate behavioral decisions under safe conditions (following, stopping, going around, passing the club, quiet mode, recording rounds, etc.).

[0003] Existing technologies for behavior control solutions in golf caddy-related equipment have significant defects and shortcomings: Voice commands lack golf semantic understanding and parameterization: Existing solutions often treat voice as a simple remote control switch, only responding to basic commands such as forward / stop / follow, and cannot parse golf intentions in natural language into structured parameters (such as hole number, club number, distance, stage preference, etc.), resulting in low interaction efficiency and high misunderstanding rate; Lack of behavioral decision-making based on "striking phase / scenario constraints": Golf has clear phase divisions (positioning, club selection, preparation, practice swing, hitting, etc.), and existing robots often ignore the differences between phases, which may cause interference or even safety risks when the user is preparing to hit the ball, by approaching, making noise or moving. Lack of authorization and anti-false triggering mechanisms under multi-user and multi-sound source conditions: The noisy stadium environment, such as the voices of others, wind noise, and ball cart noise, can easily lead to robot misidentification. The existing system lacks a combination mechanism of wake word, voiceprint authorization, near-field sound pickup, and context confirmation, resulting in frequent erroneous actions. Lack of "safety interlock + club / human approach" linkage strategy: When the robot is handing over the club, approaching the golf bag, and avoiding obstacles, it does not interlock control the club swing area, human approach distance, slope stability, etc., which poses a risk of collision, tripping, and interference with the swing. Lack of round binding and service loop: Most existing devices only stay at the level of mobility and carrying, and have not formed a closed loop of "voice command - behavior decision - service execution - result recording", so they cannot improve personalized services by learning user preferences; Incomplete handling of abnormal operating conditions: The existing system lacks explainable degradation and recovery strategies for situations such as lost target, GPS / positioning drift, path occupancy, sandpit / slope impact, and no-entry green.

[0004] Therefore, how to construct a caddie robot behavior decision-making method that is adapted to the special scenario of golf, can accurately understand semantics, make intelligent decisions based on scenarios and stages, and has high security and personalized service capabilities has become an urgent technical problem to be solved. Summary of the Invention

[0005] This invention addresses the shortcomings of the existing technology by providing a language-interactive behavior decision-making method for golf caddy robots, thus solving the problems of language interaction, behavior decision-making, safety control, and personalized service in existing golf caddy-related equipment.

[0006] The objective of this invention is achieved through the following technical solutions: A language-interactive-based behavior decision-making method for golf caddie robots, applied to caddie robots equipped with a mobile chassis, a cargo / bag carrying mechanism, voice interaction, and sensors, includes the following steps: Voice Acquisition and Noise Reduction Preprocessing S1: Acquire speech signals via a near-field microphone array or wearable microphone; S2: Perform noise reduction, echo cancellation, voice activity detection (VAD), and wind noise suppression to improve speech signal quality; S3: Recognize wake-up words, and upon wake-up, enter the instruction receiving window, receiving subsequent instructions only within the window; S4: Optionally perform voiceprint / device binding authorization to verify user legitimacy and obtain the authorization mark AUTH; Output: Text instruction T + authorization status AUTH + confidence level C_asr.

[0007] Golf semantic understanding and intent parameterization S5: Perform semantic parsing on the text instruction T, outputting the intent type I and the structured parameter set P; where the intent type I includes, but is not limited to: follow / stop / detour / maintain distance, pass / retrieve / recover / open bag, next hole / go to tee / go to ball / go to green, quiet mode / announcement mode / encouragement mode, record round / start training / review summary; the structured parameter set P includes, but is not limited to: club number (7 iron, pitching wedge, 60°, etc.), target hole number / target point (3rd tee, right fairway edge, etc.), distance and direction (2 meters away from me, stand to my right rear, etc.), stage preference (I am preparing to hit the ball, I am putting, etc.), user preference (quiet, don't get too close, etc.), parameters can be empty or partially defaulted; S6: If the voice confidence C_asr or semantic confidence C_sem is lower than the set threshold, trigger the clarification strategy, verify with the user with a short confirmation statement to avoid erroneous actions; output: (I, P, C_sem).

[0008] To address the issues of ambiguity, colloquialism, and ellipsis in natural language within a golf context, this step further introduces a golf semantic disambiguation and context correction mechanism: Golf semantic dictionary and synonym mapping: The system has a built-in domain semantic dictionary, which includes club synonym mapping (e.g., “7 iron / 7-iron / 7i” is mapped to Club=Iron7), action reference mapping (e.g., “hand it to me / take it” is mapped to Action=Deliver), spatial fuzzy reference mapping (e.g., “stand further away / don’t get too close” is mapped to Distance=Increase), and phase reference mapping (e.g., “I’m going to hit it / don’t make noise” is mapped to Phase=PreSwing), reducing ambiguity and misjudgment; Context-driven semantic disambiguation rules: When there are multiple candidate intentions or parameters in the parsing result, disambiguation is performed based on the context state X, including phase constraint disambiguation (e.g., when Phase = preparation / trial swing, "take the club" is preferentially parsed as "approach the currently used club"), spatial constraint disambiguation (e.g., when Zone = green, "come a little closer" is not parsed as approaching the putting line direction), and history preference disambiguation (preferentially adopting behaviors that have been selected multiple times in the user's historical rounds). Confidence-driven clarification interaction: If the confidence level is still low after semantic disambiguation, enter the minimum interference clarification mode. Only one clarification or silent abandonment is allowed during the ball preparation phase to avoid interfering with the swing.

[0009] Context state estimation S7: Infer the context state X using the robot's onboard sensors (vision sensors, IMU, GPS, distance sensors, etc.) and historical interaction data; X includes at least: Phase of the game: walking / club selection / preparation / practice swing / hitting / observation / moving to the landing point / putting, etc. Zones include: tee box, fairway, rough, bunker, green, no-entry zone, etc. User's relative position and movement status: distance, orientation, speed, and direction; Environmental risk status: slope, slipperiness, obstacle density, population density; Service load status: ball bag weight, remaining battery power, communication quality, etc.; The criteria for determining the stage include changes in user speed, pause time, club swing events (visual / IMU / sound feature recognition), user voice keywords, and location semantics (arrival at the tee / green). Output: X={Phase,Zone,Pose_user,Risk,...}.

[0010] Behavioral decision-making based on behavioral state machine S8: Construct a set of behavioral states B, which contains at least the following states: B1 Follow Walk: Follows the user at the set target distance and direction; B2 Side Rear Standby: Automatically retreats to the safe sector when the user is ready to hit the ball; B3 Quiet Mode: Disables voice broadcasting, prohibits approaching, and prohibits movement (or allows very low-speed fine-tuning). B4 Passing / Retrieving: Performing actions such as passing a club, opening a golf bag, and passing water; B5 Detour: Executed when the path is blocked or a restricted area is encountered; B6 Navigation to Target Point: Proceed to the designated waiting point near the tee / ball position; B7 Abnormal Degradation: Enters safe docking and requests user confirmation when the target is lost, the location drifts, or communication fails; S9: Trigger state transition based on intent type I, structured parameter set P, and context state X. The transition rules must include at least: When Phase = Ready / Practice Swing / Hit, force yourself into B2 / B3 (back off + quiet) and restrict your movement direction to avoid entering the swing sector; When I = pass and Zone = green, limit the driving speed and select low noise mode; When Zone = No Entry / High-Risk Slope, entry is prohibited, the nearest reachable waiting point is selected, and the user is prompted. When a user issues a command to "get farther / closer", the following distance parameter is updated in real time and written into the preference model; S10: Optional behavior scoring function Score(Bk|I,P,X) is introduced to calculate the score of each behavior state by comprehensively considering four indicators: safety, efficiency, interference, and energy consumption, and the behavior with the highest score is selected as the execution target. Output: Behavioral command Cmd (including target speed, target point, distance parameters, voice strategy, etc.).

[0011] Safety interlocks and action execution S11: Before executing the action instruction Cmd, perform multi-factor interlock checks to ensure action safety. Authorization interlock: AUTH must be true or near-field confirmation must be satisfied; otherwise, execution will be refused. Distance interlock: The distance between the detected cue and the human body must be greater than the safety threshold D_safe; Area interlocking: Adhere to scene rules such as no entry to the green / no entry to the bunker; Stable interlock: Limits speed or stops movement when the slope exceeds a threshold; Abnormal interlock: The machine will stop immediately when abnormalities such as chuck, collision, or sensor malfunction are detected; S12: If all interlock conditions are met, execute motion control and mechanism actions; if not, enter a safe state and prompt the user with an explainable reason (e.g., "The current area is a no-entry zone on the green, and you are waiting nearby").

[0012] To further enhance safety, this step introduces a dynamic safety zone generation and constraint method based on swing motion: Swing safety zone generation is based on: user body posture and orientation, current club type (long iron / short iron / wood / putter), swing phase (preparation / practice swing / downswing / follow-off), and historical swing amplitude and rhythm model (personalized parameters) to generate the swing safety zone Z_safe; Dynamic safety zone modeling method: adopt one or a combination of the following: fan-shaped area model (centered on the user's shoulder or ball position, changing with the club length), elliptical / polygonal model (describing the swing plane and follow-through space), and time-extended safety zone (expanding the radius of the no-entry zone in advance when predicting the downswing phase), and the safety zone is continuously updated over time. Linking safety zones with behavioral decisions: The robot's current position and target path must not intersect with Z_safe; when the user's posture changes rapidly (such as a sudden trial swing), the movement is immediately frozen and the robot retreats to the nearest safe point; during the ball-hitting phase, Z_safe has higher priority than all service behaviors; Anomaly and Degradation Handling: When the safe zone covers the entire reachable area or the sensor uncertainty is too high, enter the safe and still mode and notify the user in a low-interference manner (such as "I will wait here for you to finish playing").

[0013] Round binding records and strategy updates S13: When a "start / end of round" event is detected, generate a unique round IDR_i and bind the following data: Speech intent and parameters (I_i, P_i); Context state X_i (stage, region, risk, etc.); Execution behavior B_i and execution log; Optional: Shot result (ball trajectory / landing point) and motion analysis summary (if linked to the training system); S14: Based on continuous round data {R_i}, statistical analysis of user preferences (such as preference for silence when hitting the ball, standing 2.5m to the right rear) is performed, and decision strategy parameters (following distance, broadcast frequency, clarification threshold, etc.) are updated for the next behavior decision, thereby achieving personalized service optimization.

[0014] In summary, compared with the prior art, the present invention has the following advantages: 1. Accurate understanding of golf semantics: Through a built-in domain semantic dictionary, synonym mapping and context disambiguation mechanism, natural language instructions are parsed into structured intents and parameters, which greatly reduces the misunderstanding rate and improves interaction efficiency; 2. Intelligent adaptation of stages and scenarios: Based on the driving behavior state machine migration of the ball hitting stage and scenario area, it realizes "appropriate" services, such as automatically retreating and muting during the preparation / hitting stage, and moving slowly and quietly in the green area, which significantly reduces interference and safety risks. 3. High security and anti-false triggering capability: Through wake-up words, voiceprint authorization, near-field data acquisition, multi-factor safety interlocking and dynamic swing safety zone design, multiple protections are provided to avoid false actions and dangerous collisions, making it suitable for noisy and multi-user environments on golf courses; 4. Personalized service closed loop: By recording user behavior preferences through round-based binding, the strategy learns itself, and the service becomes more and more in line with user habits the more it is used, forming a platform-level barrier of "accompaniment + training".

[0015] 5. Strong interpretability and robustness: It has clear degradation strategies and reason prompts in abnormal working conditions, so that users can clearly understand the robot's behavior logic and improve the user experience; 6. Multi-scenario adaptation: The training ground, court, and simulator can all make consistent behavioral decisions based on semantics and context, making it highly versatile. Attached Figure Description

[0016] Figure 1 This is a diagram illustrating the overall framework of the method of this invention; Figure 2 This is a flowchart illustrating the behavioral decision-making process of the present invention. Figure 3 This is a diagram illustrating the stance and silencing strategy driven by the striking phase of this invention. Detailed Implementation

[0017] The technical solution of the present invention will be further described in detail below through specific embodiments, but the present invention is not limited to these embodiments.

[0018] Example 1: Automatic retreat and silence before hitting the ball Interaction scenario: A user is preparing to hit the ball in the fairway area and gives the robot the command: "I'm ready to hit, stand back and don't make a sound." Voice acquisition and preprocessing: The robot acquires voice signals through a near-field microphone array. After noise reduction and wind noise suppression, it recognizes the wake word and confirms the user's voiceprint authorization (AUTH=true), and outputs the text command T=“I’m about to call, stand back and don’t make noise”, with confidence C_asr=0.92; Semantic understanding and parameterization: The parsed intent type I = keep distance + quiet mode is obtained, the structured parameter P = {distance = 3m}, and the semantic confidence C_sem = 0.95 (above the threshold), requiring no clarification; Context state estimation: By detecting the user's pause time, posture (standing grip) and position (fairway area) through sensors, we infer that Phase = ready, Zone = fairway, and the user's relative position: 1.5m away, directly in front; Behavioral decision: Based on the state transition triggered by (I,P,X), enter B3 quiet mode + B2 side rear standby state, plan a safe sector 3m to the right rear of the user as the waiting point, and calculate the highest safety score for this behavior using the behavior scoring function. Safety Interlock and Execution: Verify that the authorization is valid (AUTH=true), the distance threshold is met, there are no restrictions on entry into the fairway area, the slope is stable, the interlock is passed, execute the movement to the target waiting point, and turn off the voice broadcast; Round binding and update: After the ball-hitting round ends, generate R_i and record "Phase=Ready, user prefers 3m distance + quiet mode". This strategy will be automatically triggered when the user enters the ready phase later.

[0019] Example 2: Voice relay and restricted area constraints Interaction scenario: A user needs to change clubs near the green and issues the instruction: "Give me the 7-iron"; Voice acquisition and preprocessing: The wearable microphone acquires the signal, which is then preprocessed to identify the wake word. Once authorization is granted (AUTH=true), the output T = "Hand me the No. 7 iron", C_asr = 0.94; Semantic understanding and parameterization: The parsing yields I = pass rod, P = {rod = 7 iron}, C_sem = 0.96, no further clarification is needed; Context state estimation: Inferring Zone = the area around the green, Phase = pole selection, and environmental risk state = low slope, no crowds through GPS and visual recognition; Behavioral decision: According to the migration rule, I = pass and Zone = green, trigger the B4 pass state, limit the travel speed to ≤0.3m / s, plan a path that does not cross the putting line, and the target point is the reachable area at the edge of the green; Safety Interlock and Execution: Verify that the authorization is valid, the distance to the user is 2m (meeting the safety threshold), the user has not entered the restricted area of ​​the green, and the slope is stable. Once the interlock is successful, move slowly to the target point and prompt the user: "7-iron is ready, please take the club." Round binding and update: Record the round "Zone = parrying around the green, user has no additional distance preference", and update the preference model.

[0020] Example 3: Degradation Handling of Abnormal Operating Conditions (No Entry to the Green) Interaction scenario: A user mistakenly instructs the robot to enter the green, issuing the command: "Come and pass me a ball." Voice acquisition and preprocessing: Authorization approved, output T = "Come and pass me a ball", C_asr = 0.91; Semantic understanding and parameterization: Parsing yields I = delivery object, P = {item = ball}, C_sem = 0.93; Context state estimation: Inferring Zone = green (no entry area), Phase = putting, environmental risk state = low slope; Behavioral decision: According to the migration rules, Zone = restricted area, triggering the B5 detour avoidance state, and selecting the nearest reachable waiting point on the green edge; Safety Interlock and Execution: If the area interlock judgment is not met (green restricted), enter the safety standby state and prompt the user: "The current area is a green restricted zone. We are waiting at the green. Please come and retrieve your ball." Round binding and update: Record the handling result of this abnormal working condition. If the user issues an instruction to enter the restricted area again in the future, the same avoidance strategy will be directly adopted and the prompts will be simplified.

[0021] Example 4: Preference Learning and Automatic Adaptation Initial interaction: During the user's first three rounds of hitting the ball, the robot issued the command "Stand 2 meters behind and to my left" during the preparation phase. The robot executed this action each time and recorded the round data. Preference learning: Based on data statistics from 3 consecutive rounds, the system identifies the user's fixed preferences: Phase = ready, following direction = left rear, following distance = 2m, and silent mode; Automatic adaptation: When the user enters the preparation phase for the fourth time (without issuing distance and orientation commands), the robot recognizes Phase = ready through context state estimation, and combines it with the preference model to automatically trigger the B2 side rear standby state, move to a position 2m to the left rear of the user and enter quiet mode, without requiring the user to repeat commands.

[0022] The embodiments of the present invention are not limited to those described above. Without departing from the spirit and scope of the present invention, those skilled in the art can make various changes and improvements to the present invention in form and detail, and these are all considered to fall within the protection scope of the present invention.

Claims

1. A language-interactive-based behavior decision-making method for a golf caddie robot, applied to a caddie robot with a mobile chassis, a cargo / bag mechanism, voice interaction, and sensors, characterized in that... Includes the following steps: S1. Voice Acquisition and Noise Reduction Preprocessing: Acquires voice signals through a near-field microphone array or wearable microphone. After noise reduction, echo cancellation, voice activity detection, and wind noise suppression, the wake-up word is recognized and the command receiving window is entered. Optional voiceprint / device binding authorization can be performed, and the text command T, authorization status AUTH, and confidence level C_asr are output. S2. Golf Semantic Understanding and Intent Parameterization: Perform semantic parsing on the text instruction T, outputting the intent type I and the structured parameter set P. If the confidence level is lower than the threshold, a clarification strategy is triggered. The semantic parsing process introduces golf semantic disambiguation and context correction mechanisms. S3. Context state estimation: Infer the context state X through sensor and historical interaction, wherein X includes at least the phase of hitting the ball, the scene area zone, the user's relative position and motion state, the environmental risk state and the service load state; S4. Behavior decision based on behavior state machine: Construct a set of behavior states B, trigger state transitions based on intent type I, structured parameter set P and context state X, optionally introduce a behavior scoring function to select the optimal behavior, and output behavior instruction Cmd; S5. Safety Interlock and Action Execution: Before executing the behavior instruction Cmd, a multi-factor interlock judgment is performed. If the interlock condition is met, the action is executed; otherwise, a safe state is entered and the reason is displayed. A dynamic safety zone generation mechanism based on the swing action is introduced. S6. Round Binding Records and Strategy Updates: Detect the start / end events of a round of play, generate a round ID and bind relevant data, and update decision-making strategy parameters based on user preferences statistics for consecutive rounds.

2. The method according to claim 1, characterized in that, The golf semantic disambiguation and context correction mechanism described in step 2 includes: Built-in semantic dictionary for the golf domain, including club synonym mapping, action reference mapping, spatial fuzzy reference mapping and stage reference mapping; Execution phase constraint disambiguation, spatial constraint disambiguation, and historical preference disambiguation are based on context state X; When the semantic confidence C_sem is below the threshold, the system enters the minimum interference clarification mode, allowing only one clarification or silent abandonment during the ball preparation phase.

3. The method according to claim 1, characterized in that, The striking phase described in step 3 includes walking, club selection, preparation, practice swing, striking, observation, heading to the landing point, and putting. The phase determination criteria include changes in user speed, pause time, club swing events, user voice keywords, and location semantics.

4. The method according to claim 1, characterized in that, The behavioral state set B mentioned in step 4 includes at least: following, waiting in the side and rear, quiet mode, passing / retrieving objects, detouring and avoiding, navigating to the target point, and abnormal degradation; The state transition rules include at least the following: When the Phase is ready / practice swing / hit, it forces the player into a side-back standby / quiet mode and restricts movement to avoid entering the swing sector; When I is the pass and Zone is the green, limit the driving speed and select low noise mode; When the Zone is restricted to entry or has a high-risk slope, entry is prohibited and the user is prompted to select the nearest reachable waiting point. When a user issues a distance adjustment command, the follow distance parameter is updated in real time and written into the preference model.

5. The method according to claim 1, characterized in that, The multi-factor interlocking judgment in step 5 includes authorized interlocking, distance interlocking, area interlocking, stable interlocking, and abnormal interlocking; The dynamic security zone generation mechanism includes: The swing safety zone Z_safe is generated based on the user's body posture and orientation, current club type, swing phase, and historical swing amplitude and rhythm model. Modeling is performed using one or a combination of sector-shaped region models, elliptical / polygonal models, and time-extended safe zones. Behavioral decisions must satisfy the following conditions: the robot's current position and the target path do not intersect with Z_safe; when the user's posture changes rapidly, the movement is immediately frozen and the robot retreats to the nearest safe point; during the ball-hitting phase, Z_safe has higher priority than all service behaviors.

6. The method according to claim 1, characterized in that, The relevant data bound in step 6 includes voice intent and parameters, context state, execution behavior and logs, and optionally includes hitting results and action analysis summaries; the preference model is updated based on continuous round data to achieve personalized optimization of service behavior.

7. The method according to claim 1, characterized in that, The intent type I mentioned in step 2 includes follow / stop / detour / maintain distance, pass / retrieve / collect / recover / open bag, next hole / go to tee / go to ball / go to green, quiet mode / announcement mode / encouragement mode, record round / start training / review summary; the structured parameter set P includes club number, target hole number / target point, distance and direction, stage preference, user preference, and parameters can be empty or partially defaulted.

8. The method according to claim 1, characterized in that, The abnormal interlock triggering conditions mentioned in step 5 include wheel jamming, collision, and sensor abnormality, which will immediately stop the machine after triggering; the behavior scoring function mentioned in step 4 comprehensively considers safety, efficiency, interference, and energy consumption indicators to select the behavior with the highest score.

Images