Large model long and short chain cooperative reasoning autonomous mobile platform control method

Abstract

The invention discloses a large-model long-chain and short-chain cooperative reasoning autonomous mobile platform control method. The method comprises the following steps: capturing environment dynamic data and user intention data in real time; preprocessing the environment dynamic data and the user intention data to obtain a smooth environment semantic graph; inputting the smooth environment semantic graph and the user intention data into a Transform large model to obtain a reasoning result; and the long and short thinking chain generates a decision result through the reasoning result. According to the large-model long-chain and short-chain collaborative reasoning autonomous mobile platform control method, environment dynamic data and user intention data are jointly expressed in unified coding, cross-modal understanding and decision consistency are improved, stable passing is guaranteed, decision optimization can be conducted in real time according to the environment dynamic data and the user intention data according to the actual situation, and the user experience is improved. And the lack of interpretability and self-adaptive capability of the control logic is avoided.

Description

Technical Field

[0001] This invention relates to the field of automatic control technology, and in particular to a control method for a large-scale model long-short chain collaborative reasoning autonomous mobile platform. Background Technology

[0002] Existing intelligent autonomous mobile platforms often rely on threshold triggering or local replanning strategies for obstacle avoidance and path planning, neglecting the deep integration of environmental semantic understanding and user operational intent. On the one hand, simple obstacle avoidance mechanisms are prone to "over-avoidance" or "obstacle avoidance failure" in complex road conditions, making it difficult to guarantee smooth passage. On the other hand, control logic that relies solely on fixed rules lacks interpretability and adaptability, making it difficult to continuously optimize decision-making performance over long-term use. Summary of the Invention

[0003] In view of this, the purpose of this invention is to provide a control method for an autonomous mobile platform with large-scale model long and short chain collaborative reasoning, which unifies the encoding of environmental dynamic data and user intent data into a joint representation, improves cross-modal understanding and decision consistency, ensures smooth passage, and can optimize decisions in real time based on the actual situation according to the environmental dynamic data and user intent data, thus avoiding the lack of interpretability and adaptability of the control logic.

[0004] This invention provides a control method for an autonomous mobile platform for large-scale model long-short chain collaborative inference, comprising the following steps:

[0005] Real-time capture of dynamic environmental data and user intent data;

[0006] The environmental dynamic data and the user intent data are preprocessed to obtain a smooth environmental semantic graph;

[0007] The smoothed environment semantic graph and the user intent data are input into the Transformer large model to obtain inference results;

[0008] The long and short thought chains generate decision results through the reasoning results.

[0009] In one embodiment, the wire feeding assembly includes a driving member, a transmission wheel, and a wire feeding wheel. The driving member is connected to the main body, and the output end of the driving member is connected to the transmission wheel. The wire feeding wheel is rotatably mounted on the main body, and the wire feeding wheel and the transmission wheel are positioned opposite each other, with a preset gap reserved between the wire feeding wheel and the transmission wheel.

[0010] In one embodiment, the real-time capture of environmental dynamic data and user intent data further includes:

[0011] Real-time acquisition of 3D point cloud data, depth images, inertial data, and voice data;

[0012] The voice data is converted into text format as the user intent data.

[0013] In one embodiment, the preprocessing of the environmental dynamic data and the user intent data to obtain a smoothed environmental semantic graph further includes:

[0014] Match the corresponding 3D point cloud data, depth image, and inertial data based on the timestamp of the user intent data;

[0015] The corresponding inertial data is filtered and bias drift is compensated to obtain preliminary inertial data;

[0016] Median filtering with a fixed window size is applied to the depth dimension of the corresponding 3D point cloud data and the pixel domain of the depth image to remove isolated noise points and abrupt pixels from the 3D point cloud data and the depth image, respectively.

[0017] The preliminary inertial data is smoothed using a Kalman filter to obtain the final inertial data;

[0018] The final inertial data is fed back into the 3D point cloud data and the depth image for matching and filtering to obtain a preliminary environmental semantic map;

[0019] The preliminary environmental semantic map is subjected to multimodal processing to obtain the smoothed environmental semantic map.

[0020] In one embodiment, the step of performing multimodal processing on the preliminary environmental semantic map to obtain the smoothed environmental semantic map further includes:

[0021] The initial environmental semantic map is divided into several image patches of fixed size;

[0022] The image patch is subjected to radial linear projection to obtain a sequence of visual feature vectors.

[0023] User intent data is segmented and embedded into word vectors, and then positional encoding is added to form a sequence of language feature vectors.

[0024] The visual feature vector sequence and the language feature vector sequence are concatenated in sequence, and separator markers are inserted and hierarchical normalization is performed to obtain the smooth environment semantic map.

[0025] In one embodiment, the long and short thought chains generate a decision result through the reasoning result, further including:

[0026] The inference cycle of the Transformer large model is divided into several short time windows, and the inference results are collected in each time window.

[0027] Based on the collected reasoning results, local reasoning prompts and global prompts are constructed;

[0028] Call the inference interface of the Transformer large model and output the local obstacle avoidance action sequence based on the local inference prompts;

[0029] The top k actions with the highest confidence in the local obstacle avoidance action sequence are retained;

[0030] By identifying major peer regions, planning key intermediate points leading to the destination, connecting key points to generate continuous paths, and generating multi-step inference chains, a smooth and executable global path is generated by applying Bézier curve fitting to the node sequence.

[0031] The confidence scores of the top k actions with the highest confidence scores in the local obstacle avoidance action sequence and the confidence scores of the actions in the global path are normalized to obtain normalized confidence scores.

[0032] The actions of the obstacle avoidance sequence are linearly combined with the key target points of the global path according to the normalized confidence to generate the decision result.

[0033] In one embodiment, the control method for a large-scale model long-short chain collaborative inference autonomous mobile platform further includes:

[0034] The gold standard for generating structured control commands;

[0035] By minimizing the KL divergence with the output, the decision result is mapped to a command template containing fields such as target coordinates, velocity, and obstacle avoidance priority.

[0036] In one embodiment, the control method for a large-scale model long-short chain collaborative inference autonomous mobile platform further includes:

[0037] Continuously record the semantic graph of the smooth environment, long and short thought chains, motion control commands and decisions;

[0038] Monitoring operational quality based on obstacle avoidance success rate and path deviation indicators;

[0039] When a decline in operational quality or a significant change in the environment is detected, background micro-batch incremental training is triggered to fine-tune the long and short thought chains online.

[0040] The finely tuned long and short thought chains are automatically deployed to the real-time inference engine to achieve closed-loop optimization of the strategy.

[0041] This invention provides a control method for an autonomous mobile platform with large-scale model long and short chain collaborative reasoning. It unifies the encoding of environmental dynamic data and user intent data into a joint representation, which improves cross-modal understanding and decision consistency, ensures smooth passage, and can optimize decisions in real time based on the actual situation according to the environmental dynamic data and user intent data, thus avoiding the lack of interpretability and adaptability of the control logic. Attached Figure Description

[0042] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0043] Figure 1 This is a flowchart illustrating a control method for an autonomous mobile platform with long and short chain collaborative inference in a large model, as provided by the present invention.

[0044] Figure 2 This is a schematic diagram of collaborative reasoning between long and short thought chains provided by the present invention.

[0045] Figure 3 The network architecture diagram for the teacher-student distillation provided by this invention is shown. Detailed Implementation

[0046] The specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings. Obviously, the described embodiments are merely some, not all, of the embodiments of the present invention. Based on the description of the present invention, all other embodiments obtained by those skilled in the art without inventive effort are within the scope of protection of the present invention.

[0047] In the description of this invention, unless otherwise explicitly specified and limited, the terms "set," "install," "connect," etc., should be interpreted broadly. For example, they can refer to a fixed connection, a detachable connection, or an integral connection; they can refer to a mechanical connection or an electrical connection; they can refer to a direct connection or an indirect connection through an intermediate medium. Those skilled in the art can understand the specific meaning of the above terms according to the specific circumstances.

[0048] The terms “upper,” “lower,” “left,” “right,” “front,” “back,” “top,” “bottom,” “inner,” and “outer,” etc., indicate the orientation or positional relationship based on the orientation or positional relationship shown in the accompanying drawings, or the orientation or positional relationship commonly used when the product of the invention is in use. They are only for the convenience of description and simplification, 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 the present invention.

[0049] The terms “first,” “second,” “third,” etc., are used merely to distinguish elements with similar properties, not to indicate or imply relative importance or a specific order.

[0050] The terms “include,” “comprising,” or any other variation thereof are intended to cover non-exclusive inclusion, which includes not only the elements listed but also other elements not expressly listed.

[0051] Please see Figure 1 The present invention provides a control method for an autonomous mobile platform for long-short chain collaborative inference in a large model, comprising the following steps:

[0052] S1 captures dynamic environmental data and user intent data in real time.

[0053] It is understandable that the above steps further include:

[0054] S101 acquires 3D point cloud data, depth images, inertial data, and voice data in real time;

[0055] S102, convert the voice data into text format as user intent data.

[0056] It is known that a 16-line rotating LiDAR is used to perform a 360° point cloud scan of the surrounding environment. The rotation frequency is 10Hz, and about 200,000 distance points are generated per frame, generating high-precision three-dimensional point cloud data in real time. The voice interaction module converts the user's voice commands into voice data through automatic speech recognition (ASR) and transmits them to the feature fusion engine.

[0057] S2 preprocesses environmental dynamic data and user intent data to obtain a smooth environmental semantic graph.

[0058] It is understandable that the above steps may further include:

[0059] S201, Match the corresponding 3D point cloud data, depth image and inertial data according to the timestamp of the user intent data;

[0060] S202, the corresponding inertial data is filtered and the bias drift is compensated to obtain preliminary inertial data;

[0061] S203, a median filter with a fixed window size is used in the depth dimension of the corresponding 3D point cloud data and the pixel domain of the depth image to remove isolated noise points and abrupt pixels in the 3D point cloud data and the depth image respectively.

[0062] S204: The preliminary inertial data is smoothed using a Kalman filter to obtain the final inertial data;

[0063] S205, the final inertial data is fed back to the 3D point cloud data and depth image for matching filtering to obtain a preliminary environmental semantic map;

[0064] S206. Perform multimodal processing on the preliminary environmental semantic map to obtain a smooth environmental semantic map.

[0065] It is known that the acquired 3D point cloud data undergoes real-time parameter calibration and distortion correction to ensure the accuracy and continuity of spatial coordinates. Specifically, the distortion correction utilizes the rotation rate provided by the IMU. and scan time Time compensation is performed for each point; the depth camera acquires depth images within the field of view, and the original depth map is first subjected to bilateral filtering. For missing values, nearest neighbor interpolation is used to supplement the sparse point cloud or blind spots caused by the LiDAR in corners, low-reflection surfaces, or multipath interference areas; the inertial measurement unit (IMU) outputs triaxial acceleration and triaxial angular velocity at a sampling frequency of over 200Hz. Combined with zero-bias estimation and attitude fusion algorithms, it provides the attitude angle and linear acceleration of the autonomous mobile platform in real time for dynamic motion state estimation and motion supplementation; the voice interaction module integrates a multi-microphone array and a noise suppression front end. It uses a DNN-HMM model trained with the Kaldi toolbox and utilizes Automatic Speech Recognition (ASR) technology to transcribe the user's natural language commands into plain text in real time. The ASR text is input into a BERT-based intent classifier to extract and further analyze the user's intent. A timestamp t and confidence classification p are added to each command to form a structured record. Finally, the instruction text, after being timestamped and semantically enhanced, is uniformly transmitted to the feature fusion engine, providing accurate and reliable high-level input for subsequent multimodal fusion and decision-making;

[0066] The system precisely aligns multiple data streams, including LiDAR, depth camera, IMU, and voice commands, in the time dimension using timestamps. It then uses Kalman filtering and smoothing median filtering to remove various sensor noises and sudden anomalies. The purified 3D point cloud data, depth map, and pose information are input into a convolutional neural network to generate a rasterized environment map with semantic labels such as "passable area," "static obstacle," and "dynamic pedestrian." This environment semantic map, along with the user command text, is then concatenated as a multimodal sequence and fed into the Transformer large model for joint encoding and inference.

[0067] Periodically use PTP to calibrate the device clock to eliminate clock skew. During alignment, nearest neighbor frame matching is used. For each instruction text, find the frame with the closest timestamp in the LiDAR point cloud, depth map, and IMU data, and form a unified time slice.

[0068] Noise and outlier filtering is performed to construct an IMU state-space model, and the triaxial acceleration and angular velocity sequences are filtered to remove high-frequency noise and compensate for bias drift. Median filtering with a fixed window size is applied to the depth dimension of the point cloud and the pixel domain of the depth map to remove isolated noise points and abrupt pixel changes. A cascaded strategy is adopted: the IMU data is first smoothed by Kalman filtering (KF), and then the filtered attitude information is fed back to the median filtering (MF) processing of the depth map and point cloud to ensure dual smoothness in both space and time.

[0069] The environmental semantic map generation process involves rasterization projection, where filtered 3D point cloud data is projected onto a ground plane to generate a height raster map. The height raster map, depth map, and pose-corrected camera view are then stitched together as a multi-channel input. Convolutional inference is used to input the data into a trained multi-layer convolutional neural network, which outputs the semantic category and corresponding confidence score for each raster. Post-processing applies a conditional random field to refine the boundaries, ultimately resulting in a smoothed environmental semantic map.

[0070] S3 inputs the smoothed environment semantic graph and user intent data into the Transformer large model to obtain inference results.

[0071] It is understandable that the above steps may further include:

[0072] S301, the initial environmental semantic map is divided into several image patches of fixed size;

[0073] S302, After performing radial linear projection on the image patch, a sequence of visual feature vectors is obtained;

[0074] S303, the user intent data is segmented and word vectors are embedded, and positional encoding is added to form a sequence of language feature vectors.

[0075] As we know, the multimodal input assembly visual feature embedding divides the environmental semantic map into fixed-size image patches, and obtains a visual feature vector sequence after radial linear projection; the text feature encoding performs word segmentation and word vector embedding on the instruction text, and adds positional encoding to form a language feature vector sequence; the visual and language features are concatenated in sequence, with necessary delimiters inserted and hierarchical normalization performed; the concatenated multimodal sequence is input into the Transformer large model to provide a unified and synchronous semantic representation for reasoning between long and short thought chains.

[0076] S4, the long and short thought chains generate decision results through reasoning results;

[0077] The above steps may further include:

[0078] S401 divides the inference cycle of the large Transformer model into several short time windows and collects inference results within each time window.

[0079] S402, Construct local reasoning prompts and global prompts based on the collected reasoning results;

[0080] S403 calls the inference interface of the Transformer large model and outputs the local obstacle avoidance action sequence based on the local inference prompts;

[0081] S404: The top k actions with the highest confidence in the local obstacle avoidance action sequence are retained;

[0082] S405 generates a multi-step inference chain by identifying major peer areas, planning intermediate key points leading to the destination, connecting key points to generate continuous paths, and applying Bézier curve fitting to the node sequence to generate a smooth and executable global path.

[0083] S406, normalize the confidence of the top k actions with the highest confidence in the local obstacle avoidance action sequence and the confidence of the actions in the global path to obtain normalized confidence.

[0084] S407 linearly combines the actions of the obstacle avoidance sequence with the key target points of the global path according to the normalized confidence level to generate the decision result.

[0085] Understandably, short-chain CoT units divide the inference cycle into several short time windows. :

[0086]

[0087] Obstacle state vector construction: In the k-th time window, collect the states of the most recent m obstacles.

[0088]

[0089] in, The relative coordinates of the obstacle, Its radial velocity.

[0090] Within each time window, the state of surrounding obstacles and a simple text command Ut are collected, and a vector is obtained through simple text embedding. Construct local inference clues:

[0091]

[0092] Call the Transformer inference interface:

[0093]

[0094] Each action Including steering angle and acceleration values, the generated multiple candidate actions are filtered according to safety constraints, retaining the top k actions with the highest confidence. For each candidate action... examine:

[0095]

[0096] Confidence ranking and Top-k retention: Transformer outputs each confidence level Select the first k actions in descending order:

[0097]

[0098] S302: The long thought chain unit reads the current environment semantic graph Gt and the destination semantic label Ld as global input. It generates a multi-step reasoning chain by identifying the main peer areas, planning intermediate key points leading to the destination, and connecting key points to generate continuous paths, thus constructing global prompt words.

[0099]

[0100] Perform nth-order Bézier curve fitting on the key point {si}:

[0101]

[0102] Generates smooth, executable global paths. Primarily designed for global planning and complex, long text processing.

[0103] S303: The confidence-weighted mechanism integrates the results of short and long thought chains, obtaining confidence scores cs and cl from the outputs of short and long thought chains respectively, performing normalization, and calculating:

[0104]

[0105] The short-chain recommendation action 'as' is linearly combined with the long-chain recommendation next global key target point 'gl' according to weights:

[0106]

[0107] in, This represents the action vector from the current position to the next global keypoint.

[0108] The fusion results were obtained After dynamic safety checks to ensure that kinematic and safety constraints are met, the data is sent to the control execution unit.

[0109] S5, the gold standard for generating structured control commands;

[0110] By minimizing the KL divergence with the output, the decision result is mapped to a command template containing fields such as target coordinates, velocity, and obstacle avoidance priority.

[0111] Understandably, the intent distillation module employs a typical "teacher-student" network architecture, such as... Figure 3 As shown. Teacher model inference and label generation, input: user instruction text U; teacher model: a large model fine-tuned for the usage scenario. Reasoning under temperature parameter T:

[0112]

[0113] in, Let logits be the logits of the i-th token in the command vocabulary for the teacher model.

[0114] Gold Standard Tag: Teacher Distribution As a "soft tag," and to record the highest-level general field to generate structured commands. .

[0115] A lightweight decoder is trained as a student model. Upon receiving natural language commands from the user, it learns to map the natural language into command templates containing fields such as target coordinates, velocity, and obstacle avoidance priority by minimizing the KL divergence between its own output distribution and the teacher model's output distribution. Student model structure and initialization; decoder architecture: lightweight Transformer decoder S; output distribution:

[0116]

[0117] Initial vocabulary list: a candidate set pre-set based on teaching data This includes common target coordinates, velocity ranges, and priority labels.

[0118] During the distillation process, not only are the network parameters of the student model optimized, but the vocabulary of command templates used are also dynamically adjusted to ensure that control commands can be generated efficiently and accurately during online inference.

[0119] Command template vocabulary optimization involves analyzing frequently used command fields and value ranges in student output during training iterations. The vocabulary is then updated by adding, deleting, or reordering entries, retaining the most commonly used command words and value ranges while identifying redundant and rarely used items. Dynamic vocabulary updates help reduce the decoder's output space and accelerate online inference.

[0120] For each training sample, distillation loss and potential auxiliary supervision losses, such as cross-entropy loss for certain key fields, are calculated. Backpropagation is used to simultaneously update the student model parameters and vocabulary mapping, ensuring optimal collaboration between the model and the vocabulary. After training, a lightweight student model is deployed for real-time instruction parsing. Due to the compact vocabulary and small model parameter count, inference latency is significantly reduced, while the output command template maintains accuracy similar to the teacher model. This "teacher-student" distillation process significantly improves online inference efficiency and resource utilization while ensuring command quality control. The total loss for a single sample U is defined as:

[0121]

[0122] KL divergence term (soft label distillation, ):

[0123]

[0124] Cross-entropy term (hard-label supervision):

[0125]

[0126] A smooth probability distribution is adopted with temperature T>1.

[0127] S6 continuously records smooth environmental semantic graphs, long and short thought chains, motion control commands and decisions;

[0128] Monitoring operational quality based on obstacle avoidance success rate and path deviation indicators;

[0129] When a decline in operational quality or a significant change in the environment is detected, background micro-batch incremental training is triggered to fine-tune the long and short thought chains online.

[0130] The finely tuned long and short thought chains are automatically deployed to the real-time inference engine to achieve closed-loop optimization of the strategy.

[0131] Understandably, the following metrics are defined: obstacle avoidance success rate SR = number of successful obstacle avoidance attempts / total number of obstacle avoidance triggers; average path deviation PD = average minimum distance between trajectory point and planned path; and smoothness SM = standard deviation of control command acceleration / steering angle change rate. These metrics are used in the performance evaluation module, which calculates them using a sliding window and compares them with preset thresholds (SR0, PD0, SM0). When any metric is below or above the safe range, a "retraining request" is issued, and the incremental training process begins.

[0132] Extract the k most recent training samples that meet the triggering conditions from the latest logs, including the input and the "ideal" inference chain and command labels; fine-tune the prompt template vector and a small number of parameters in the last layer of the thought chain module, control the training rounds to 1-5 rounds to ensure fast convergence, and use mini-batch online gradient updates; evaluate the three metrics SR, PD, and SM on the reserved validation set to ensure no backsliding.

[0133] The automated model deployment includes secure switching, performance verification, and rollback mechanisms. Through A / B traffic splitting, a small portion of inference requests are switched to the fine-tuned model for online monitoring. If the new model outperforms the original model in T consecutive calls, a full switch is completed. If the new model's performance is substandard, it automatically reverts to the old model, and anomaly logs are recorded. After deployment, the new model's performance continues to be monitored according to the S502 process, and performance reports are generated periodically to provide optimization decision-making support for operations and development.

[0134] As described above, the large-scale model long-short chain collaborative reasoning autonomous mobile platform control method provided by this invention unifies the encoding of environmental dynamic data and user intent data into a joint representation, improving cross-modal understanding and decision consistency, ensuring smooth passage, and enabling real-time decision optimization based on actual conditions according to environmental dynamic data and user intent data, thus avoiding the lack of interpretability and adaptability of control logic.

[0135] The above description is merely a specific embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the technical scope disclosed in the present invention should be included within the scope of protection of the present invention. Therefore, the scope of protection of the present invention should be determined by the appended claims.

Claims

1. A control method for an autonomous mobile platform with large-scale model long-short chain collaborative reasoning, characterized in that, Includes the following steps: Real-time capture of dynamic environmental data and user intent data; The environmental dynamic data and the user intent data are preprocessed to obtain a smooth environmental semantic graph; The smoothed environment semantic graph and the user intent data are input into the Transformer large model to obtain inference results; The long and short thought chains generate decision results through the reasoning results.

2. The control method for a large-scale model long-short chain collaborative inference autonomous mobile platform as described in claim 1, characterized in that, The real-time capture of dynamic environmental data and user intent data further includes: Real-time acquisition of 3D point cloud data, depth images, inertial data, and voice data; The voice data is converted into text format as the user intent data.

3. The control method for a large-scale model long-short chain collaborative inference autonomous mobile platform as described in claim 1, characterized in that, The preprocessing of the environmental dynamic data and the user intent data to obtain a smoothed environmental semantic graph further includes: Match the corresponding 3D point cloud data, depth image, and inertial data based on the timestamp of the user intent data; The corresponding inertial data is filtered and bias drift is compensated to obtain preliminary inertial data; Median filtering with a fixed window size is applied to the depth dimension of the corresponding 3D point cloud data and the pixel domain of the depth image to remove isolated noise points and abrupt pixels from the 3D point cloud data and the depth image, respectively. The preliminary inertial data is smoothed using a Kalman filter to obtain the final inertial data; The final inertial data is fed back into the 3D point cloud data and the depth image for matching and filtering to obtain a preliminary environmental semantic map; The preliminary environmental semantic map is subjected to multimodal processing to obtain the smoothed environmental semantic map.

4. The control method for a large-scale model long-short chain collaborative reasoning autonomous mobile platform as described in claim 3, characterized in that, The step of performing multimodal processing on the preliminary environmental semantic map to obtain the smoothed environmental semantic map further includes: The initial environmental semantic map is divided into several image patches of fixed size; The image patch is subjected to radial linear projection to obtain a sequence of visual feature vectors. User intent data is segmented and embedded into word vectors, and then positional encoding is added to form a sequence of language feature vectors. The visual feature vector sequence and the language feature vector sequence are concatenated in sequence, and separator markers are inserted and hierarchical normalization is performed to obtain the smooth environment semantic map.

5. The control method for a large-scale model long-short chain collaborative inference autonomous mobile platform as described in claim 1, characterized in that, The long and short thought chains generate decision results through the reasoning results, and further include: The inference cycle of the Transformer large model is divided into several short time windows, and the inference results are collected in each time window. Based on the collected reasoning results, local reasoning prompts and global prompts are constructed; Call the inference interface of the Transformer large model and output the local obstacle avoidance action sequence based on the local inference prompts; The top k actions with the highest confidence in the local obstacle avoidance action sequence are retained; By identifying major peer regions, planning key intermediate points leading to the destination, connecting key points to generate continuous paths, and generating multi-step inference chains, a smooth and executable global path is generated by applying Bézier curve fitting to the node sequence. The confidence scores of the top k actions with the highest confidence scores in the local obstacle avoidance action sequence and the confidence scores of the actions in the global path are normalized to obtain normalized confidence scores. The actions of the obstacle avoidance sequence are linearly combined with the key target points of the global path according to the normalized confidence to generate the decision result.

6. The control method for an autonomous mobile platform with long and short chain collaborative reasoning in a large model as described in claim 1, characterized in that, The aforementioned control method for a large-scale model long-short chain collaborative inference autonomous mobile platform also includes: The gold standard for generating structured control commands; By minimizing the KL divergence with the output, the decision result is mapped to a command template containing fields such as target coordinates, velocity, and obstacle avoidance priority.

7. The control method for a large-scale model long-short chain collaborative inference autonomous mobile platform as described in claim 5, characterized in that, The control method for a large-scale model long-short chain collaborative inference autonomous mobile platform also includes: Continuously record the semantic graph of the smooth environment, long and short thought chains, motion control commands and decisions; Monitoring operational quality based on obstacle avoidance success rate and path deviation indicators; When a decline in operational quality or a significant change in the environment is detected, background micro-batch incremental training is triggered to fine-tune the long and short thought chains online. The finely tuned long and short thought chains are automatically deployed to the real-time inference engine to achieve closed-loop optimization of the strategy.

Images