A control method for embodied intelligence adaptive communication and cloud collaboration

Abstract

The application discloses a kind of self-adapting communication and cloud collaborative control methods for embodied intelligence, belong to wireless communication network and artificial intelligence cross technical field;Through real-time data acquisition of multi-modal sensor, the robot working state is divided into high load, medium load, light load, based on deep Q network (DQN) scheduling network communication parameter, the communication resource allocation of different load is realized;Establish end-edge-cloud collaborative mechanism, according to task computation, data volume and delay constraint adaptive decision task unloading target, form the joint optimization closed loop of communication and collaboration;The application realizes the fine configuration of communication resource and on-demand optimization of cloud computing call of embodied intelligent robot under different working modes by two core mechanisms of adaptive communication resource scheduling driven by working state and dynamic decision of cloud collaborative, significantly reduces the deployment cost of communication module, reduces communication energy consumption, while meeting the ultra-low latency and high reliability requirements of high-load tasks.

Description

Technical Field

[0001] This invention relates to the technical field of the intersection of wireless communication networks and artificial intelligence, and in particular to a control method for adaptive communication and cloud collaboration oriented towards embodied intelligence. Background Technology

[0002] With the development of technology, embodied intelligence is at a critical juncture of transitioning from "virtual" to "real." However, the large-scale commercialization of embodied intelligent robots still faces the following challenges: 1. The value of communication links is seriously underestimated and the deployment cost is high. Currently, robot communication modules still generally adopt standardized solutions of "general industrial components" and lack dedicated optimization for embodied intelligence scenarios. As a result, the deployment cost of communication links accounts for 15% to 20% of the total cost of the robot, which seriously restricts the large-scale promotion of robots in civilian scenarios.

[0003] 2. There is a severe mismatch between communication resource allocation and business needs. Embossed intelligent robots operate in various modes, including high load, medium load, and standby. In high load mode (such as precision assembly or human-robot collaborative operations), the uplink bandwidth (high-definition video + sensor data transmission) requirement is extremely high, needing to reach >100Mbps, end-to-end latency <10ms, and reliability of 99.99%. In standby mode, the robot only needs to maintain a basic heartbeat connection and low-power listening; a data transmission rate of less than 1Mbps is sufficient. Existing network resource allocation strategies mostly use fixed configurations or simple periodic adjustments, failing to achieve fine-grained adaptive scheduling based on the robot's real-time task status and channel conditions. This results in a significant waste of high-bandwidth resources in standby mode, increasing communication power consumption and operating costs.

[0004] 3. The lack of dynamic optimization mechanisms in cloud collaboration means that current embodied intelligent robots mostly adopt a fixed "edge-cloud" architecture: relying on cloud computing power for large model inference under high-load tasks, and processing locally under light-load tasks. However, existing solutions lack the ability to dynamically adjust cloud collaboration strategies, and cannot adaptively determine which tasks are executed in the cloud and which are executed on the edge based on factors such as real-time network quality, cloud load, and task urgency. This "one-size-fits-all" approach neither fully leverages the advantages of cloud computing power nor ensures the robot's basic autonomous operation capabilities in weak network environments. Summary of the Invention

[0005] The purpose of this invention is to provide a control method for adaptive communication and cloud collaboration for embodied intelligence. It establishes two core mechanisms: adaptive communication resource scheduling driven by working state and dynamic decision-making for cloud collaboration. This enables fine-grained configuration of communication resources and on-demand optimization of cloud computing power in different working modes of embodied intelligent robots, significantly reducing the deployment cost of communication modules and reducing communication energy consumption, while meeting the requirements of ultra-low latency and high reliability for high-load tasks.

[0006] To achieve the above objectives, the present invention provides a control method for adaptive communication and cloud-based collaboration oriented towards embodied intelligence, comprising the following steps: Step S1: Establish an embodied intelligent working structure, including a cloud center, base station edge nodes, and robots connected in sequence; based on the robot's multimodal sensor data, classify the robot's working state into three levels, including high load, medium load, and light load; Step S2: The robot's communication module sets communication resources based on the working status and realizes dynamic scheduling of communication resources; Step S3: Establish a collaborative dynamic decision-making mechanism, including the terminal side, edge side, and cloud side, and set corresponding task offloading decision algorithms; Step S4: Combine the adaptive communication resource scheduling in step S2 and the dynamic decision-making mechanism in step S3 into a joint optimization closed loop, and dynamically adjust the task offloading strategy according to the communication conditions. Step S5: Based on the decision result of the task offloading strategy adjustment, feed it back to the communication module to pre-schedule uplink resources.

[0007] Preferably, in step S1, the robot's multimodal sensors include an inertial measurement unit, a joint torque sensor, a vision sensor, a microphone array, and a channel quality monitoring module on the network side. The collected data includes joint motor current, with a sampling rate of 1kHz, end effector load force, robot movement speed, environmental noise level, and channel signal-to-noise ratio. Preferably, the robot's working states are divided in step S1 as follows: Criteria for determining high load state: Total power of the robot's joint motors 60% of rated power, or end effector load. 20N, or speed of movement 1.5 m / s, or the frequency of interaction 10 times / minute.

[0008] Medium load condition determination criteria: The total power of the robot's joint motors is between 20% and 60% of the rated power, and there is no high load condition triggering.

[0009] Light load condition determination criteria: Total power of the robot's joint motors Rated power 20%, movement speed The working status level is updated every 100ms. The update uses a sliding window average with a window length of 5 sampling points. The hysteresis margin for status switching is set to ±5% of the power threshold to avoid frequent status switching caused by instantaneous fluctuations.

[0010] Preferably, in step S2, the robot's communication module sets communication resources based on its working state, as follows: For high loads, the uplink target rate requirement is: End-to-end target latency requirements Target reliability requirements The sensing signal period is set to The scheduling priority is high; For medium load, the uplink target rate requirement is: End-to-end target latency requirements Target reliability requirements The sensing signal period is set to The scheduling priority is medium. For light loads, the uplink target rate requirement is: End-to-end target latency requirements Target reliability requirements The sensing signal period is set to The scheduling priority is low; Preferably, in step S2, the process of dynamically scheduling communication resources is as follows: A smart resource scheduling algorithm based on Deep Q-Network (DQN) is adopted to realize the dynamic allocation of communication resources, including the state space and action space, as follows: The state space includes: current operating state level, channel quality SNR and BLER, buffer queue length, and historical resource utilization; the action space includes: the number of uplink resource blocks allocated, with a value range of [value missing]. Modulation coding scheme level, with a value range of: ; Adjust the transmit power by setting the step size to Adjust the range to set The formula for the corresponding reward function is as follows: ; In the above formula, To represent the weight coefficients of different sub-items, Indicates rate reward, Indicates a delay reward. Indicates a reliability bonus. The formulas for different components of the normalized transmit power cost bonus are as follows: ; ; ; ; In the above formula, =0.001 represents the computation constant; the deep Q-network structure is as follows: input layer dimension is set to 12, hidden layers are set to 2 fully connected layers with 128 and 64 nodes respectively, activation function is ReLU, output layer dimension = 18, experience replay pool capacity = 10000, exploration rate = The power is linearly reduced from 1.0 to 0.05, with the number of attenuation steps set to 5000 and the target network update frequency set to 100 steps. The robot communication resource scheduling execution cycle is aligned with the scheduling cycle of the base station edge node. Each scheduling is based on the action configuration of the DQN output, including uplink resource block allocation, modulation and coding scheme level, and transmit power, and the scheduling results are sent to the robot terminal.

[0011] Preferably, in step S3, the process of establishing a collaborative dynamic decision-making mechanism is as follows: the edge is set as a robot, a lightweight inference engine is deployed, the computing power requirement is in the range of 4-8 TPS, and the response latency requirement is... The edge side is set as the base station edge node, deploying MEC servers, with a computing power requirement of 50-200 TPS and a response latency requirement of 10-50ms; the cloud side is set as the cloud center, deploying large language models and large visual models, with a computing power requirement of >1000 TPS and a response latency requirement of 100-500ms. The corresponding task unloading decision algorithm is as follows: For each computational task generated by the robot... Define the following parameters: task computation amount The unit is GFLPs, and the setting range is 0.1 to 1000. This represents the amount of input data for the task. The unit is MB, and the setting range is 0.1 to 100; task latency constraint. The unit is milliseconds (ms), and the setting range is 5 to 1000; task importance weight. Determined by the task type, the optimization problem for task unloading decisions is as follows: ; The corresponding constraint formulas are as follows: In the above formula, The computation latency depends on the allocated computing resources; For transmission delay, The energy consumption cost of local computation on the edge is denoted by λ=0.5, which is a trade-off coefficient. A greedy algorithm combined with a priority queue is used for real-time decision-making. The algorithm complexity is set to (N·lgN), where N is the number of concurrent tasks, and the decision cycle is set to 100ms. Tasks are weighted according to their importance. Sort the solutions from high to low and try edge computing, edge offloading and cloud offloading in turn. Select the solution that meets the latency constraints and has the lowest energy consumption cost. The update strategy for setting up a collaborative dynamic decision-making mechanism is as follows: every 5 minutes, the task processing success rate, average latency, and resource utilization of the cloud, edge, and terminal sides are counted, the window size is set to 10 periods, and the unloading decision threshold is dynamically adjusted based on the sliding average.

[0012] Preferably, in step S4, the process of forming a joint optimization closed loop is as follows: The communication module provides real-time channel quality and available uplink bandwidth. Real-time channel quality includes SNR, latency, and packet loss rate. A collaborative dynamic decision-making mechanism dynamically adjusts the task offloading strategy based on communication conditions, as detailed below: When the uplink bandwidth is ≥50Mbps and the latency is ≤20ms, the computationally intensive tasks are offloaded to the cloud side; when the uplink bandwidth is limited to <10Mbps or the latency is >50ms, the tasks are switched to the edge or end side for processing; the collaborative dynamic decision-making mechanism feeds back the offloading decision results to the communication module, which is used to pre-schedule the uplink resources on the end side and reduce the scheduling delay caused by sudden data transmission.

[0013] Therefore, the above-mentioned adaptive communication and cloud-based collaborative control method for embodied intelligence has the following advantages: (1) Significantly reduce communication deployment costs: By scheduling communication resources, the resource occupation in standby mode is reduced to the minimum, which can reduce the real-time communication requirements of base station edge nodes and robots. According to calculations, after adopting the scheduling scheme of the present invention, the hardware cost of the robot communication module can be reduced from 15% to 20% of the total cost to 10% to 14%, a reduction of 30% to 40%. (2) Improve resource utilization efficiency: Compared with the fixed configuration of the communication scheme, the present invention dynamically adjusts the bandwidth and latency target according to the working status, and reduces the uplink bandwidth usage by about 95% in standby mode (from 100Mbps to 1Mbps), and reduces communication energy consumption by about 70% to 80%.

[0014] (3) Ensure the ultimate performance of high-load tasks: Under high load conditions, the system automatically allocates the highest priority resources to ensure that the end-to-end latency is ≤10ms, the uplink rate is ≥100Mbps, and the reliability is ≥99.99%, meeting the stringent requirements of precision industrial operation and real-time human-machine interaction.

[0015] (4) Improve robustness in weak network environment: When uplink bandwidth is limited or latency deteriorates, the collaborative dynamic decision-making mechanism automatically offloads computing tasks from the cloud to the edge or end side, avoiding "robot malfunction" caused by network problems. According to actual tests, the system can still maintain basic autonomous operation capability in weak network environment, and the task completion rate is ≥85%.

[0016] The technical solution of the present invention will be further described in detail below with reference to the accompanying drawings and embodiments. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the overall process of an adaptive communication and cloud-based collaborative control method for embodied intelligence according to the present invention. Figure 2 This is a flowchart illustrating the working state of the embodied intelligence working structure in the adaptive communication and cloud-based collaborative control method for embodied intelligence according to the present invention. Figure 3 This is a flowchart illustrating the execution of the task offloading decision algorithm in the adaptive communication and cloud-based collaborative control method for embodied intelligence according to the present invention. Detailed Implementation

[0018] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations. Specific model specifications need to be selected and determined according to the actual specifications of the device, etc. The specific selection calculation method adopts existing technology in the art, and therefore will not be described in detail.

[0019] Example like Figure 1 As shown, this invention provides a control method for adaptive communication and cloud-based collaboration oriented towards embodied intelligence, comprising the following steps: Step S1: As Figure 2 As shown, an embodied intelligent working structure is established, including a cloud center, base station edge nodes, and robots connected in sequence; based on the robot's multimodal sensor data, the robot's working state is divided into three levels, including high load, medium load, and light load; The robot's multimodal sensors include an inertial measurement unit, joint torque sensors, vision sensors, a microphone array, and a channel quality monitoring module on the network side. The collected data includes joint motor current at a sampling rate of 1kHz, end effector load force, robot movement speed, environmental noise level, and channel signal-to-noise ratio.

[0020] The robot's working states are divided as follows: Criteria for determining high load state: Total power of the robot's joint motors 60% of rated power, or end effector load. 20N, or speed of movement 1.5 m / s, or the frequency of interaction 10 times / minute; Medium load condition determination criteria: The total power of the robot's joint motors is between 20% and 60% of the rated power, and there is no high load condition triggering. Light load condition determination criteria: Total power of the robot's joint motors Rated power 20%, movement speed The working status level is updated every 100ms. The update uses a sliding window average with a window length of 5 sampling points. The hysteresis margin for status switching is set to ±5% of the power threshold to avoid frequent status switching caused by instantaneous fluctuations.

[0021] Step S2: The robot's communication module sets communication resources based on the working status and realizes dynamic scheduling of communication resources; The robot's communication module sets communication resources based on its working status, as follows: For high loads, the uplink target rate requirement is: End-to-end target latency requirements Target reliability requirements The sensing signal period is set to The scheduling priority is high; For medium load, the uplink target rate requirement is: End-to-end target latency requirements Target reliability requirements The sensing signal period is set to The scheduling priority is medium. For light loads, the uplink target rate requirement is: End-to-end target latency requirements Target reliability requirements The sensing signal period is set to The scheduling priority is low.

[0022] The corresponding table is as follows: Table 1

[0023] The process of achieving dynamic scheduling of communication resources is as follows: A smart resource scheduling algorithm based on Deep Q-Network (DQN) is used to dynamically allocate communication resources, including a state space and an action space, as detailed below: The state space includes: current operating state level, channel quality SNR and BLER, buffer queue length, and historical resource utilization; the action space includes: the number of uplink resource blocks allocated, with a value range of [value missing]. Modulation coding scheme level, with a value range of: ; Adjust the transmit power by setting the step size to Adjust the range to set The formula for the corresponding reward function is as follows: ; In the above formula, To represent the weight coefficients of different sub-items, Indicates rate reward, Indicates a delay reward. Indicates a reliability bonus. The formulas for different components of the normalized transmit power cost bonus are as follows: ; ; ; ; In the above formula, =0.001 represents the computation constant; the deep Q-network structure is as follows: input layer dimension is set to 12, hidden layers are set to 2 fully connected layers with 128 and 64 nodes respectively, activation function is ReLU, output layer dimension = 18, experience replay pool capacity = 10000, exploration rate = The value decreases linearly from 1.0 to 0.05, with the number of decay steps set to 5000 and the target network update frequency set to 100 steps. The training process for Deep Q-Network (DQN) is as follows: Initially, initialize the experience replay pool (capacity 10000), the main network Q, and the target network Q' (parameter synchronization). In each update state, execute: Obtain the current state s, which includes 12 dimensions: 3-dimensional working state level ne-ht encoding, 1-dimensional SNR normalization, 1-dimensional BLER normalization, 1-dimensional cache queue length normalization, 3-dimensional historical resource utilization, and the remaining 3 dimensions are reserved. Use an ε-greedy strategy to select actions. : In random number < Random actions, when the random number is greater than 100. When using this formula, the following method is employed: ; Set the exploration rate ε to decay linearly from 1.0 to 0.05 (decay steps = 5000), execute action a, including configuring RB, MCS, and power, obtain reward r (calculated according to the reward function), and enter the next state. Experience Store the data in the experience replay pool, randomly sample a mini-batch (size 64) from the replay pool, and calculate the target value using the following formula: ; • Update the Q parameters of the main network using gradient descent, and copy the main network parameters to the target network Q' every 100 steps until a trained Q network is obtained; The robot communication resource scheduling execution cycle is aligned with the scheduling cycle of the base station edge nodes. Each scheduling is based on the uplink resource block allocation, modulation and coding scheme level, and transmit power output by DQN, and the scheduling results are sent to the robot terminal to achieve dynamic allocation of communication resources. The online scheduling process is as follows: At the start, obtain the current state s (same as the training input), call the trained Q-network, perform forward propagation to obtain the Q-values ​​of all actions, and select the action with the largest Q-value. Output actions: RB allocation number (1~100), MCS level (0~15), transmit power adjustment (-10~+10dBm). These are sent to the end side via RRC signaling; the system waits for the next scheduling cycle (20ms), executes cyclically, with an experience replay pool capacity of 10000, a target network update frequency of 100 steps, a learning rate of 0.001, an optimizer named Adam, and a loss function set to mean squared error.

[0024] Step S3: Establish a collaborative dynamic decision-making mechanism, including the terminal side, edge side, and cloud side. Based on the task computation volume, input data volume, and latency constraints, set up corresponding task offloading decision algorithms and output task allocation results that meet latency requirements and minimize costs. The process of establishing a collaborative dynamic decision-making mechanism is as follows: The edge is set up as a robot, a lightweight inference engine is deployed, the computing power requirement is in the range of 4-8 TPS, and the response latency requirement is... The edge side is set as the base station edge node, deploying MEC servers, with a computing power requirement of 50-200 TPS and a response latency requirement of 10-50ms; the cloud side is set as the cloud center, deploying large language models and large visual models, with a computing power requirement of >1000 TPS and a response latency requirement of 100-500ms. Corresponding design task unloading decision algorithm, such as Figure 3 As shown, specifically as follows: For each computational task generated by the robot Define the following parameters: task computation amount The unit is GFLPs, and the setting range is 0.1 to 1000. This represents the amount of input data for the task. The unit is MB, and the setting range is 0.1 to 100; task latency constraint. The unit is milliseconds (ms), and the setting range is 5 to 1000; task importance weight. Determined by the task type, the optimization problem for task unloading decisions is as follows: ; The corresponding constraint formulas are as follows: In the above formula, The computation latency depends on the allocated computing resources; For transmission delay, The energy consumption cost of local computation on the edge is denoted by λ=0.5, which is a trade-off coefficient. A greedy algorithm combined with a priority queue is used for real-time decision-making. The algorithm complexity is set to (N·lgN), where N is the number of concurrent tasks, and the decision cycle is set to 100ms. Tasks are weighted according to their importance. Sort the solutions from highest to lowest performance, and select the solution that satisfies the latency constraint and has the lowest energy consumption cost based on three methods: edge computing, edge offloading, and cloud offloading. The specific process is as follows: Edge computing conditions: Current available computing power on the edge (≤4~8 TPS), local computing latency: ; like Calculation cost And record the plan; Side-mounted unloading: Uplink transmission delay (Obtained from the communication module), delay calculated on the side. (Assuming 50 TPS shared). If Then calculate the total cost. Record the side scheme.

[0025] Cloud-side offloading: Uplink transmission latency (Same side), cloud-side computation delay (Assuming 1000 TPS shared) Backhaul latency Ignore (or fix for 10ms). If Then calculate the total cost. Record cloud-based solutions.

[0026] Choose the best option: Select the option with the lowest cost from the feasible options. If multiple options have the same cost, prioritize the terminal side option based on bandwidth, then the edge side option, and finally the cloud side option.

[0027] The update strategy for setting up a collaborative dynamic decision-making mechanism is as follows: every 5 minutes, the task processing success rate, average latency, and resource utilization of the cloud, edge, and terminal sides are counted, the window size is set to 10 periods, and the unloading decision threshold is dynamically adjusted based on the sliding average.

[0028] Assign tasks to selected execution units (end / edge / cloud), update resource status: deduct computing resources (temporarily occupied) of the corresponding execution unit, record the estimated completion time of the task, and process the next task until the queue is empty.

[0029] Periodic triggering: The above process is re-executed every 100ms, including processing newly arrived tasks and updating the status of assigned but incomplete tasks. Edge computing power is 4-8 TPS, cloud computing power is 50-200 TPS, and cloud computing power is >1000 TPS; task computation volume Range 0.1–1000 GFLPs; Data volume Range 0.1–100 MB; latency constraints Range 5–1000 ms.

[0030] Step S4: Combine the communication resource scheduling in step S2 and the dynamic decision-making mechanism in step S3 into a joint optimization closed loop, and dynamically adjust the task offloading strategy according to the communication conditions. The process of forming a joint optimization closed loop is as follows: The communication module provides real-time channel quality and available uplink bandwidth. Real-time channel quality includes SNR, latency, and packet loss rate. The collaborative dynamic decision-making mechanism dynamically adjusts the task offloading strategy according to communication conditions, as detailed below: When the uplink bandwidth is ≥50Mbps and the latency is ≤20ms, the computationally intensive tasks are offloaded to the cloud side; when the uplink bandwidth is limited to <10Mbps or the latency is >50ms, the tasks are switched to the edge or end side for processing; the collaborative dynamic decision-making mechanism feeds back the offloading decision results to the communication module, which is used to pre-schedule the uplink resources on the end side and reduce the scheduling delay caused by sudden data transmission.

[0031] Step S5: Based on the decision result of the task unloading strategy adjustment, feedback is sent to the communication module to pre-schedule uplink resources. This process is repeated until the robot shuts down, as detailed below: The communication module provides the collaborative dynamic decision-making mechanism with real-time channel quality (SNR, latency, packet loss rate) and available uplink bandwidth. The collaborative dynamic decision-making mechanism adjusts the offloading strategy based on communication conditions and feeds back the predicted data upload volume to the communication module for pre-scheduling (reserving bandwidth 1-2 scheduling cycles in advance), then loops back to step 1. It continues running until shutdown.

[0032] The specific experimental scenarios are provided below: Experimental scenario (1), welding robot in industrial production scenario: A 6-axis welding robot was selected for testing in an automobile manufacturing plant. The robot needs to transmit high-definition video (1080p@60fps, bitrate about 50Mbps) and sensor data (6-axis joint angle, torque, welding gun current, data volume about 10Mbps) in real time during welding operations (under high load), and receive welding trajectory correction instructions sent from the cloud.

[0033] Parameter configuration: The system automatically identifies the robot's state as high-load, configuring an uplink target rate ≥100Mbps, end-to-end target latency ≤10ms, and reliability ≥99.99%. DQN algorithm output RB allocation = 80 (total system RB = 100), MCS level = 12 (64QAM), transmit power +3dBm. Cloud-based collaborative decision-making offloads the welding quality assessment task (computational cost approximately 200GFLPs) to the cloud, while retaining real-time obstacle avoidance judgment (computational cost approximately 5GFLPs) for edge processing.

[0034] Test Results: After 8 hours of continuous operation, approximately 300 welding tasks were completed. Average uplink speed = 112Mbps, 95th percentile latency = 8.7ms, reliability = 99.992%. Compared to a fixed configuration (always allocated 100Mbps bandwidth), communication power consumption was reduced by approximately 65% ​​(with a significant decrease in power consumption during standby), and the welding quality pass rate was improved by approximately 2% (thanks to the more accurate identification of weld defects by the cloud-based large model).

[0035] Experimental scenario (2), shopping mall guide robot in commercial service scenario: Five guide robots are deployed in a large shopping mall. Their work includes welcoming and guiding (S1 high load state, requiring real-time voice interaction and gesture recognition), autonomous navigation (S2 medium load state, regular navigation and obstacle avoidance), and standby charging (S3 light load state).

[0036] Parameter Configuration: The system automatically switches parameters based on real-time status. Peak period (S1): Uplink rate ≥ 100Mbps, latency ≤ 10ms; Cruise period (S2): Uplink rate = 30Mbps, latency ≤ 20ms; Standby period (S3): Uplink rate ≤ 1Mbps, latency ≤ 100ms. Cloud-based collaborative decision-making offloads multi-turn dialogue tasks to the cloud and gesture recognition (computational load approximately 30GFLPs) to the edge MEC.

[0037] Test Results: After 7 days of continuous operation (12 hours per day), the average uplink bandwidth usage was 18.5 Mbps (an 81.5% reduction compared to a fixed 100 Mbps solution), and the power consumption of the communication module decreased by approximately 72%. Voice interaction success rate was 98.7%, average response latency was 1.2 seconds, and user satisfaction rating was 4.6 / 5.0. When 5 robots were running concurrently, the total uplink throughput remained between 150 and 200 Mbps, with no latency jitter caused by resource conflicts.

[0038] Experimental scenario (3), companion robot in home service scenario: A companion robot is deployed in a home environment. Its work includes health monitoring of the elderly, voice companionship, and entertainment interaction (S1 / S2 / S3 states alternate). Special attention is paid to the robustness test in weak network environment - the uplink bandwidth in areas with weak Wi-Fi signal in the home (such as the corner of the bedroom) is measured to be about 5 to 8 Mbps, and the SNR is about 6 to 10 dB.

[0039] Parameter configuration: When the channel quality degrades (SNR < 10dB), the system automatically switches the task offloading strategy from "cloud-side priority" to "end-side priority". The large model dialogue task (computational cost of approximately 500GFLPs) that was originally offloaded to the cloud is downgraded to a small model on the end side (parameters of approximately 0.5B), while the health monitoring and emergency call functions are retained in the cloud backup.

[0040] Test results: In areas with weak Wi-Fi (uplink bandwidth ≈ 6Mbps), the system was still able to maintain basic voice interaction and health monitoring functions, with a task completion rate of approximately 87%. When the robot moved to an area with good signal (uplink bandwidth restored to over 80Mbps), the system automatically resumed cloud-based large-scale model dialogue, with dialogue quality and richness improving by approximately 40%.

[0041] Therefore, this invention adopts an adaptive communication and cloud-based collaborative control method for embodied intelligence, establishing two core mechanisms: adaptive communication resource scheduling driven by working state and dynamic decision-making for cloud collaboration. This enables refined configuration of communication resources and on-demand optimization of cloud computing power calls for embodied intelligent robots in different working modes, significantly reducing the deployment cost of communication modules and reducing communication energy consumption, while meeting the ultra-low latency and high reliability requirements of high-load tasks.

[0042] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit them. Although the present invention has been described in detail with reference to preferred embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the technical solutions of the present invention, and these modifications or equivalent substitutions cannot cause the modified technical solutions to deviate from the spirit and scope of the technical solutions of the present invention.

Claims

1. A control method for adaptive communication and cloud-based collaboration oriented towards embodied intelligence, characterized in that: Includes the following steps: Step S1: Establish an embodied intelligent working structure, including a cloud center, base station edge nodes, and robots connected in sequence; based on the robot's multimodal sensor data, classify the robot's working state into three levels, including high load, medium load, and light load; Step S2: The robot's communication module sets communication resources based on the working status and realizes dynamic scheduling of communication resources; Step S3: Establish a collaborative dynamic decision-making mechanism, including the edge, cloud, and terminal sides, based on the task's computational load. Input data volume and delay constraints A corresponding task unloading decision algorithm is set, and the output is the task allocation result that satisfies the latency requirement and minimizes the cost. Step S4: Combine the communication resource scheduling in step S2 and the dynamic decision-making mechanism in step S3 into a joint optimization closed loop, and dynamically adjust the task offloading strategy according to the communication conditions. Step S5: Based on the decision result of the task unloading strategy adjustment, feed it back to the communication module to pre-schedule the uplink resources and repeat the process until the robot is shut down.

2. The control method for adaptive communication and cloud collaboration oriented towards embodied intelligence according to claim 1, characterized in that: In step S1, the robot's multimodal sensors include an inertial measurement unit, joint torque sensors, vision sensors, a microphone array, and a channel quality monitoring module on the network side. The collected data includes joint motor current, with a sampling rate of 1kHz, end effector load force, robot movement speed, environmental noise level, and channel signal-to-noise ratio.

3. The control method for adaptive communication and cloud collaboration oriented towards embodied intelligence according to claim 2, characterized in that: In step S1, the robot's working states are divided as follows: Criteria for determining high load state: Total power of the robot's joint motors 60% of rated power, or end effector load. 20N, or speed of movement 1.5 m / s, or the frequency of interaction 10 times / minute; Medium load condition determination criteria: The total power of the robot's joint motors is between 20% and 60% of the rated power, and there is no high load condition triggering. Light load condition determination criteria: Total power of the robot's joint motors Rated power 20%, movement speed The working status level is updated every 100ms. The update uses a sliding window average with a window length of 5 sampling points. The hysteresis margin for state switching is set to ±5% of the power threshold to avoid frequent state switching caused by instantaneous fluctuations.

4. The control method for adaptive communication and cloud collaboration oriented towards embodied intelligence according to claim 3, characterized in that: In step S2, the robot's communication module sets communication resources based on its working state, as follows: For high loads, the uplink target rate requirement is: End-to-end target latency requirements Target reliability requirements The sensing signal period is set to The scheduling priority is high; For medium load, the uplink target rate requirement is: End-to-end target latency requirements Target reliability requirements The sensing signal period is set to The scheduling priority is medium. For light loads, the uplink target rate requirement is: End-to-end target latency requirements Target reliability requirements The sensing signal period is set to The scheduling priority is low.

5. The control method for adaptive communication and cloud collaboration oriented towards embodied intelligence according to claim 1, characterized in that: In step S2, the process of dynamically scheduling communication resources is as follows: A smart resource scheduling algorithm based on Deep Q-Network (DQN) is adopted to realize the dynamic allocation of communication resources, including the state space and action space, as follows: The state space includes: current operating state level, channel quality SNR and BLER, buffer queue length, and historical resource utilization; the action space includes: the number of uplink resource blocks allocated, with a value range of [value missing]. Modulation coding scheme level, with a value range of: ; Adjust the transmit power, set the step size to Adjust the range to set The formula for the corresponding reward function is as follows: ； In the above formula, To represent the weight coefficients of different sub-items, Indicates rate reward, Indicates a delay reward. Indicates a reliability bonus. The formulas for different components of the normalized transmit power cost bonus are as follows: ； ； ； ； In the above formula, =0.001 represents the computation constant; the deep Q-network structure is as follows: input layer dimension is set to 12, hidden layers are set to 2 fully connected layers with 128 and 64 nodes respectively, activation function is ReLU, output layer dimension = 18, experience replay pool capacity = 10000, exploration rate = The power is linearly reduced from 1.0 to 0.05, with the number of attenuation steps set to 5000 and the target network update frequency set to 100 steps. The robot communication resource scheduling execution cycle is aligned with the scheduling cycle of the base station edge node. Each scheduling is based on the action configuration of the DQN output, including uplink resource block allocation, modulation and coding scheme level, and transmit power, and the scheduling results are sent to the robot terminal.

6. The control method for adaptive communication and cloud collaboration oriented towards embodied intelligence according to claim 5, characterized in that: In step S3, the process of establishing a collaborative dynamic decision-making mechanism is as follows: the edge is set as a robot, a lightweight inference engine is deployed, the computing power requirement is in the range of 4-8 TPS, and the response latency requirement is... The edge side is set as the base station edge node, deploying MEC servers, with a computing power requirement of 50-200 TPS and a response latency requirement of 10-50ms; the cloud side is set as the cloud center, deploying large language models and large visual models, with a computing power requirement of >1000 TPS and a response latency requirement of 100-500ms. The corresponding task unloading decision algorithm is as follows: For each computational task generated by the robot... Define the following parameters: task computation amount The unit is GFLPs, and the setting range is 0.1 to 1000. This represents the amount of input data for the task. The unit is MB, and the setting range is 0.1 to 100; task latency constraint. The unit is milliseconds (ms), and the setting range is 5 to 1000; task importance weight. Determined by the task type, the optimization problem for task unloading decisions is as follows: ； The corresponding constraint formulas are as follows: In the above formula, The computation latency depends on the allocated computing resources; For transmission delay, The energy consumption cost is calculated locally on the edge side, with λ=0.5 as a tradeoff coefficient; a greedy algorithm combined with a priority queue is used for real-time decision-making, and the algorithm complexity is set to... N represents the number of concurrent tasks, and the decision period is set to 100ms; tasks are weighted according to their importance. Sort the solutions from high to low and try the three solutions in turn: edge computing, edge offloading, and cloud offloading. Select the solution that meets the latency constraints and has the lowest energy consumption cost. The update strategy for setting up a collaborative dynamic decision-making mechanism is as follows: every 5 minutes, the task processing success rate, average latency, and resource utilization of the cloud, edge, and terminal sides are counted, the window size is set to 10 periods, and the unloading decision threshold is dynamically adjusted based on the sliding average.

7. The control method for adaptive communication and cloud collaboration oriented towards embodied intelligence according to claim 6, characterized in that: In step S4, the process of forming a joint optimization closed loop is as follows: The communication module provides real-time channel quality and available uplink bandwidth. Real-time channel quality includes SNR, latency, and packet loss rate. A collaborative dynamic decision-making mechanism dynamically adjusts the task offloading strategy based on communication conditions, as detailed below: When the uplink bandwidth is ≥50Mbps and the latency is ≤20ms, the computationally intensive tasks are offloaded to the cloud side; when the uplink bandwidth is limited to <10Mbps or the latency is >50ms, the tasks are switched to the edge or end side for processing; the collaborative dynamic decision-making mechanism feeds back the offloading decision results to the communication module, which is used to pre-schedule the uplink resources on the end side and reduce the scheduling delay caused by sudden data transmission.

Images