An asynchronous LLM-MPC robot control method for non-deterministic inference latency
By using the collaborative control of edge-side MPC and latency predictor, the instability of robot systems caused by the nondeterministic latency of large cloud models is solved, achieving efficient task completion and safety assurance under different working conditions.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- GUANGDONG UNIV OF TECH
- Filing Date
- 2026-04-07
- Publication Date
- 2026-07-24
AI Technical Summary
Existing technologies struggle to effectively manage the nondeterministic inference latency of large cloud-based models, leading to issues such as reference trajectory misalignment, geometric mismatch, and task termination during critical segment switching in robot systems. In particular, it is difficult to distinguish the impact of different scenarios and channel states in long-tail late events.
By adopting edge-side MPC as the execution core, a cloud-edge collaborative two-layer control architecture is constructed. Using a concurrent preemption mechanism and a latency predictor, combined with rolling release time and AoI segmented decision, the waiting time in the cloud is transformed into a bounded reference disturbance that can be tolerated on the edge side, thereby realizing asynchronous LLM-MPC control.
It significantly reduced the task failure rate, improved the stability and safety of the robot system under different working conditions, maintained the gains of high-level cloud planning, and reduced the risk of waiting in the cloud.
Smart Images
Figure CN122449919A_ABST
Abstract
Description
Technical Field
[0001] This invention belongs to the fields of intelligent robot control, networked control, and artificial intelligence fusion control technology. Specifically, it relates to an asynchronous LLM-MPC robot control method, system, and storage medium for addressing the nondeterministic inference latency of large cloud-based models. More particularly, it relates to a scheduling and control technology that reconstructs long-tail waiting in the cloud into bounded reference fluctuations that can be tolerated by the edge model predictive controller through latency prediction, dual-channel concurrent preemption, AoI segmented decision, and last-mile adaptation. The term "nondeterministic inference latency" as used in this specification refers to the random fluctuations and extreme delays exhibited by the response latency of large cloud-based models due to the combined effects of queuing, network jitter, and heavy-tail bursts. In preferred embodiments, this type of latency is primarily characterized by "long-tail latency" with significant tail expansion. Unless otherwise stated, "nondeterministic inference latency" and "long-tail latency" are used synonymously or are context-equivalent. Background Art
[0002] As large language models are increasingly used for high-level decision-making in mobile robots, more and more robot systems are adopting a layered architecture of "generating high-level parameter packages in the cloud and executing low-level closed-loop control at the edge." The cloud side can output high-level parameters such as target speed, path shape, switching position, and control weights based on natural language tasks, environmental semantics, and scene context, while the edge side needs to continuously solve MPC at a fixed control cycle to ensure the real-time performance, stability, and safety of the robot's underlying motion.
[0003] However, the end-to-end response process of large cloud-based models is typically affected by request queuing, inference load fluctuations, message packet reassembly, network congestion, and sudden retransmissions. In such scenarios, latency no longer manifests as the small-disturbance bounded latency of traditional networked control, but rather as a complex long-tailed process composed of piecewise steady-state reference terms, random drift, and heavy-tailed bursts. When an extreme late event happens to coincide with a segment switching window, it can easily lead to reference trajectory misalignment, geometric mismatch at curve entrances or parking anchor points, decreased optimizer hot-start quality, and even solution infeasibility and task abort.
[0004] Existing solutions often address latency issues by using fixed timeouts, simple moving averages, or passively waiting for cloud results. These solutions struggle to differentiate the impact of different scenarios, request structures, and channel states on latency, and they also fail to answer the crucial question of whether waiting for the current cloud result is worthwhile. Furthermore, simply introducing AoI metrics without integrating them with remaining runtime, planned release time, and execution layer degradation logic is insufficient to mitigate the risk of misaligned implementation during critical segment switching.
[0005] Therefore, an asynchronous control method is needed to proactively manage the waiting value and freshness of parameter packages for nondeterministic inference latency, so as to ensure the real-time performance, recursive feasibility, and security of the edge robot's closed loop while preserving the high-level planning gains of the large cloud model. (Summary of the Invention)
[0006] I. Technical problems to be solved
[0007] The technical problem this invention aims to solve is to address the issues of nondeterministic inference latency and long-tail delay in the output of large cloud models in cloud-edge collaborative robot control. It provides an asynchronous LLM-MPC control method with edge-side MPC as the execution core and cloud-based large model channels and template fast channels as the source of concurrent parameters. Through latency prediction, dual-channel concurrent preemption, rolling release time, AoI segmented decision, and last-mile adaptation, the uncontrollable cloud waiting is reconstructed into bounded reference disturbances that the underlying controller can withstand, thereby reducing the task failure rate and improving stability across operating conditions.
[0008] II. Technical Solution
[0009] To solve the above-mentioned technical problems, the present invention adopts the following technical solution:
[0010] 1. Construct a cloud-edge collaborative dual-layer control architecture. On the edge side, the differential mobile robot MPC runs with a fixed sampling period; on the higher layer side, a concurrent preemptive mechanism is used to generate parameter packets between the template channel and the large model channel in the cloud. In dual-channel mode, the first task segment is quickly started using an immediate trigger packet, and subsequent task segments are updated asynchronously around the switching anchor point.
[0011] 2. Establish a Ridge+EWMA delay predictor. Extract features such as amplitude, frequency, target velocity, request length, previous round's actual delay, last-mile switch, channel identifier, and scene key from the candidate parameter packet. After standardization and categorical one-hot expansion, first calculate the basic predicted delay, then perform EWMA correction based on the residual table of the same scene, and obtain the final predicted delay through upper and lower bound clipping.
[0012] 3. Construct a two-stage scheduling logic that combines pre-start, deadline waiting, and planned release. The system calculates the pre-start conditions, waiting deadline, and planned release time based on the remaining running time from the current position to the switching anchor point. When a subsequent task meets the pre-start conditions, the template thread and the cloud thread are started and parameter packages are generated. Within the deadline waiting window, the cloud parameter package that is actually completed is used first; otherwise, the template parameter package is used.
[0013] 4. Set up an information freshness defense line on the execution side. After the parameter package is published, it enters the buffer, and when the trigger anchor point is reached, a segmented freshness judgment is performed based on the publication time as the AoI benchmark; when the AoI exceeds the threshold, either the currently applied parameter package is maintained or a safe stopping package is executed according to degrade=hold or degrade=fallback.
[0014] 5. Introducing rolling release and last-mile adaptation in rolling mode. Rolling mode continuously refreshes the planned release time. When the remaining release time window is less than the adaptation window or the remaining runtime enters the AoI threshold semantic window, the release time is stopped from being pushed back, and bounded fine-tuning is performed on the switching position, target velocity, position item weight, and angular velocity change rate weight.
[0015] 6. In the edge-side model predictive controller, solve the optimization problem containing position error term, heading error term, linear velocity control term, angular velocity control term, control rate of change term, and terminal term, and adopt intelligent warm start in the first control cycle after the new parameter package takes effect to improve the initial value quality of the model predictive control solver.
[0016] III. Beneficial Effects
[0017] Compared with the prior art, the present invention has the following beneficial effects:
[0018] 1. This invention does not passively wait for cloud results, but rather uses a Ridge+EWMA latency predictor, a dual-channel concurrent preemption mechanism, and a real completion determination within the deadline window to collaboratively assess the waiting risk and timing of cloud results, thereby significantly reducing the probability of critical segment switching being hit by extreme delays.
[0019] 2. This invention uses the parameter package release time as the AoI benchmark, unifying the rolling release time strategy with the execution-side AoI defense line, and can maintain the timeliness boundary of the underlying control when there is uncertainty in the arrival of the high-level parameter package.
[0020] 3. This invention, through last-mile adaptation, performs bounded corrections only on the switching position, target velocity, position term weights, and angular velocity rate of change weights. It does not require reconstructing the underlying solver, thus transforming the long-tail waiting problem into a bounded reference disturbance problem that the low-level controller can tolerate.
[0021] 4. In high-noise, long-tail benchmark experiments, the Pred+EWMA strategy reduces the task failure rate from... The relative decline reached ;That The noise levels at low, medium, and high are respectively: Cross-noise fluctuations are only .
[0022] 5. The cloud exposure rate of Pred+EWMA in the entire sample is approximately This indicates that the present invention does not sacrifice security by drastically reducing cloud usage, but rather achieves a better security-intelligence trade-off while maintaining a high level of cloud intelligence exposure. Attached Figure Description
[0023] Figure 1 This is a schematic diagram of the system structure of the cloud-edge collaborative dual-channel long-tail delay scheduling and edge model predictive control of the present invention;
[0024] Figure 2 This is a schematic diagram of the long-tail delay prediction process based on the superposition of Ridge regression and scene residual EWMA in this invention;
[0025] Figure 3 Based on deadline in this invention Planned release time With AoI threshold A schematic diagram of the two-stage scheduling logic;
[0026] Figure 4 This is a schematic diagram of the parameter packet freshness segmentation and degradation process after the execution layer anchor point is triggered in this invention; Detailed Implementation
[0027] The present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. In the accompanying drawings, Figure 1 Used to illustrate the structure of a cloud-edge collaborative dual-channel system. Figure 2 This is used to illustrate the long-tail delay prediction process. Figure 3 Used to illustrate the two-stage scheduling logic, Figure 4 This is used to illustrate the AoI segmentation and degradation process at the execution layer. It should be understood that the following embodiments are for illustrative purposes only and are not intended to limit the scope of protection of this invention.
[0028] Example 1: Cloud-Edge Collaborative Asynchronous Control Architecture
[0029] This embodiment uses a differential speed mobile robot as an example, and the edge-side control cycle is taken as... The robot state and control variables are denoted as follows:
[0030]
[0031] The discrete kinematic model on the edge side satisfies:
[0032]
[0033] Instead of directly issuing time-by-time control inputs, the higher-level side issues parameter packets containing information such as target velocity, trajectory amplitude, trajectory frequency, switching position, and control weights. In dual-channel mode, the first task is an immediately triggered packet; subsequent tasks update parameters asynchronously around the end point of the current segment or the starting anchor point of the next segment.
[0034] Example 2: Delay Prediction and Dual-Channel Concurrent Scheduling
[0035] In this embodiment, Figure 2 focuses on illustrating the calculation process of the delay predictor. The system first calculates the time delay based on the robot's current time $t$ and the start of the next task segment. Estimated remaining run time The calculation formula for two-stage scheduling satisfies:
[0036] in, The pre-start conditions, deadline waiting windows, and planned release times for subsequent task segments are further explained in Figure 3. That is, when... At the same time, both the template thread and the cloud thread are started to generate candidate parameter packages, and then only when...
[0037]
[0038] The system waits for the cloud thread to complete within a defined waiting window. If the cloud thread actually completes within this window, the cloud parameter package is selected; otherwise, the template parameter package is selected. Therefore, the final channel selection in this embodiment is based on the actual availability before the deadline, rather than solely on the predicted value itself.
[0039] Example 3: Rolling Release Time and AoI Segmentation Decision Parameter Packet After the actual release, the release time is used as the basis for decision. As an AoI benchmark, the instantaneous information age is calculated as follows: For non-parking packages, the system determines whether to apply candidate parameter packages at the trigger anchor point according to the AoI segmentation decision logic, and applies the parameter packages accordingly. satisfy:
[0040] For explicit parking packets (i.e., those with target_speed=0), since they inherently carry out conservative safety actions, the system directly executes the parking packet upon reaching the trigger anchor point without performing an AoI threshold comparison. Example 4: Last-mile Adaptation and Edge MPC Solving
[0041] In this embodiment, after the high-level parameter packet is sent, the target velocity in the parameter packet is first... trajectory amplitude Trajectory frequency and cut
[0042] Change position The local reference trajectory is generated, and its calculation formula satisfies: To mitigate geometric shifts and optimizer oscillations caused by switching high-level parameter packages, this embodiment performs last-mile adaptation on candidate parameter packages before official release. After adaptation, the edge-side MPC adopts the following objective function:
[0043]
[0044] Under the nominal parameter configuration, the following parameters can be selected: Example 5: Intelligent Warm-up and Safety Assurance
[0045] In this embodiment, the edge side performs a smart warm-start in the first control cycle after receiving a new parameter packet. The system first constructs an initial sequence of linear velocity values based on the target velocity, and then constructs an initial sequence of angular velocity values based on the difference between adjacent reference headings. Subsequently, starting from the current real state, the system performs forward simulation of the state sequence according to the discrete kinematics model to form dynamically consistent initial state values.
[0046]
[0047] If the controller is configured in conventional mode, the optimal solution obtained in the previous step is used as the initial value for warm-start; if it is configured in cold-start mode, no initial guess value is set.
[0048] Furthermore, this embodiment can also be configured with safety protection logic for situations where new parameter packets have not been applied for an extended period. For example, if the system does not receive a new valid parameter packet within a preset time and the robot is still in motion, a safety parking packet can be automatically constructed to bring the robot to a stop, thereby further ensuring physical safety boundaries.
[0049] Example 6: Experimental Setup and Results
[0050] This embodiment also provides a predictive control system for mobile robots with long-tail latency in large cloud-based models. This system is deployed on an edge computing node (such as a Raspberry Pi hardware platform) and mainly includes: a high-level scheduling module for calculating the two-stage scheduling time based on the estimated remaining runtime; a latency prediction module for predicting arrival times based on a hybrid Ridge and EWMA algorithm; a dual-channel parameter generation module for asynchronously generating parameters for the cloud and template fast channels; a parameter buffering and information age discrimination module for performing AoI freshness discrimination when the anchor point is triggered; and an edge model predictive control module for performing low-level closed-loop tracking at fixed intervals. In a set of experiments on a TurtleBot-RaspberryPi compatible ROS / Gazebo closed-loop platform with interfaces consistent with actual deployment, a high-noise baseline configuration defined by run_noiseH_sweep_v2.sh was used. Simultaneously set high noise switching probability. Maintain length Secondary requests, heavy-tailed burst probability , And the sudden upper realm .
[0051] Experimental results show that, under high-noise long-tailed benchmarks, both Pred+EWMA and Pred-only reduce the task failure rate to a minimum. Among them, Pred+EWMA The noise levels at low, medium, and high are respectively: Cross-noise fluctuations are only Its full sample cloud exposure rate is approximately The above experimental results verify that the asynchronous scheduling and edge control collaborative mechanism described in this invention can effectively reduce the task failure rate with a high cloud participation rate and maintain the stability of end-to-end delivery tail latency under different noise conditions.
Claims
1. An asynchronous LLM-MPC robot control method for nondeterministic reasoning delay, characterized in that, Includes the following steps: Step 1: Construct a cloud-edge collaborative two-layer control architecture, where the edge side runs a model prediction controller with a fixed sampling period, and the higher layer side consists of a cloud-based large model parameter generation channel and an edge template fast channel; generate an immediate trigger parameter package for the first task segment, and trigger anchor points for subsequent task segments. To switch positions and initiate dual-channel concurrent preemption; Step 2: Extract the first Feature vectors of candidate parameter packs First, calculate the base prediction latency, then superimpose the residual EWMA from the same scenario to obtain the trimmed prediction latency. and predicted arrival time ; Step 3: Based on the robot's current time Estimated remaining running time to the start of the next task segment Calculate the deadline for cloud-based results. With the planned release time If the arrival time is predicted If the cloud-based parameter package is selected first, then the template fast-track parameter package will be used first; otherwise, the template fast-track parameter package will be actively switched to. Step 4: Before the plan is released, perform last-mile adaptation to fine-tune the switching position, target velocity, position weight, and angular velocity change rate weight of the candidate parameter packages. Step 5: On the execution side, using the parameter packet release time as the AoI benchmark, when the underlying executor reaches the trigger anchor point, it receives the parameter packet in the buffer and performs an information freshness determination: if the instantaneous information age... If so, then apply this parameter package; if If the current parameter package is retained or the configuration is switched to a safe stop rollback parameter package, then the downgrade configuration will be determined accordingly. The freshness defense threshold; Step 6: Input the parameter package confirmed in Step 5 into the edge-side model predictive controller to drive the robot to move along the updated local reference trajectory, thereby reconstructing the uncontrollable cloud long tail into a bounded reference fluctuation that the underlying model predictive controller can withstand.
2. The method according to claim 1, characterized in that, The feature vector is in, For trajectory amplitude, For trajectory frequency, For the target speed, For the request length, For the previous round of actual latency, Toggle content adaptation switch For channel identification, The scene key is used; the Ridge basic prediction and residual correction satisfy:
3. The method according to claim 1, characterized in that, The remaining runtime, pre-start conditions, waiting deadline, and planned release time are satisfied as follows: in, For cloud waiting deadline margin, For the target release time margin, For pre-startup margin; in dual-channel mode, the template thread and the cloud thread run concurrently. The cloud thread is only waited for within the deadline waiting window. If it does not complete within the window, the template fast channel result is used instead.
4. The method according to claim 1, characterized in that, The freshness determination of the execution-side information satisfies: Among them, if the candidate parameter package is an explicit parking package, that is, its If the trigger anchor point is reached, the parking package will be executed directly without performing an AoI threshold comparison.
5. The method according to claim 1, characterized in that, The local reference trajectory is defined in the parameter package.
6. The method according to claim 1, characterized in that, The model predictive controller solves the following optimization problem for the discrete kinematics model of the differential mobile robot: in, The controller simultaneously satisfies .
7. The method according to claim 1, characterized in that, The rolling release strategy continuously refreshes the planned release time in rolling mode and stops postponing the release time when any of the following conditions are met: in, Adapt the content to the window; stop shifting the release time and then perform last-mile adaptation and complete the parameter package release.
8. The method according to claim 1, characterized in that, The last-mile adaptation includes:
9. The method according to claim 1, characterized in that, In the first control cycle after receiving a new parameter packet at the edge, an intelligent warm-start method is used to generate initial values for the model predictive control solver, where: The discrete kinematic model is then subjected to forward simulation based on the current real state to construct a dynamically consistent sequence of initial state values. When configured in conventional mode, the optimal solution from the previous step is used as the initial value, while no initial value is set when configured in cold-start mode.
10. An asynchronous LLM-MPC robot control system for nondeterministic reasoning delay, characterized in that, include: The method comprises a latency prediction module, a dual-channel concurrent generation module, a rolling release module, a parameter buffering and AoI determination module, a last-mile adaptation module, and an edge model prediction control module; each module is configured to perform the method described in any one of claims 1 to 9.
11. A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the method of any one of claims 1 to 9.