Robot path smoothing control method based on artificial intelligence
By employing asynchronous parallelism, data fusion, and hierarchical motion planning, the problems of inference delay and motion jitter in real-time robot control were solved, generating a continuous and stable control command flow and improving the continuity and stability of robot motion.
Patent Information
- Application Number
- CN202511725927.6
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-24
- Publication Date
- 2026-03-06
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Existing technologies in real-time robot control suffer from motion interruptions caused by inference delays and motion jitter caused by sparse commands. Existing solutions cannot solve both problems simultaneously.
By employing an asynchronous parallel and data fusion approach, and defining a look-ahead step size L and a trigger threshold T, asynchronous triggering and smooth transition of robot commands are achieved. Combined with hierarchical motion planning, spline interpolation algorithm is used to generate smooth trajectories, hiding inference delays and satisfying robot dynamic constraints.
It achieves continuity and stability of robot motion, eliminates motion jitter, generates intelligent, smooth and efficient control command flow, and improves the continuity and reliability of motion.
Smart Images
Figure CN121613822A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the fields of robot real-time control and artificial intelligence technology, and more specifically, to an artificial intelligence-based robot path smoothing control method. Background Technology
[0002] Applying complex artificial intelligence (AI) models, especially large-scale deep neural network models, to real-time robot control faces two long-standing technical bottlenecks:
[0003] 1. Motion interruptions caused by inference delays: The time consumed by a single forward inference in an AI model is significant (typically hundreds of milliseconds to seconds). In the traditional synchronous "perception-reasoning-execution" loop, the robot must wait for the next inference result after each action, causing its motion process to be periodically interrupted, resulting in a significant "stop-and-go" phenomenon, which seriously undermines the continuity and efficiency of the operation.
[0004] 2. Motion jitter caused by sparse commands: For the sake of generalization and computational efficiency, AI models typically output control commands or path points at a low frequency (e.g., 5-10Hz). However, the robot's underlying actuators require high-frequency (e.g., 100-500Hz) control inputs. The current common practice is to use APIs such as MoveL (linear motion) provided by the robot controller to perform simple linear interpolation between the sparse points output by the AI. This interpolation method ignores dynamic constraints, resulting in discontinuous velocity and acceleration curves. This causes vibrations and impacts during robot operation, affecting not only task accuracy but also accelerating mechanical wear.
[0005] Existing solutions either focus on optimizing model computation speed (such as model pruning and quantization) or on performing trajectory filtering in the backend, neither of which can simultaneously solve the two problems at the system architecture level. Summary of the Invention
[0006] In view of the problems existing in the prior art, the purpose of this invention is to provide a robot path smoothing control method based on artificial intelligence.
[0007] To solve the above problems, the present invention adopts the following technical solution;
[0008] An artificial intelligence-based robot path smoothing control method includes: enabling the robot to execute instructions smoothly through asynchronous parallelism and data fusion, with the specific workflow as follows:
[0009] The parameters are defined as follows: two key parameters are defined: the look-ahead step size L and the trigger threshold T, where T < L;
[0010] Initialization and first inference: The system starts up, the AI model performs its first inference, and generates an action sequence of length L. And send it to the execution buffer;
[0011] Execution and asynchronous triggering: The robot executes the first T actions (A1, A2, A3) in the buffer. When the Tth action begins to execute, the next model inference is triggered asynchronously.
[0012] Trajectory fusion, when the robot completes its execution After the Lth action, it does not immediately jump to A7' in the text, but the system in The LT step and Between the first LT steps, linear weighted fusion is performed to generate a new set of smoothly transitioned fused actions;
[0013] The fusion formula, taking step i as an example:
[0014] ;
[0015] The weighting factor α gradually changes from 1 to 0, where 1 indicates complete trust in the old sequence and 0 indicates complete trust in the new sequence.
[0016] The process continues in a loop, with the robot continuing to execute the fused actions A4'' to A6'. At the same time, the system triggers reasoning for subsequent sequences; this process repeats continuously, achieving a smooth and uninterrupted flow of instructions.
[0017] As a further description of the above technical solution:
[0018] The robot motion planning adopts a method that decouples high-level decision-making from low-level execution. The layered motion planning is divided into an outer loop and an inner loop. The outer loop is the AI task planning layer, and the inner loop is the physical trajectory smoothing layer.
[0019] As a further description of the above technical solution:
[0020] The outer loop operates at a low frequency and performs macroscopic, sparse path point planning in the task space based on high-level sensing information, outputting a sequence of sparse path points. .
[0021] As a further description of the above technical solution:
[0022] The low-frequency operation of the outer ring is consistent with the inference frequency of the AI model, and the advanced perceptual information includes vision and touch.
[0023] As a further description of the above technical solution:
[0024] The inner loop operates at high frequency and receives sparse path points output by the outer loop. The inner loop contains a trajectory optimizer based on a physical model, which uses a spline interpolation algorithm to generate a trajectory that passes through or approximates the physical model. The trajectory optimizer converts the smooth trajectory τ(t) into low-level control commands and sends them to the robot servo driver. The time-optimal smooth trajectory τ(t) that strictly satisfies the robot dynamics constraints is converted into low-level control commands.
[0025] As a further description of the above technical solution:
[0026] The trajectory optimizer uses one of the following spline interpolation algorithms: cubic spline, B-spline, and Model Predictive Control (MPC).
[0027] As a further description of the above technical solution:
[0028] The robot dynamics constraints include velocity, acceleration, and jerk.
[0029] Compared with the prior art, the advantages of this invention are:
[0030] (1) This scheme completely hides the inference delay of the AI model through the "look-ahead inference" and "asynchronous triggering" mechanisms, eliminates all pauses in the motion process, and achieves true continuity; the innovative "weight fusion" mechanism effectively alleviates the prediction uncertainty or jump between two independent inferences, making the connection of the action sequence natural and smooth, and improving the stability and reliability of the motion.
[0031] (2) This scheme combines the intelligence of AI with the robustness of classical control through a hierarchical planning architecture. The resulting trajectory is not only intelligent, but also smooth, efficient and safe, completely eliminating the jitter phenomenon. As a hardware and software middleware, this architecture does not depend on a specific AI model or robot platform and has good versatility and integrability. Attached Figure Description
[0032] Figure 1 This is a schematic diagram illustrating the principle of robot path smoothing control in this invention;
[0033] Figure 2 This is a timing flowchart of the overlapping look-ahead and weighted fusion mechanism of the present invention;
[0034] Figure 3 This is a schematic diagram illustrating the hierarchical planning principle of the present invention. Detailed Implementation
[0035] The technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention;
[0036] Please see Figure 1-3The present invention provides an embodiment 1 based on an asynchronous control architecture of overlapping look-ahead inference and trajectory fusion:
[0037] An artificial intelligence-based robot path smoothing control method includes: enabling the robot to execute instructions smoothly through asynchronous parallelism and data fusion, with the specific workflow as follows:
[0038] The parameters are defined as follows: two key parameters are defined: the look-ahead step size L and the trigger threshold T, where T < L;
[0039] Initialization and first inference: The system starts up, the AI model performs its first inference, and generates an action sequence of length L. And send it to the execution buffer;
[0040] Execution and asynchronous triggering: The robot executes the first T actions (A1, A2, A3) in the buffer. When the Tth action begins to execute, the next model inference is triggered asynchronously.
[0041] Trajectory fusion, when the robot completes its execution After the Lth action, it does not immediately jump to A7' in the text, but the system in The LT step and Between the first LT steps, linear weighted fusion is performed to generate a new set of smoothly transitioned fused actions;
[0042] The fusion formula, taking step i as an example:
[0043] ;
[0044] The weighting factor α gradually changes from 1 to 0, where 1 indicates complete trust in the old sequence and 0 indicates complete trust in the new sequence. The innovative "weight fusion" mechanism effectively alleviates the prediction uncertainty or jump between two independent inferences, making the connection between action sequences natural and smooth, and improving the stability and reliability of the motion.
[0045] The process continues in a loop, with the robot continuing to execute the fused actions A4'' to A6'. At the same time, the system triggers reasoning for subsequent sequences; this process repeats continuously, achieving a smooth and uninterrupted flow of instructions.
[0046] For example, if the lookahead step size L is set to 6 and the trigger threshold T is 3, satisfying T < L, then the action sequence generated by the AI model during its first inference is: And send it to the buffer;
[0047] The robot executes the first T actions (A1, A2, A3) in the buffer. When T=3 actions begin execution, the next model inference is asynchronously triggered. This inference uses the latest environment observations as input to calculate the action sequence for the next L steps. When the T-th action begins execution, the next model inference is asynchronously triggered. This inference uses the latest environmental observations as input to calculate the action sequence for the next L steps. ;
[0048] When the robot finishes its task The A6 action in the sequence does not immediately jump to... It's not A7' in the system, but rather the system in... The LT step after and The front LT step Linear weighted fusion is performed between these elements to generate a new set of smoothly transitioning fused actions.
[0049] The robot continues to execute the fused actions A4'' to A6''. At the same time, the system triggers inference on subsequent sequences, such as A7' to A12' of Seq_new. This process repeats continuously, achieving a smooth and uninterrupted instruction flow.
[0050] By employing "look-ahead reasoning" and "asynchronous triggering" mechanisms, the inference latency of the AI model is completely hidden, eliminating all pauses in the motion process and achieving true continuity. As a hardware-software middleware, this architecture does not depend on a specific AI model or robot platform and has good versatility and integrability.
[0051] Please see Figure 1-3 Based on Example 1, the present invention also includes Example 2:
[0052] To address the issue of robot motion jitter caused by sparse commands in the control system, this invention employs a hierarchical motion planning mechanism to decouple high-level decision-making from low-level execution, ensuring the physical feasibility of the final motion.
[0053] The outer ring, or AI task planning layer, operates at a low frequency, consistent with the AI model's inference frequency, such as 5-10Hz. It is responsible for macroscopic, sparse path point planning in the task space based on high-level perceptual information (such as vision), and outputs a sequence. This ring encompasses the latency and non-smoothness of the AI model.
[0054] The inner loop, or physical trajectory smoothing layer, operates at a high frequency, consistent with the robot's control frequency (e.g., 100-500Hz), and receives sparse path points output from the outer loop. It incorporates a physical model-based trajectory optimizer that uses spline interpolation algorithms, such as cubic splines, B-splines, or Model Predictive Control (MPC), to generate a trajectory that passes through or approximates a given path. The trajectory optimizer optimizes the time-optimal trajectory τ(t) that strictly satisfies the robot dynamics constraints, namely the smooth trajectory τ(t) that satisfies velocity, acceleration, jerk, etc. It converts sparse and potentially abrupt AI instructions into high-frequency, continuous, and safe low-level control instructions, which are then sent directly to the robot servo driver.
[0055] By combining the intelligence of AI with the robustness of classical control through a "layered planning" architecture, the resulting trajectory is not only intelligent, but also smooth, efficient, and safe, completely eliminating jitter.
[0056] The above description is merely a preferred embodiment of the present invention; however, the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and its improved concepts, should be covered within the scope of protection of the present invention.
Claims
1. A method for controlling a path of a robot based on artificial intelligence, the method comprising: Comprise: Through asynchronous parallel and data fusion, the robot has executable instructions, and the instruction transition is smooth, the specific workflow is as follows: Parameter definition, define two key parameters: look-ahead step L and trigger threshold T, and T < L; Initialization and first inference, system startup, the AI model performs the first inference, generates an action sequence of length L and into the execution buffer; Execution and asynchronous trigger, the robot executes the first T actions (A1, A2, A3) in the buffer, and triggers the next model inference when the Tth action starts to execute; Trajectory fusion, when the robot finishes the Lth action in , it does not immediately jump to A7' in , but the system performs a linearly weighted fusion between the last L-T steps of and the first L-T steps of , generating a new set of fused actions that are smooth and transitional. Fusion formula, take the ith step as an example: Where the weight factor α gradually changes from 1 to 0, 1 represents complete trust in the old sequence, and 0 represents complete trust in the new sequence; Loop forward, the robot continues to execute the fused actions A4'' to A6', while the system triggers the inference of the subsequent sequence; This process is repeated, realizing continuous and smooth instruction flow. 2.The robot path smoothing control method based on artificial intelligence of claim 1, wherein: The robot motion planning decouples high-level decision-making from low-level execution, and the hierarchical motion planning is divided into outer and inner rings. The outer ring is the AI task planning layer, and the inner ring is the physical trajectory smoothing layer. 3.The AI-based robot path smoothing control method of claim 2, wherein: The outer loop runs at low frequency and performs macro, sparse path point planning of the task space based on high-level perception information, outputting a sequence of sparse path points .
4. The robot path smoothing control method based on artificial intelligence according to claim 3, characterized in that: The outer ring runs at a low frequency consistent with the AI model inference frequency, and the high-level perception information includes vision and touch.
5. The robot path smoothing control method based on artificial intelligence according to claim 3, characterized in that: The inner loop runs at high frequency, receiving the sparse path points output by the outer loop The inner loop contains a trajectory optimizer based on physical model, which generates a smooth trajectory τ(t) passing through or approaching The trajectory optimizer converts the smooth trajectory τ(t) into underlying control commands and sends them to the robot servo driver.
6. The robot path smoothing control method based on artificial intelligence according to claim 5, characterized in that: The spline interpolation algorithm used by the trajectory optimizer uses one of cubic spline, B-spline and model predictive control (MPC).
7. The robot path smoothing control method based on artificial intelligence according to claim 5, characterized in that: The robot dynamics constraints include velocity, acceleration and jerk.