Robot motion control method, system, and robot using the same

Abstract

This invention discloses a robot motion control method, system, and robot. The method includes: acquiring a robot's historical state sequence, motion hidden feature sequence, and time delay-related information; performing weighted temporal encoding on the historical state sequence based on the time delay-related information to obtain compensated state features; identifying the robot's current task stage to obtain stage features; extracting a summary from the motion hidden feature sequence and generating a gating vector based on the compensated state features and stage features, then filtering to obtain gated motion features; generating scaling modulation parameters and bias modulation parameters based on the gated motion features, compensated state features, and stage features; performing constraint projection on the modulation parameters according to safety constraints to obtain target modulation parameters; and using the target modulation parameters to linearly modulate the motion hidden feature sequence to output the robot's control action. This invention improves the stability, stage adaptability, and operational safety of robot motion control through state delay compensation, stage adaptive gating, and safety constraint projection.

Description

Technical Field

[0001] This invention relates to the field of robot intelligent control technology, and in particular to a robot motion control method, system, and robot. Background Technology

[0002] With the widespread application of robots in tasks such as grasping, contact assembly, surface treatment, and collaborative handling, control systems increasingly need to generate control actions in real time based on multi-source sensing information and task context. Existing technologies already include schemes that construct policy inputs based on current and historical observations and utilize conditional modulation networks to generate robot actions. For example, WO2024178241A1 discloses a robot control scheme where the policy input can include the current observation image, earlier observation images, and target localization output. The encoding network can also employ a FiLM conditional layer, and the policy network outputs action values ​​for each dimension of the robot's actions. Therefore, generating robot actions based on multimodal, temporal observation information is a known technical approach in existing intelligent robot control.

[0003] However, in practical robotic systems, historical state information does not necessarily correspond precisely to the current control moment. The paper "Measuring and Modelling Delays in Robot Manipulators for Temporally Precise Control using Machine Learning" points out that both execution and response delays in robot control affect precise timing control. Response delays cause the robot's perceived state to lag behind its actual state, resulting in a discrepancy between state perception and the actual motion state. Furthermore, experiments on platforms such as the KUKA KR 5 Sixx and UR10 show that significant delays do indeed exist in standard control loops. In other words, when a control system directly uses historical states for action generation without compensating for their timeliness, a mismatch between the historical state and the current action calculation is likely to occur.

[0004] On the other hand, for contact-type robotic tasks such as wiping, peeling, cutting, and assembly, the reliance on control information varies at different stages. Literature such as *Force-Aware Reactive Policy for Contact-Rich Robotic Manipulation* points out that these tasks typically include both non-contact and contact phases. Existing methods often continuously integrate force / torque and visual information throughout the operation, neglecting the fact that force / torque information primarily plays a role in the contact phase. Furthermore, in the non-contact phase, force / torque sensor noise can interfere with strategy decisions. Therefore, using a uniform motion feature processing and modulation method for different task stages can easily introduce unnecessary disturbances during stage switching or contact establishment, affecting the stability and accuracy of the motion output.

[0005] In summary, while existing technologies have disclosed solutions for robot motion generation based on time-series observation, robot control delay modeling, and stage-related fusion for contact tasks, they primarily address single aspects and lack a unified technical solution that organically combines historical state delay compensation, task stage identification, and motion feature modulation during robot motion control. Especially when robots perform contact or dynamically switching tasks, existing technologies are prone to motion oscillations, contact overshoot, and accumulated trajectory deviations due to mismatches between historical states and current control moments, and the use of uniform processing methods across different task stages, further compressing the robot's safety margin. Therefore, it is necessary to propose a new robot motion control technology to address these issues. Summary of the Invention

[0006] To address the issues in existing technologies where robots performing tasks such as grasping, contact assembly, surface treatment, and other dynamically switching tasks often experience time delay mismatches between historical state information and the current control moment, and where different task stages have significantly different motion control requirements, and where the motion generation process lacks unified constraints on safe operating boundaries, leading to problems such as motion oscillation, contact overshoot, accumulated trajectory deviations, and insufficient safety margins, this invention provides a robot motion control method, system, robot, and computer-readable storage medium to achieve adaptive modulation of robot motion characteristics, thereby improving the stability of robot motion output, adaptability to stage switching, and control safety during complex tasks.

[0007] To achieve the above objectives, the present invention adopts the following technical solution:

[0008] This invention provides a robot motion control method, comprising: acquiring a robot historical state sequence, a motion hidden feature sequence, and time delay related information corresponding to the robot historical state sequence; performing weighted temporal coding on the robot historical state sequence based on the time delay related information to obtain compensated state features; identifying the task stage of the robot based on robot operation information to obtain stage features; extracting a summary from the motion hidden feature sequence to obtain a motion summary vector, and generating a gating vector based on the compensated state features and the stage features; filtering the motion summary vector using the gating vector to obtain gated motion features; generating motion feature modulation parameters based on the gated motion features, the compensated state features, and the stage features, wherein the motion feature modulation parameters include at least scaling modulation parameters and bias modulation parameters; performing constraint projection on the motion feature modulation parameters according to the robot's safety constraints to obtain target modulation parameters; linearly modulating the motion hidden feature sequence using the target modulation parameters to obtain a modulated motion feature sequence, and outputting robot control actions based on the modulated motion feature sequence. By using the above technical solutions, historical state compensation, task stage identification, action gating screening, modulation parameter generation, and safety constraint correction are organically linked into a unified action control link, thereby reducing the risk of state mismatch during robot action generation and improving the stability, stage adaptability, and safety of action output.

[0009] Preferably, the delay-related information includes timestamp information, sampling interval information, state reliability information, and / or frame loss marker information. The step of weighted temporal encoding of the robot's historical state sequence based on the delay-related information to obtain compensated state features specifically includes: determining a delay attenuation factor based on the time difference between the acquisition time of state data at each moment and the current control time; determining a reliability correction factor based on the sampling interval information, state reliability information, and / or frame loss marker information; determining the state weight corresponding to the state data at each moment based on the delay attenuation factor and the reliability correction factor; and weighted aggregation of the encoding results of the robot's historical state sequence based on the state weights to obtain the compensated state features. This allows the contribution of historical states to current action control to match their timeliness and reliability, thereby reducing the interference of lagging states, abnormal sampling states, or low-reliability states on the action generation results and improving the temporal consistency during the control process.

[0010] Preferably, the weighted temporal encoding of the robot's historical state sequence includes: inputting the robot's historical state sequence into a temporal encoding network to extract temporal correlation features from the historical states; and then performing weighted summation, weighted averaging, or end-state selection on the encoding results according to the state weights corresponding to the state data at each time point to obtain the compensated state features. The temporal encoding network can be a recurrent neural network encoder, which can be a gated recurrent unit encoder and / or a long short-term memory network encoder. By encoding the temporal correlations in the historical states and combining them with state weights for compensation and aggregation, the ability to represent state evolution trends can be further improved, thereby providing a state basis more consistent with the current control time for subsequent motion feature modulation.

[0011] Preferably, the step of identifying the task stage of the robot based on robot operation information and obtaining stage characteristics includes: identifying the task stage of the robot based on at least one of the following: end effector speed, target distance change rate, contact trigger signal, force / torque change rate, and drive current change; the task stage includes at least two of the following: free motion stage, target approach stage, contact establishment stage, contact holding stage, and release stage; and generating stage identifiers and / or stage weights as stage characteristics based on the identification results. By introducing a task stage identification mechanism, the robot can adopt different motion modulation criteria at different stages, thereby avoiding the use of a uniform motion control strategy for the entire task process, which helps to reduce motion oscillations, contact overshoot, and trajectory deviation accumulation during stage switching.

[0012] Preferably, the step of extracting a summary of the hidden action feature sequence to obtain an action summary vector includes performing mean pooling, max pooling, or weighted pooling along the time dimension on the hidden action feature sequence to form an action summary vector reflecting the overall expression of the current action sequence. The step of generating a gating vector based on the compensated state features and the stage features includes mapping the compensated state features and the stage features to a preset projection space and inputting them into a gating parameter generation network, outputting a gating vector that corresponds one-to-one with each feature dimension of the action summary vector; then multiplying the gating vector with the action summary vector dimension by dimension to obtain the gated action features. Therefore, based on the current compensated state and task stage, action components related to the current control objective can be selectively strengthened, while inappropriate action components can be suppressed, thereby improving the matching degree between the action feature expression and the current task requirements.

[0013] Preferably, the step of generating motion feature modulation parameters based on the gated motion features, the compensated state features, and the stage features includes: concatenating and fusing the gated motion features, the compensated state features, and the stage features through a linear mapping fusion to obtain joint features; inputting the joint features into a scaling parameter generation branch and an offset parameter generation branch respectively, and outputting the scaling modulation parameters and the offset modulation parameters, wherein the scaling parameter generation branch and the offset parameter generation branch can be implemented using a multilayer perceptron. By combining the gated motion information, the compensated state information, and the stage information for modulation parameter generation, the modulation parameters can simultaneously reflect the robot's current state, historical evolution trend, and task stage requirements, thereby improving the targeting and adaptability of motion modulation.

[0014] Preferably, the step of constraining and projecting the motion feature modulation parameters according to the robot's safety constraints to obtain target modulation parameters includes: acquiring constraint indices reflecting the robot's safe operating envelope, wherein the constraint indices include at least one of joint limit margin, drive torque margin, end-effector contact force upper limit, jerk upper limit, and singularity index; determining the allowable range of the scaling modulation parameters and the bias modulation parameters according to the constraint indices; and restricting the scaling modulation parameters and the bias modulation parameters within the corresponding allowable ranges using truncation, cropping, or normalized projection methods to obtain the target modulation parameters. By introducing a constraint projection process after the modulation parameter output, the motion features can be avoided from being over-amplified or experiencing unreasonable deviations, thereby reducing the risk of control actions exceeding the robot's safe operating envelope and improving the safety and engineering feasibility of the control process.

[0015] Preferably, the linear modulation of the motion hidden feature sequence using the target modulation parameters includes: extending the target modulation parameters to each time step of the motion hidden feature sequence, and performing dimension-wise scaling and translation modulation on the feature vector of each time step to obtain the modulated motion feature sequence. Further, joint position commands, joint velocity commands, joint torque commands, and / or end effector control commands can be output from the modulated motion feature sequence via a motion decoding network, a control head network, or a control mapping function as the robot control actions. Thus, the aforementioned state compensation, stage recognition, and safety constraint correction results can be ultimately implemented at the robot control output layer, thereby ensuring that the robot motion generation results balance control accuracy, stage adaptability, and safety boundary requirements.

[0016] This invention also provides a robot motion control system, including a data acquisition module, a state compensation encoding module, a stage recognition module, a gating screening module, a modulation parameter generation module, a constraint projection module, and a motion output module. These modules cooperate to execute the aforementioned robot motion control method. Through systematic module configuration, this invention can be easily deployed on various robot platforms such as robotic arms, collaborative robots, and mobile robots, improving the engineering implementation convenience of the solution.

[0017] This invention also provides a robot, including a sensor assembly, an actuator, and a controller. The controller is configured to execute the aforementioned robot motion control method to generate robot control actions and control the actuator's movements. By applying this invention to a specific robot body, the stability and safety of the robot's motion control in complex tasks such as grasping, assembling, wiping, and polishing can be improved.

[0018] The present invention also provides a computer-readable storage medium having a computer program stored thereon. When the computer program is executed by a processor, it implements the above-described robot motion control method, thereby facilitating the deployment of the present invention in software form on existing robot control platforms.

[0019] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0020] This invention introduces a time-delay-related historical state weighted timing coding mechanism, which enables historical states to be compensated based on time lag, sampling stability, reliability, and frame loss before participating in the generation of the current action. This reduces the risk of mismodulation caused by the mismatch between historical states and the current control time, and improves the timing consistency and stability of action control.

[0021] This invention identifies the task stage of the robot and incorporates the stage characteristics into the motion gating and modulation parameter generation process. This allows the motion feature processing method to adapt to different stages such as free movement, approaching the target, contact establishment, contact maintenance or release. This improves the problems of motion oscillation, contact overshoot and trajectory deviation accumulation during stage switching, and enhances the robot's adaptability to contact-type and dynamically switching tasks.

[0022] This invention generates a gating vector by utilizing compensated state features and stage features, selectively filters the action summary vector, and then jointly generates scaling modulation parameters and bias modulation parameters based on the gating action features, state features, and stage features. This enables the action hidden feature sequence to obtain a dimension-wise modulation result that better matches the current control context, thereby improving the consistency between action feature representation and the robot's real-time control requirements.

[0023] This invention further introduces a safety constraint projection mechanism after the modulation parameter output, incorporating robot operation safety boundaries such as joint limit margin, drive torque margin, end contact force upper limit, jerk upper limit, and singularity index into the modulation parameter correction process. This can effectively prevent the control action from exceeding the safe operation envelope due to over-modulation of motion characteristics, thereby improving the safety and feasibility of the robot control process.

[0024] Furthermore, this invention organically links historical state compensation, stage recognition, action gating, modulation parameter generation, and safety constraint correction into a unified action control link. It is not only applicable to contact-rich tasks such as robotic arm grasping, assembly, wiping, and polishing, but also to other robot control scenarios with time-dependent and multi-source state inputs, demonstrating good versatility and engineering application value. Attached Figure Description

[0025] Figure 1 This is a schematic diagram of the process steps of the present invention;

[0026] Figure 2 This is a diagram illustrating the system module composition and data flow logic of the present invention.

[0027] Figure 3 This is a schematic diagram of the process and sub-steps of steps S1 to S4 of the present invention;

[0028] Figure 4 This is a schematic diagram of the process and sub-steps of steps S5 to S7 of the present invention. Detailed Implementation

[0029] The present invention will be further described in detail below with reference to embodiments. It should be understood that the following embodiments are only used to explain the technical solutions of the present invention and are not intended to limit the scope of protection of the present invention. Equivalent substitutions, simple modifications, or conventional adjustments made by those skilled in the art based on the content of this specification without departing from the concept of the present invention should all fall within the scope of protection of the present invention.

[0030] In this invention, the "delay-related information" includes at least timestamp information, sampling interval information, state reliability information, and / or frame loss marker information; the "state weight" is determined by a combination of the delay attenuation factor and reliability correction factor corresponding to the state data at each time step; the "linear modulation" refers to performing dimension-wise scaling and translation processing on the feature vector of each time step of the action hidden feature sequence using target scaling modulation parameters and target bias modulation parameters; the "constraint projection" refers to the process of determining the allowable range of modulation parameters according to robot safety constraints and limiting the modulation parameters to the allowable range through cropping, truncation, or normalized projection.

[0031] In some implementations, the compensated state features and stage features can be first mapped to a preset projection space, and then used for gating vector generation and modulation parameter generation, respectively; the scaling parameter generation branch and the bias parameter generation branch are modeled using independent network parameters. Further, the robot motion control system may include a data acquisition module, a state weight determination module, a state compensation encoding module, a stage recognition module, a motion summary extraction module, a gating generation and filtering module, a modulation parameter generation module, a constraint projection module, and a motion modulation output module. Each module can be implemented through software, hardware, or a combination of both.

[0032] Example 1

[0033] This embodiment uses a six-degree-of-freedom robotic arm performing a plug-in assembly task as an example to describe the method of the present invention in detail. The robotic arm includes a joint encoder, a vision sensor, a six-dimensional force / torque sensor, a current sensor, a servo driver, and a controller. The controller can be an industrial controller, an embedded processor, or an edge computing device with a graphics processing unit, and the control cycle is preferably 10 to 20 milliseconds. When the robotic arm performs a plug-in assembly task, it generally goes through a free movement phase, a target approach phase, a contact establishment phase, a contact holding phase, and a release phase in sequence. The focus of motion control differs in each phase, making this a suitable application scenario for the present invention.

[0034] As shown in the figure, at the current control moment, the data acquisition module first collects the robot's historical state sequence, action hidden feature sequence, and time delay related information corresponding to the historical state sequence.

[0035] The historical state sequence consists of state data from several consecutive sampling moments prior to the current control moment. Taking a control cycle of 20 milliseconds and a historical window length of 10 cycles as an example, the historical state sequence covers the state changes within the most recent 200 milliseconds. The state data at each moment can include one or more of the following: joint angle, joint angular velocity, drive current, end effector position, end effector attitude, end effector velocity, force / torque sensor output, end effector pose error relative to the target hole, and contact marker. In this embodiment, six joint angles, six joint angular velocities, six drive currents, six-dimensional end effector pose error, six-dimensional force / torque value, and one contact marker can be concatenated to form a historical state vector. For scenarios with visual perception, the target hole relative pose features extracted by the visual network can also be incorporated into the state vector.

[0036] The action hidden feature sequence is output by the upstream action generation network. In this embodiment, the upstream action generation network can be an imitation learning policy network, a diffusion action policy network, or a multimodal policy network. The action hidden feature sequence is used to characterize the feature representation of the action to be output at several future control moments, rather than direct joint commands. The action hidden feature sequence preferably contains action feature vectors for multiple time steps, such as 8 or 10 time steps, and the action feature dimension of each time step can be set to 64, 128, or 256 dimensions.

[0037] Delay-related information corresponding to historical state sequences is used to characterize the timeliness and reliability of state data at each moment. In this embodiment, the delay-related information includes at least the following: timestamp information corresponding to each historical state, sampling interval information between adjacent samples, confidence information of the corresponding state, and frame loss marker information indicating whether frame loss has occurred. The timestamp information describes the state acquisition time; the sampling interval information reflects whether the sampling of each sensor is stable; the confidence information can be given by the sensor self-test results, anomaly detection results, or the confidence level output by the sensing network; the frame loss marker information indicates whether a certain state data is filled by historical values, interpolated, or has a transmission interruption. Since the update frequency and transmission delay of the visual sensor, force sensor, encoder, and current acquisition link are different, the above-mentioned delay-related information provides a basis for subsequent historical state compensation.

[0038] In this embodiment, the system establishes the following control link: first, input data is acquired, and then the state weight determination module and state compensation encoding module complete historical state compensation; then, the stage identification module identifies the task stage; then, the action summary extraction module and the gating generation and filtering module complete the action summary generation and action feature filtering; next, the modulation parameter generation module outputs the action feature modulation parameters; then, the constraint projection module corrects the modulation parameters; finally, the action modulation output module completes linear modulation and outputs the robot control action.

[0039] S2. After obtaining the historical state sequence and corresponding time delay information, the state weight determination module first determines the state weight of each historical state data, and then the state compensation coding module generates the compensated state features.

[0040] Specifically, for each moment's state data in the historical window, a delay attenuation factor is first determined based on the time difference between its acquisition time and the current control time. State data closer to the current control time has a larger delay attenuation factor; state data farther from the current control time has a smaller delay attenuation factor. This setting aims to highlight state information closer to the current moment and reduce the adverse effects of older states on the generation of current actions.

[0041] Subsequently, the reliability correction factor is determined based on the stability of the sampling interval, the reliability of the state data, and the frame loss situation. If the state data at a certain moment comes from normal sampling, the sampling period is stable, the reliability is high, and there are no frame losses, then its reliability correction factor can be set to a higher value; if the state data at a certain moment has a significant delay, the sampling interval deviates from the nominal control period, the perceived confidence is reduced, or there are frame losses, then its reliability correction factor is reduced.

[0042] Next, the delay attenuation factor and the reliability correction factor are combined to obtain the state weights of the state data at the corresponding time step. The combination method can be a product method or a weighted summation method, for example, using the state weights as the attention weights output by the encoder at each time step and performing a weighted summation. In this embodiment, a product method is preferred to reflect both the timeliness and reliability aspects simultaneously.

[0043] After obtaining the state weights, the historical state sequence is input into the temporal coding network. In this embodiment, a gated recurrent unit network (GRU) is preferably used as the temporal coding network; in other embodiments, a long short-term memory (LSTM) network can also be used. The temporal coding network is used to extract temporal correlation features from the historical states, such as the trend of gradually decreasing end-effector velocity, the trend of rapidly increasing contact force, and the trend of gradually converging target pose error. The hidden states at each time step output by the temporal coding network are then weighted and aggregated with the aforementioned state weights to obtain the compensated state features. The weighted aggregation can be implemented by weighted summation, weighted averaging, or by combining the state weights with the corrected end-effector hidden states as the output.

[0044] The compensated state features obtained by the above method can more accurately reflect the real state that can be referenced at the current control moment compared to directly averaging the historical states or directly taking the state of the last frame. This reduces the risk of state mismatch caused by visual link delay, local sampling anomalies and frame loss.

[0045] S3. After obtaining the compensated state characteristics, the stage identification module identifies the current task stage and outputs the stage characteristics.

[0046] In this embodiment, the task phase includes at least two of the following: free movement phase, target approach phase, contact establishment phase, contact holding phase, and release phase, preferably including all five phases. Phase identification can be determined based on one or more of the following: end effector speed, rate of change of distance between the end effector and the target hole, contact trigger signal, rate of change of force / torque, and change of drive current.

[0047] For example, when the robotic arm is still far from the target hole, the contact marker is not in contact, and the end effector speed is higher than the preset slow speed threshold, it can be determined as the free movement stage; when the robotic arm has entered the area near the target hole and the speed begins to decrease, it can be determined as the approaching target stage; when the contact marker appears for the first time and the force sensor detects a significant increase in contact force, it can be determined as the contact establishment stage; when contact continues and the contact force fluctuates stably within a certain range, it can be determined as the contact holding stage; when contact is released and the end effector begins to withdraw from the insertion area, it can be determined as the release stage.

[0048] Stage features can be output as stage identifiers, stage weights, or a combination of both. In this embodiment, a one-hot encoded stage identifier is preferred; in some implementations, the confidence level of the corresponding stage can also be output and used as the stage weight. If a lightweight stage classification network is used, the stage category and its confidence level can be directly output by the classification network.

[0049] By setting up a phase recognition process, subsequent motion screening and motion modulation can distinguish whether the robotic arm is currently in a large-scale movement phase or a fine contact phase, thereby avoiding the use of the same motion processing method for different phases.

[0050] S4. After obtaining the compensated state features and stage features, the action summary extraction module and the gating generation and filtering module complete the action summary vector generation and action filtering.

[0051] First, the action hidden feature sequence is pooled to obtain an action summary vector. The pooling process can employ mean pooling, max pooling, or weighted pooling. In this embodiment, mean pooling is preferred to form an action summary vector that reflects the overall trend of the current action sequence. When the task focuses more on certain action steps closer to the current time step, time-weighted pooling can also be used to increase the contribution of action features from later time steps to the action summary vector.

[0052] Then, the compensated state features and stage features are mapped to a preset projection space. The purpose of the preset projection space is to make the state-side information and the action summary vector more closely match in terms of dimension and feature scale, facilitating subsequent gating generation. Next, the mapped and compensated state features and stage features are input into the gating parameter generation network, which outputs gating vectors that correspond one-to-one with each feature dimension of the action summary vector. The gating parameter generation network can employ a one-layer or multi-layer perceptron structure, and an activation function is used at the output to constrain the value range of each dimension of the gating vector, ensuring it falls within a preset interval.

[0053] Next, the gating vector is applied to the action summary vector to obtain the gated action features. This process can be understood as selectively preserving, enhancing, or suppressing different feature dimensions of the action summary vector. Taking a plug-in assembly task as an example, during the approach phase, the gating vector can increase the weight of action components related to end-effector attitude adjustment and low-speed approach; during the contact establishment phase, the gating vector can increase the weight of action components related to minor position correction and contact buffering; and during the free movement phase, it can reduce the weight of some action components related to contact fine-tuning but not currently important. Through this filtering process, the subsequent modulation parameter generation can be more focused on the action trends truly relevant to the current state and the current task stage.

[0054] S5. After obtaining the gated motion characteristics, the modulation parameter generation module generates motion characteristic modulation parameters. The motion characteristic modulation parameters include at least scaling modulation parameters and bias modulation parameters.

[0055] Specifically, the gated action features, compensated state features, and stage features are first fused to obtain joint features. The fusion method can be direct concatenation or linear mapping followed by fusion. In this embodiment, concatenation fusion is preferred to preserve the integrity of the three types of information. Then, the joint features are input into the scaling parameter generation branch and the bias parameter generation branch, respectively. The two branches are modeled using independent network parameters to ensure that the scaling modulation parameters and bias modulation parameters can learn different modes of operation. Both the scaling parameter generation branch and the bias parameter generation branch can be implemented using a multilayer perceptron.

[0056] The scaling modulation parameter controls the enhancement or reduction magnitude of each feature dimension of the action hidden feature sequence, while the bias modulation parameter applies state-related and stage-related offset corrections to each feature dimension of the action hidden feature sequence. By simultaneously generating the scaling modulation parameter and the bias modulation parameter, the action feature modulation can possess both gain adjustment capability and offset compensation capability, thereby improving the flexibility and adaptability of the modulation.

[0057] S6. After outputting the scaling modulation parameters and the bias modulation parameters, the constraint projection module performs constraint projection on the modulation parameters according to the robot safety constraints to obtain the target scaling modulation parameters and the target bias modulation parameters.

[0058] In this embodiment, the safety constraint indicators may include one or more of the following: joint limit margin, drive torque margin, end-effector contact force limit, acceleration or jerk limit, and singularity indicator. The joint limit margin is used to prevent the robotic arm from outputting large-amplitude movement commands when the joints are near their limit positions; the drive torque margin is used to avoid motor overload; the end-effector contact force limit is used to prevent excessive contact impact during the contact establishment and contact holding phases; the acceleration or jerk limit is used to suppress high-frequency abrupt movements; and the singularity indicator is used to reduce the unstable modulation tendency when the robotic arm approaches singular configurations.

[0059] After determining the safety constraints, the allowable ranges for scaling and bias modulation parameters are determined based on these constraints. These allowable ranges can be fixed or dynamically adjusted according to the current stage and safety margins. For example, when the robotic arm is in a free-movement phase and the safety margins are large, the allowable range for modulation parameters can be relatively loose; when the robotic arm is in the contact establishment phase and the upper limit of the contact force is low, the allowable range for modulation parameters related to rapid approach and significant correction is correspondingly compressed; when the robotic arm approaches a singular configuration, some high-gain modulation parameters are tightened overall.

[0060] Subsequently, by employing cropping, truncation, or normalized projection methods, the original scaling modulation parameters and original bias modulation parameters are constrained within their respective allowable ranges to obtain the target scaling modulation parameters and target bias modulation parameters. This process binds the motion feature modulation to the robot's actual executable boundaries, preventing the modulation trends directly output by the network from causing boundary violations at the physical execution level.

[0061] S7. After obtaining the target scaling modulation parameters and the target bias modulation parameters, the motion modulation output module performs linear modulation on the motion hidden feature sequence and outputs the robot control motion.

[0062] Specifically, the target scaling modulation parameters and target bias modulation parameters are first extended to each time step of the action hidden feature sequence. Then, a dimensional scaling and translation process is performed on the action feature vector at each time step. For any time step and any feature dimension, its modulated value is equal to the original value of that feature dimension multiplied by one, summed with the corresponding target scaling modulation parameter, and then added to the corresponding target bias modulation parameter. After completing the linear modulation of the action features at all time steps, the modulated action feature sequence is obtained.

[0063] The modulated motion feature sequence is then input into a motion decoding network, a control head network, or a control mapping function to output robot control actions. These actions can be one or more of the following: joint position increment commands, joint velocity commands, joint torque commands, end-effector pose adjustment commands, and gripper control commands. The controller then sends these control actions to the servo driver, which drives the robotic arm to execute the corresponding actions.

[0064] Through the above process, this embodiment realizes a complete control path from input acquisition, historical state compensation, stage identification, action filtering, modulation parameter generation, safety constraint correction to control action output.

[0065] To further illustrate the execution process of this invention, a specific execution scenario example is given below for further explanation.

[0066] Assume a robotic arm is performing an assembly task involving inserting a pin into a hole. The control cycle is 20 milliseconds, the historical state window length is 10, the motion hidden feature sequence length is 8, and the motion hidden feature dimension is 128. In the most recent 10 frames of historical states before the current moment, the joint encoder and current acquisition link are updated normally, but the vision link has a delay of approximately 60 milliseconds due to image processing time, and one frame of image features has low confidence. At this time, the state weight determination module identifies a large time difference between the visually relevant historical states and the current control moment based on timestamp information, and identifies one frame of visual features with low confidence based on confidence information, thus reducing the state weight of these frames accordingly. Meanwhile, the most recent frames of states from the joint encoder and force sensor have higher state weights due to their smaller time delay and stable sampling. After time-series encoding and weighted aggregation, the resulting compensated state features are more biased towards the robotic arm's current true end-effector position, speed, and contact trend, without being excessively interfered with by delayed visual information.

[0067] In this example, the robotic arm has entered the vicinity of the hole, the distance between its end effector and the center of the hole has decreased to within 10 mm, the contact marker has not yet been triggered, and the end effector speed has dropped below the preset slow speed threshold. Therefore, the stage recognition module determines the current task stage as the approaching target stage and outputs the corresponding stage features. The action summary extraction module performs mean pooling on the action hidden feature sequence to obtain an action summary vector. The gating generation and filtering module generates a gating vector based on the compensated state features and the approaching target stage features. This gating vector increases the weight of action feature dimensions related to attitude fine-tuning and low-speed approach, and decreases the weight of action feature dimensions related to large swings and rapid movements, resulting in gated action features.

[0068] Subsequently, the modulation parameter generation module fuses the gated motion characteristics, compensated state characteristics, and stage characteristics, outputting scaling modulation parameters and bias modulation parameters respectively. At this point, since the robotic arm has not yet contacted the hole edge, the upper limit of the end-effector contact force in the safety constraint indicators has not yet played a dominant role. The constraint projection module only applies slight restrictions to the modulation parameters based on the joint limit margin and drive torque margin. The motion modulation output module completes linear modulation based on the target modulation parameters, and the motion decoding network outputs the joint position fine-tuning command in the next control window, allowing the robotic arm to continue approaching the target hole at a relatively low speed.

[0069] When the robotic arm's end effector first contacts the edge of the hole, the force / torque sensor detects a rapid increase in contact force, and the contact indicator changes from no contact to contact. The stage identification module then switches the task stage to the contact establishment stage. At this time, because the contact establishment stage has higher requirements for impact suppression, the stage characteristics change. The gating vector correspondingly increases the weight of motion feature dimensions related to minor position corrections and contact buffering, and further suppresses the weight of motion feature dimensions related to rapid propulsion. At the same time, the constraint projection module significantly tightens the allowable range of some modulation parameters based on the upper limit of the end effector contact force, the drive torque margin, and the singularity index of the robotic arm's current posture, making the overall target modulation parameters more conservative. The motion modulation output module then generates fine-tuning control commands with smaller amplitudes based on the target modulation parameters, prompting the robotic arm to make minor corrections along the edge of the hole, reducing the risk of contact overshoot.

[0070] When the robotic arm has entered a stable insertion state and the contact force fluctuation remains within the allowable range for several consecutive control cycles, the stage identification module identifies the task stage as the contact holding stage. At this time, the gating vector retains more motion feature dimensions related to stable holding, slight compensation, and force control. The overall amplitude of the scaling modulation parameters output by the modulation parameter generation module is further reduced, while the bias modulation parameters are mainly used to compensate for minor assembly errors. The constraint projection module continues to control the parameter range based on the upper limit of the contact force and the joint margin. The final output robot control action is a small-amplitude, stable, and continuous holding action, avoiding jitter and secondary impacts during the insertion process.

[0071] When assembly is complete and the exit action is executed, the contact flag is released, and the stage identification module switches the task stage to the release stage. At this time, the gating vector will again increase the weight of the motion components related to the smoothness of the evacuation path, and the modulation parameter range can also be widened accordingly, allowing the robotic arm to smoothly exit the target area. Therefore, this invention can dynamically adjust the motion feature processing method according to the stage changes within the same task, and maintain the stability and safety of the output motion through state compensation and safety constraints.

[0072] Example 2

[0073] Based on Embodiment 1, this embodiment illustrates alternative implementations of the present invention in terms of parameters and structure.

[0074] In some implementations, the temporal coding network is not limited to gated recurrent unit networks (GRUs) and can also employ long short-term memory (LSTM) networks. For tasks with longer state dependency spans, LSM networks can retain state evolution information over a longer time range. Action summary extraction methods are not limited to mean pooling; max pooling or weighted pooling can also be used. Weighted pooling is particularly suitable for scenarios that focus more on subsequent action steps. Stage recognition methods can employ threshold rules or lightweight classification networks; when task stage boundaries are ambiguous, classification networks are more beneficial for improving robustness. Constraint projection methods are not limited to simple pruning; dynamic compression of parameter ranges based on safety margins can also be used. When the robot's safety margin is sufficient, a larger modulation amplitude is allowed; when the safety margin is insufficient, the modulation amplitude is automatically reduced. These adjustments do not change the core technical concept of this invention.

[0075] Example 3

[0076] In some implementations, the robot motion control system includes: a data acquisition module, a state weight determination module, a state compensation encoding module, a stage recognition module, a motion summary extraction module, a gating generation and filtering module, a modulation parameter generation module, a constraint projection module, and a motion modulation output module. Specifically, the data acquisition module acquires historical state sequences, motion hidden feature sequences, and time delay-related information; the state weight determination module determines the state weights corresponding to the state data at each time step based on a time delay attenuation factor and a reliability correction factor; the state compensation encoding module obtains the compensated state features; the stage recognition module outputs stage features; the motion summary extraction module generates motion summary vectors; the gating generation and filtering module outputs gating vectors and obtains gating motion features; the modulation parameter generation module generates scaling modulation parameters and bias modulation parameters; the constraint projection module obtains target modulation parameters; and the motion modulation output module executes linear modulation and outputs the robot control motion.

[0077] Example 4

[0078] In some implementations, the robot includes sensor components, actuators, and a controller. The sensor components may include joint encoders, vision sensors, force / torque sensors, current sensors, displacement sensors, or other sensors used to acquire historical robot states and environmental perception information; the actuators may include servo motors, reduction gears, end effectors, wheel drive systems, or other actuators; the controller is configured to execute the aforementioned robot motion control method to generate robot control actions and control the actuator movements.

[0079] Example 5

[0080] In other embodiments, a computer-readable storage medium stores a computer program that is deployed in a robot body controller, an edge computing device, an industrial computer, or a host computer, and is executed by a processor to implement the above-described robot motion control method.

Claims

1. A robot motion control method, characterized in that, include: S1. Obtain the robot's historical state sequence, action hidden feature sequence, and time delay related information corresponding to the robot's historical state sequence; S2. Based on the time delay related information obtained in S1, the robot's historical state sequence is weighted and time-series encoded to obtain the compensated state features; S3. Identify the task stage the robot is in based on the robot's operational information and obtain the stage characteristics; S4. Extract a summary from the action hidden feature sequence to obtain an action summary vector, and generate a gating vector based on the compensated state features and the stage features. Use the gating vector to filter the action summary vector to obtain gated action features. S5. Generate motion feature modulation parameters based on the gated motion features, the compensated state features, and the stage features. The motion feature modulation parameters include at least scaling modulation parameters and bias modulation parameters. S6. Project the motion feature modulation parameters according to the robot's safety constraints to obtain the target modulation parameters; S7. Linearly modulate the motion hidden feature sequence using the target modulation parameters to obtain the modulated motion feature sequence, and output the robot control action based on the modulated motion feature sequence.

2. The robot motion control method according to claim 1, characterized in that, The delay-related information includes timestamp information, sampling interval information, state reliability information, and / or frame loss marker information; The step of weighting and temporally encoding the robot's historical state sequence based on the time delay-related information to obtain compensated state features includes: The delay attenuation factor is determined based on the time difference between the acquisition time of the state data at each moment and the current control time; the reliability correction factor is determined based on the sampling interval information, state reliability information and / or frame loss marking information; the state weight corresponding to the state data at each moment is determined based on the delay attenuation factor and the reliability correction factor; the encoding results of the robot's historical state sequence are weighted and aggregated based on the state weight to obtain the compensated state features.

3. The robot motion control method according to claim 1 or 2, characterized in that, The weighted temporal encoding of the robot's historical state sequence includes: The robot's historical state sequence is input into a recurrent neural network encoder to extract historical state temporal features; the historical state temporal features are then weighted and summed, weighted and averaged, or weighted and selected based on the state weights to obtain the compensated state features. The recurrent neural network encoder is a gated recurrent unit encoder and / or a long short-term memory network encoder.

4. The robot motion control method according to claim 1, characterized in that, The process of identifying the task stage of the robot based on its operational information and obtaining stage characteristics includes: The robot's task stage can be identified by at least one of the following: end effector speed, target distance change rate, contact trigger signal, force / torque change rate, and drive current change. The task phase includes at least two of the following: free movement phase, target approach phase, contact establishment phase, contact maintenance phase, and release phase; The stage identifier and / or stage weight are generated based on the identification results as the stage features.

5. The robot motion control method according to claim 1, characterized in that, The process involves extracting a summary of the hidden action feature sequence to obtain an action summary vector, generating a gating vector based on the compensated state features and the stage features, and then using the gating vector to filter the action summary vector to obtain gated action features, including: The action hidden feature sequence is subjected to mean pooling, max pooling, or weighted pooling along the time dimension to obtain the action summary vector; the compensated state features and the stage features are mapped to a preset projection space and then input into a gating parameter generation network to output a gating vector that corresponds one-to-one with each feature dimension of the action summary vector; the gating vector is multiplied dimension by dimension with the action summary vector to obtain the gated action features.

6. The robot motion control method according to claim 1, characterized in that, The generation of motion feature modulation parameters based on the gated motion features, the compensated state features, and the stage features includes: The gated action features, the compensated state features, and the stage features are spliced ​​and fused or linearly mapped and fused to obtain joint features; the joint features are respectively input into scaling parameters to generate branches and bias parameters to generate branches, and the scaling modulation parameters and the bias modulation parameters are output. The scaling parameter generation branch and the bias parameter generation branch are both implemented using a multilayer perceptron.

7. The robot motion control method according to claim 1, characterized in that, The step of constraining and projecting the motion feature modulation parameters according to the robot's safety constraints to obtain target modulation parameters includes: Obtain constraint indices reflecting the safe operation envelope of the robot. The constraint indices include at least one of joint limit margin, drive torque margin, end contact force upper limit, jerk upper limit, and singularity index. Determine the allowable range of the scaling modulation parameter and the bias modulation parameter based on the constraint indices. Limit the scaling modulation parameter and the bias modulation parameter to the corresponding allowable range using truncation, cropping, or normalized projection methods to obtain the target modulation parameter. The method of linearly modulating the action hiding feature sequence using the target modulation parameters includes: extending the target modulation parameters to each time step of the action hiding feature sequence, and performing dimension-wise scaling and translation modulation on the feature vector of each time step.

8. A robot motion control system, characterized in that, include: The data acquisition module is used to acquire the robot's historical state sequence, action hidden feature sequence, and time delay related information corresponding to the robot's historical state sequence; The state compensation coding module is used to perform weighted temporal coding on the robot's historical state sequence based on the time delay related information to obtain the compensated state features; The stage identification module is used to identify the task stage that the robot is in based on the robot's operation information and obtain stage characteristics. The gating filtering module is used to extract a summary of the action hidden feature sequence to obtain an action summary vector, and generate a gating vector based on the compensated state features and the stage features. The gating vector is then used to filter the action summary vector to obtain gated action features. The modulation parameter generation module is used to generate motion feature modulation parameters based on the gated motion features, the compensated state features, and the stage features; the constraint projection module is used to perform constraint projection on the motion feature modulation parameters according to the robot's safety constraints to obtain target modulation parameters. The motion output module is used to linearly modulate the motion hidden feature sequence using the target modulation parameters to obtain the modulated motion feature sequence, and output robot control actions based on the modulated motion feature sequence.

9. A robot, characterized in that, include: Sensor components are used to collect historical state sequences of the robot and time-delay related information corresponding to the historical state sequences of the robot; An actuator for performing robot control actions; a controller connected to the sensor assembly and the actuator, the controller being configured to acquire a sequence of hidden action features and execute the robot motion control method according to any one of claims 1 to 7 to generate the robot control actions and control the actions of the actuator.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When the computer program is executed by the processor, it implements the robot motion control method according to any one of claims 1 to 7.

Images