Robot motion control methods and devices, robots, media, products

By using a generative motion control model, the trajectory probability distribution information is determined based on the robot's task objectives and state information, which solves the problem of insufficient compliance and robustness of robot motion control and achieves higher adaptability and stability in complex scenarios.

CN122500740BActive Publication Date: 2026-11-10AGIBOT INNOVATION (SHANGHAI) TECHNOLOGY CO LTD
View PDF 2 Cites 0 Cited by

Patent Information

Application Number
CN202610992707.8
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Filing Date
2026-07-06
Publication Date
2026-11-10
Estimated Expiration
2046-07-06

AI Technical Summary

Technical Problem

The robot's motion control lacks compliance, robustness, and adaptability to complex scenarios, making it difficult to meet the needs of practical applications.

Method used

By using a generative motion control model, based on the robot's task target information and state information, the trajectory probability distribution information is determined, and then motion control data is generated to achieve closed-loop control and recovery capability of robot motion.

Benefits of technology

It improves the compliance of robot motion control and its adaptability to complex scenarios, reduces the risk of motion instability, and enhances the robot's recovery ability in non-standard conditions.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN122500740B_ABST
    Figure CN122500740B_ABST
Patent Text Reader

Abstract

This application provides a robot motion control method and device, a robot, a medium, and a product. The method includes: determining trajectory probability distribution information based on the robot's task target information and robot state information using a generative motion control model; wherein the trajectory probability distribution information is used to characterize the probability distribution of the robot's motion trajectory in the target future time domain under the conditions corresponding to the task target information and the robot state information; determining target motion trajectory information based on the trajectory probability distribution information; and generating motion control data for the robot based on the target motion trajectory information to control the robot's motion. This application adopts the concept of modeling trajectory probability distribution information based on task target information and robot state information, thereby improving the compliance, robustness, and adaptability to complex scenarios of robot multi-degree-of-freedom motion control.
Need to check novelty before this filing date? Find Prior Art

Description

Technical Field

[0001] This application relates to the field of robotics technology, and more particularly to robot motion control methods and devices, robots, media, and products. Background Technology

[0002] With the development of robotics, artificial intelligence, sensor technology, and motion control technology, robots are increasingly being applied in various scenarios such as industrial manufacturing, warehousing and logistics, commercial services, home services, emergency rescue, medical assistance, and education and scientific research. As the number of degrees of freedom in robots increases, the complexity of robot motion control also increases significantly. Currently, the compliance, robustness, and adaptability to complex scenarios in robot motion control are insufficient, making it difficult to meet the needs of practical applications.

[0003] Based on this, embodiments of this application provide robot motion control methods and devices, robots, media, and products to improve related technologies. Summary of the Invention

[0004] The purpose of this application is to provide robot motion control methods and devices, robots, media, and products to improve the compliance, robustness, and adaptability to complex scenarios of multi-degree-of-freedom motion control of robots.

[0005] In a first aspect, embodiments of this application provide a robot motion control method, comprising: determining trajectory probability distribution information based on robot task target information and robot state information through a generative motion control model; wherein the trajectory probability distribution information is used to characterize the probability distribution of the robot's motion trajectory in the target future time domain under the conditions corresponding to the task target information and the robot state information; determining target motion trajectory information based on the trajectory probability distribution information; and generating motion control data for the robot based on the target motion trajectory information to control the robot's motion.

[0006] Secondly, embodiments of this application provide a robot motion control device, comprising: a model inference module, used to determine trajectory probability distribution information based on robot task target information and robot state information through a generative motion control model; wherein the trajectory probability distribution information is used to characterize the probability distribution of the robot's motion trajectory in the target future time domain under the conditions corresponding to the task target information and the robot state information; a trajectory determination module, used to determine target motion trajectory information based on the trajectory probability distribution information; and a motion control module, used to generate motion control data for the robot based on the target motion trajectory information, so as to control the robot's motion.

[0007] Thirdly, embodiments of this application provide a robot, which includes a control module for performing the steps of any of the methods described above.

[0008] Fourthly, embodiments of this application provide a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above methods.

[0009] Fifthly, embodiments of this application provide a computer program product, the computer program product including a computer program, which, when executed by a processor, implements the steps of any of the above methods.

[0010] This application provides a robot motion control method and device, robot, medium, and product. By determining the trajectory probability distribution information through a generative motion control model, robot motion control can be extended from a single deterministic action output to a probability distribution of the target's future motion trajectory in the time domain. By determining the target motion trajectory information based on the trajectory probability distribution information, a motion trajectory that better matches the task objective and robot state information can be selected from multiple possible motion results represented by the probability distribution. By generating motion control data based on the target motion trajectory information, the generative trajectory results can be transformed into control data that the robot can execute, thereby realizing closed-loop control of robot motion and improving the compliance and adaptability of robot motion control to complex scenarios. Moreover, when the robot's state deviates or is disturbed by external factors, determining the trajectory probability distribution information through the generative motion control model can be understood as the generative motion control model outputting multiple recovery trajectories that both meet the current task objective and have motion feasibility based on the task objective information and the robot's state information after the state deviation. By determining the target motion trajectory information from the trajectory probability distribution information, it can be understood as the robot selecting a more suitable recovery trajectory from it, thereby reducing the risk of motion instability due to single-path recovery failure and improving the robustness of robot motion control. Attached Figure Description

[0011] The embodiments of this application are further described below with reference to the accompanying drawings and specific implementation details.

[0012] Figure 1 This is a flowchart illustrating a robot motion control method provided in an embodiment of this application.

[0013] Figure 2 This is a structural block diagram of a robot motion control device provided in an embodiment of this application.

[0014] Figure 3 This is a structural block diagram of a robot provided in an embodiment of this application.

[0015] Figure 4 This is a structural block diagram of a computer device provided in an embodiment of this application. Detailed Implementation

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the embodiments of this application.

[0017] In the description of the embodiments of this application, it should be understood that the terms "first" and "second" are used for descriptive purposes only and should not be construed as indicating or implying relative importance or implicitly specifying the number of indicated technical features. Therefore, features defined with "first" and "second" may explicitly or implicitly include one or more of the stated features. In the description of the embodiments of this application, "multiple" means two or more, unless otherwise explicitly specified.

[0018] With the development of robotics, artificial intelligence, sensor technology, and motion control technology, robots are increasingly being applied in various scenarios such as industrial manufacturing, warehousing and logistics, commercial services, home services, emergency rescue, medical assistance, and education and scientific research. To enable robots to perform actions such as movement, turning, obstacle avoidance, handling, grasping, posture adjustment, and balance recovery in different task environments, it is necessary to control the robot's motion process. As the number of degrees of freedom of a robot increases, the complexity of robot motion control increases significantly. When robots face scenarios with insufficient training data coverage, external impacts, ground changes, contact anomalies, or sudden changes in the task, they may exhibit problems such as abrupt changes in motion trajectory, untimely posture recovery, uncoordinated local joint movements, or decreased motion stability.

[0019] See Figure 1 , Figure 1 This is a flowchart illustrating a robot motion control method provided in an embodiment of this application.

[0020] To improve the compliance, robustness, and adaptability to complex scenarios of multi-degree-of-freedom motion control of robots, embodiments of this application provide a robot motion control method, including steps S101 to S103.

[0021] Step S101: Based on the robot's task target information and robot state information, determine the trajectory probability distribution information through a generative motion control model. The trajectory probability distribution information characterizes the probability distribution of the robot's motion trajectory in the target future time domain under the conditions corresponding to the task target information and the robot state information.

[0022] Step S102: Determine the target motion trajectory information based on the trajectory probability distribution information.

[0023] Step S103: Generate motion control data for the robot based on the target motion trajectory information to control the robot's movement.

[0024] Robot motion control methods refer to the methods used to drive a robot to complete corresponding movements. The robot can be a humanoid robot, bipedal robot, wheeled robot, multi-legged robot, or a mobile manipulator with multiple degrees of freedom. Robot motion control methods can be used in scenarios such as walking, turning, squatting, standing up, carrying, obstacle avoidance, balance recovery, and posture adjustment, as well as in scenarios involving continuous switching between multiple actions.

[0025] Task objective information refers to information that characterizes the motion goals or action constraints that a robot needs to achieve in the current control task. Task objective information can include target movement speed, target movement direction, target position, target posture, target gait, target foot landing point, target hand position, target torso posture, target action type, target contact state, upper body stability requirements, and load-bearing requirements during transport. For example, task objective information could instruct the robot to walk forward at a speed of 0.5 m / s, or to turn 20 degrees to the left, or to maintain a torso pitch angle of less than 5 degrees during walking.

[0026] Robot state information refers to information used to characterize the robot's motion state, force state, and contact state. Robot state information can include joint positions, joint velocities, joint feedback torques, center of mass position, center of mass velocity, torso base posture, torso angular velocity, foot contact state, foot force state, overall robot velocity, overall robot angular velocity, and external disturbance estimation information. Robot state information can include real-time state information corresponding to the target control cycle, as well as historical state information corresponding to multiple historical control cycles, so that the generative motion control model can determine whether the robot is currently accelerating, decelerating, turning, swinging its legs, landing, or recovering from imbalance based on the motion trend indicated by the robot state information.

[0027] Generative motion control models are machine learning models used to determine the probability distribution of a robot's future motion trajectory given corresponding task target information and robot state information. Unlike control models that directly output a single deterministic action value, generative motion control models can predict the probability distribution of the robot's motion trajectory in the future time domain of the target, thereby providing the robot with one or more physically feasible motion trajectory options.

[0028] Trajectory probability distribution information refers to the probability distribution of a robot's future motion trajectory under the conditions of corresponding task target information and robot state information. Trajectory probability distribution information can be represented by probability distribution parameters, trajectory mean and trajectory variance, multiple sampled trajectories and their probabilities, candidate motion trajectory scores, denoised generation results, flow matching generation results, or latent variables used to describe trajectory probability density.

[0029] The target future time domain refers to a range of predicted future time extending from the start time of the target control cycle. It can cover the target control cycle and one or more subsequent control cycles. The target future time domain can include one or more control cycles, or it can include continuous time intervals. For example, with a control cycle of 20 milliseconds, the target future time domain could be 0.5 seconds in the future, corresponding to 25 future time steps; it could also be 1 second in the future, corresponding to 50 future time steps. The length of the target future time domain can be set according to the robot's motion type, control frequency, computational resources, task difficulty, or stability requirements.

[0030] A motion trajectory can refer to a sequence of motion states of a robot in the future time domain of a target. A motion trajectory can include at least one of the following: joint position trajectory, joint velocity trajectory, joint torque trajectory, torso posture trajectory, foot position trajectory, center of mass motion trajectory, and foot contact state trajectory. By modeling the probability distribution of the motion trajectory, the coupling relationships between different moving parts of the robot's body can be represented.

[0031] Target motion trajectory information can refer to trajectory information determined based on trajectory probability distribution information, used to guide the robot in performing motion. Target motion trajectory information can include the joint reference state, trunk reference state, foot reference state, center of mass reference state, or contact reference state of the target in the future time domain.

[0032] Motion control data refers to data generated based on target motion trajectory information and used to control robot motion. Motion control data can include joint position commands, joint velocity commands, joint torque commands, foot control variables, trunk control variables, center of mass control variables, gait control variables, contact switching control variables, or reference state data from lower-level controllers. For example, motion control data may include the robot's left hip joint target angle, right knee joint target angular velocity, ankle joint target torque, trunk target posture, and foot contact switching time, etc.

[0033] In this embodiment, the robot motion control method models the robot's future trajectory as a probability distribution under the conditions of task target information and robot state information. After receiving the task target information and robot state information, the generative motion control model outputs trajectory probability distribution information based on the joint change law of multiple degrees of freedom of the robot in the future time domain of the target. Subsequently, the robot motion control method determines the target trajectory information based on the trajectory probability distribution information, and then converts the target trajectory information into motion control data that the robot can execute, thereby driving the robot to move according to the motion control data. Since the trajectory probability distribution information can characterize multiple possible motion outcomes of the robot and their probabilistic relationships, when facing disturbances, action switching, or state deviations, the robot can select a trajectory that better suits the current task and current state based on the probability distribution.

[0034] Specifically, when a robot is subjected to external disturbances, changes in ground friction, load variations, abnormal foot contact, or sudden changes in the task objective, the robot's state information can reflect the deviation from its normal state after the disturbance. Generative motion control models can determine the corresponding trajectory probability distribution information based on the task objective information and the robot's state information after the state deviation. The trajectory probability distribution information can characterize multiple feasible recovery trajectories that satisfy the current task objective, such as recovery trajectories that adjust the foot landing point, recovery trajectories that lower the center of mass height, recovery trajectories that compensate for posture through upper limb swing, and recovery trajectories that re-establish the support relationship by changing stride length or stride frequency. Robot motion control methods can determine the target motion trajectory information from multiple recovery trajectories based on the trajectory probability distribution information and generate motion control data to control the robot to recover from the deviated state to a stable motion state.

[0035] This embodiment, by determining the trajectory probability distribution information through a generative motion control model based on task target information and robot state information, expands robot motion control from a single deterministic action output to a probability distribution of the target's future motion trajectory in the time domain. By determining the target motion trajectory information based on the trajectory probability distribution, a motion trajectory that better matches the task target and robot state information can be selected from multiple possible robot motion outcomes represented by the probability distribution. By generating motion control data based on the target motion trajectory information, the generative trajectory results can be transformed into executable control data for the robot, thereby achieving closed-loop control of robot motion and improving the compliance of robot motion control. It enhances the robot's adaptability to complex scenarios. Furthermore, when the robot's state deviates or is subject to external interference, the generative motion control model determines the trajectory probability distribution information. This can be understood as the generative motion control model outputting multiple recovery trajectories that meet the current task objective and have motion feasibility, thereby improving the robot's recovery capability in non-standard states. By determining the target motion trajectory information from the trajectory probability distribution information, it can be understood that the robot selects a more suitable recovery trajectory based on at least one of the following: the degree of deviation from the current state, the contact state, and the stability condition. This reduces the risk of motion instability caused by the failure of a single path recovery and improves the robustness of the robot's motion control.

[0036] In one example, the robot is a humanoid robot with 32 degrees of freedom, a control cycle of 20 milliseconds, and a target future time domain of 0.5 seconds, corresponding to 25 future time steps. Task target information includes: target movement speed of 0.6 m / s, target direction 15 degrees to the left of the robot's current orientation, and torso pitch angle maintained within the range of -5 to 5 degrees. Robot state information includes the positions of 32 joints, the velocities of 32 joints, the feedback torques of 32 joints, the position and velocity of the center of mass, the posture of the torso base, the angular velocity of the torso, and the contact status of the left and right feet. The generative motion control model, based on task target information and robot state information, determines the probability distribution of the position trajectories, velocity trajectories, torque trajectories, center of mass motion trajectory, torso posture trajectory, left foot contact state trajectory, and right foot contact state trajectory for each of the 32 joints within the next 0.5 seconds. The robot motion control method then selects the target motion trajectory from these multiple generated trajectories, considering evaluation criteria such as whether joint angles exceed joint limits, whether joint torques exceed drive limits, whether the center of mass projection meets stability conditions, and whether the left and right foot contact states meet task requirements. Subsequently, the robot motion control method extracts the reference motion state corresponding to the current control cycle from the target motion trajectory information and generates position, velocity, and torque commands for each joint to control the robot to complete a leftward turning and walking motion.

[0037] In some embodiments, determining the trajectory probability distribution information based on the robot's task target information and robot state information using a generative motion control model may include: performing the following processing using the generative motion control model: encoding the task target information to obtain task condition features; encoding the robot state information to obtain state condition features; wherein the state condition features include target state features corresponding to the target control cycle and a sequence of historical state features corresponding to multiple historical control cycles prior to the target control cycle; determining visible motion constraints and target completion motion trajectory components based on the task target information and the robot state information, and encoding the visible motion constraints and target completion motion trajectory components to obtain motion condition features; and determining the trajectory probability distribution information based on the state condition features, the task condition features, and the motion condition features.

[0038] Task condition features refer to the features obtained after encoding task target information. Task requirements such as target movement speed, target movement direction, target posture, and target contact state are converted into task condition features, facilitating processing by generative motion control models. Task condition features are primarily used to characterize the overall motion task requirements of the robot, enabling the generative motion control model to perceive the target motion state the robot needs to achieve when determining trajectory probability distribution information.

[0039] State condition features refer to the features obtained after encoding robot state information. State condition features are used to provide the generative motion control model with the robot's actual physical state and historical motion trends. State condition features include the target state features corresponding to the target control cycle and the sequence of historical state features corresponding to multiple historical control cycles preceding the target control cycle. Through state condition features, the generative motion control model can consider state factors such as whether the robot is currently in a stable supported state, whether it is swinging its legs, whether it is landing, whether its center of mass has shifted, and whether its torso posture has changed during the motion trajectory generation process.

[0040] The target control cycle refers to the control period during which the robot's motion control method needs to generate or update the trajectory probability distribution information. The target control cycle can be determined based on the robot's control frequency; for example, at a control frequency of 50 Hz, one control cycle is 20 milliseconds; at a control frequency of 100 Hz, one control cycle is 10 milliseconds. The robot state information corresponding to the target control cycle can be used to form target state features.

[0041] Target state features refer to feature representations obtained based on robot state information corresponding to the target control cycle. Target state features can include at least one of joint state features, body state features, and contact state features. For example, target state features can characterize the positions, velocities, feedback torques, center of mass position, center of mass velocity, torso base posture, torso angular velocity, left foot contact state, and right foot contact state of multiple joints during the target control cycle. Target state features are used to reflect the robot's instantaneous motion state during the target control cycle.

[0042] A historical state feature sequence can be a sequence of state features corresponding to multiple historical control cycles preceding the target control cycle, arranged in chronological order. A historical state feature sequence can include multiple historical state features, each corresponding to a historical control cycle. This sequence can be used to characterize the robot's motion trend before the target control cycle, such as whether the robot is accelerating, decelerating, lifting its foot, being recently pushed by an external force, whether its center of mass is continuously shifting to one side, and whether its foot contact state is changing.

[0043] Visible motion constraints refer to those explicitly given during the trajectory generation process and required to remain consistent in the generated result. Visible motion constraints can include at least one of the following: target movement speed, target movement direction, target turning angular velocity, target torso posture range, target foot contact state, target hand position, target foot landing point, and target center of mass height range. For example, if the task objective requires the robot to move forward and the robot's state information indicates that the left foot is in a supporting state, both the target forward speed and the left foot support constraint can be considered visible motion constraints.

[0044] Target-completed motion trajectory components refer to the parts of the future motion trajectory that are not fully specified and need to be completed by the generative motion control model, given the correspondence between the task objective information and the robot's state information. Target-completed motion trajectory components can include leg swing trajectories, arm coordinated swing trajectories, waist compensation posture trajectories, knee flexion trajectories, ankle cushioning trajectories, foot transition trajectories, or contact switching trajectories, etc. For example, if the task objective information only specifies the robot's forward velocity and torso posture range, but does not specify the specific trajectories of the hip, knee, ankle, and arm joints in future time steps, these unspecified future motion trajectory parts can be used as target-completed motion trajectory components.

[0045] Motion condition features refer to the conditional features obtained by encoding the visible motion constraints and the target-completed motion trajectory components. Motion condition features are used to characterize which motion requirements need to remain consistent and which parts of the future motion trajectory need to be completed.

[0046] This embodiment encodes task condition features by encoding task target information, converting task requirements such as robot target movement speed, target movement direction, and target posture into condition features that can be processed by the generative motion control model, thereby improving the matching degree between trajectory probability distribution information and task targets. Encoding robot state information to obtain state condition features, including target state features and historical state feature sequences, allows the generative motion control model to simultaneously perceive the robot's immediate state and historical motion trends, thus improving the state adaptability and continuity of future trajectory generation. Determining visible motion constraints and target-completed trajectory components based on task target information and robot state information, and encoding them to obtain motion condition features, distinguishes between explicitly given motion constraints and the future trajectory parts that need to be completed, thereby improving trajectory completion capabilities in scenarios with partial motion constraints. Determining trajectory probability distribution information based on state condition features, task condition features, and motion condition features allows state conditions, task conditions, and motion constraints to be used together for probabilistic modeling of future trajectories, thereby improving the expressive power of trajectory probability distribution information for complex robot motion scenarios. This embodiment addresses the problem that existing robot motion control methods struggle to simultaneously consider task objectives, robot instantaneous state, historical motion trends, and the need for partial motion constraint completion. It adopts a concept that uses task condition features, state condition features, and motion condition features to jointly drive trajectory probability distribution information modeling, thereby improving the task adaptability, state robustness, and motion coordination of the robot's future motion trajectory generation.

[0047] In one example, the robot is a humanoid robot with 32 degrees of freedom, a control frequency of 50 Hz, a control cycle of 20 milliseconds, and a target future time domain of 0.5 seconds, corresponding to 25 future time steps. The task target information includes a target forward velocity of 0.6 m / s, a target turning angular velocity of 0.15 radians / s, and a torso pitch angle range of -5 degrees to 5 degrees. The generative motion control model encodes the task target information to obtain task condition features. The robot state information includes the 32 joint positions, 32 joint velocities, 32 joint feedback torques, center of mass position, center of mass velocity, torso base posture, torso angular velocity, left foot contact state, and right foot contact state corresponding to the target control cycle, as well as the state information corresponding to the 10 historical control cycles preceding the target control cycle. The generative motion control model encodes the robot state information to obtain state condition features, which include the target state features corresponding to the target control cycle and a sequence of historical state features corresponding to the 10 historical control cycles. The generative motion control model, based on task target information and robot state information, defines the target forward velocity (0.6 m / s), target turning angular velocity (0.15 radians / s), torso pitch angle range (-5 degrees to 5 degrees), and left foot support state as visible motion constraints. It defines the leg swing trajectories, arm coordinated swing trajectories, waist compensation posture trajectories, and foot transition trajectories within the next 25 time steps as the target completion motion trajectory components. The visible motion constraints and target completion motion trajectory components are encoded to obtain motion condition features. Based on the state condition features, task condition features, and motion condition features, the generative motion control model determines the trajectory probability distribution information for the robot within the next 0.5 seconds.

[0048] In some embodiments, encoding the visible motion constraints and the target completion motion trajectory components to obtain motion condition features may include: conditionally encoding the visible motion constraints to obtain motion condition encoding information; setting occlusion flags for the target completion motion trajectory components; and determining the motion condition features based on the motion condition encoding information and the occlusion flags.

[0049] Conditional encoding can include numerical normalization, temporal and positional encoding, motion degree-of-freedom encoding, task type encoding, contact state encoding, attitude range encoding, feature mapping, and feature concatenation. Through conditional encoding, different forms of visible motion constraints can be converted into unified or different feature representations, enabling constraints such as target movement speed, target turning angular velocity, target attitude range, and target contact state to participate in subsequent trajectory probability distribution information modeling.

[0050] Motion condition coding information refers to the coded information obtained after conditionally coding visible motion constraints. Motion condition coding information is used to ensure that the generative motion control model maintains consistency with the visible motion constraints when generating future motion trajectories. Motion condition coding information can include target velocity coding information, target direction coding information, target attitude coding information, target contact state coding information, target position coding information, target gait coding information, etc.

[0051] Occlusion markers refer to identifying information used to indicate the components of a target's completed motion trajectory. Occlusion markers enable generative motion control models to distinguish between the motion content information corresponding to visible motion constraints and the future motion trajectory portion that needs to be completed. Occlusion markers can take the form of binary masks, weighted masks, time-step masks, motion degree-of-freedom masks, channel masks, or visibility markers. For example, a value of 1 can represent the motion content information corresponding to visible motion constraints, while a value of 0 can represent the components of the target's completed motion trajectory.

[0052] This embodiment, by conditionally encoding visible motion constraints, can convert different forms of motion constraints, such as target movement speed, target movement direction, target torso posture range, and target foot contact state, into motion conditional encoding information, thereby improving the ability of visible motion constraints to constrain trajectory probability distribution information. By setting occlusion markers for the target completion motion trajectory components, the generative motion control model can identify the incomplete parts of the future motion trajectory, thereby improving the generative motion control model's ability to complete the target completion motion trajectory components. By determining motion condition features based on motion conditional encoding information and occlusion markers, visible motion constraints and occlusion markers can be used as model input conditions, thereby improving the completeness and coordination of the robot's future motion trajectory generation in scenarios with partial motion constraint input. This embodiment addresses the problem that related robot motion control methods struggle to distinguish between constraint preservation and trajectory completion content when only partial motion constraints are given in the task objective. It adopts a concept combining visible motion constraint conditional encoding and target completion motion trajectory component occlusion markers, thereby improving the robot's trajectory completion ability, task adaptability, and motion coordination in scenarios with partial motion constraint input.

[0053] In one example, the robot is a humanoid robot with 32 degrees of freedom, a control cycle of 20 milliseconds, and a target future time domain of 0.5 seconds, corresponding to 25 future time steps. Visible motion constraints include the target forward velocity of 0.6 m / s, the target turning angular velocity of 0.15 radians / s, the torso pitch angle range of -5 degrees to 5 degrees, and the left foot support state. The robot motion control method conditionally encodes the target forward velocity of 0.6 m / s, the target turning angular velocity of 0.15 radians / s, the torso pitch angle range of -5 degrees to 5 degrees, and the left foot support state to obtain motion condition encoding information. The target completion motion trajectory components include the right leg swing trajectory, the coordinated arm swing trajectory, the waist compensation posture trajectory, and the right foot landing transition trajectory within the next 25 time steps. The robot motion control method sets occlusion markers for the above target completion motion trajectory components, for example, marking the channel corresponding to the visible motion constraints as 1 and the channel corresponding to the target completion motion trajectory components as 0, and determines the motion condition features based on the motion condition encoding information and the occlusion markers.

[0054] In some embodiments, determining the trajectory probability distribution information based on the state condition features, the task condition features, and the motion condition features may include: spatially encoding the features corresponding to multiple moving parts in the state condition features based on the topological relationship of the robot's motion nodes to obtain spatial coupling features; temporally encoding the historical state feature sequence based on the control cycle sequence corresponding to the historical state feature sequence to obtain temporal motion features; determining generation condition features based on the spatial coupling features, the temporal motion features, the task condition features, and the motion condition features; and determining the trajectory probability distribution information based on the generation condition features.

[0055] Motion node topology refers to the structural connections or kinematic relationships between various motion nodes of a robot. Motion nodes can include hip, knee, ankle, shoulder, elbow, wrist, waist, trunk, foot, and hand nodes, among others. Motion node topology can be determined based on the robot's mechanical structure, parent-child relationships in kinematic chains, left-right limb symmetry, dynamic coupling, or cooperative motion relationships. For example, hip, knee, and ankle nodes can form a lower limb kinematic chain, and arm swing nodes and leg stepping nodes can form a kinematic relationship for posture balance compensation.

[0056] Spatial coding refers to a processing method that encodes features corresponding to multiple moving parts in state condition features based on the topological relationships of robot motion nodes. Spatial coding can be implemented using methods such as topological adjacency coding, graph structure coding, joint hierarchy coding, limb grouping coding, spatial position coding, and spatial attention coding. Through spatial coding, generative motion control models can perceive the connection relationships and motion coupling relationships between different motion nodes, thereby extracting cooperative motion information of different parts of the robot.

[0057] Spatial coupling features refer to features obtained by spatially encoding the features corresponding to multiple moving parts in the state condition features. Spatial coupling features are used to characterize the spatial relationships between different motion nodes, different limbs, torso, feet, and center of mass of a robot. Spatial coupling features can reflect the linkage relationship between leg joints, the switching relationship between left and right foot support, the compensation relationship between arm swing and torso balance, and the relationship between foot contact and center of mass stability, etc.

[0058] The control cycle sequence refers to the temporal order between control cycles corresponding to multiple historical state features in a historical state feature sequence. This sequence can be represented as the first historical control cycle preceding the second, the second preceding the third, and so on, continuing until the last historical control cycle before the target control cycle. By understanding the control cycle sequence, generative motion control models can interpret the robot's historical state changes in chronological order, rather than treating the historical state feature sequence as a disordered set of features.

[0059] Temporal coding refers to a processing method that encodes historical state feature sequences in a time dimension based on the sequential relationship between control cycles corresponding to those sequences. Temporal coding can be implemented using methods such as temporal position coding, temporal attention calculation, time window feature fusion, and cyclic feature extraction. Through temporal coding, generative motion control models can extract the robot's motion trends over multiple past control cycles, enabling the generation of future motion trajectories to consider the robot's motion inertia and state changes.

[0060] Temporal motion features refer to the features obtained by temporally encoding a sequence of historical state features. These features characterize the dynamic changes of a robot over a period of time prior to the target control cycle. For example, temporal motion features can characterize motion trends such as the robot's center of mass continuously shifting to the left, the torso angular velocity gradually increasing, the right foot about to land, the left knee joint gradually extending, and the robot recovering from external disturbances. Temporal motion features can provide historical motion trends for modeling trajectory probability distribution information.

[0061] Generative conditional features can refer to comprehensive conditional features determined based on spatial coupling features, temporal motion features, task condition features, and motion condition features. Generative conditional features simultaneously include the spatial coupling relationships between multiple moving parts of the robot, the temporal motion trends corresponding to the robot's historical states, the overall task objective of the robot, the relationships between visible motion constraints, and the components of the target-completed motion trajectory. Generative conditional features serve as the foundation for determining the trajectory probability distribution information in generative motion control models.

[0062] This embodiment spatially encodes features corresponding to multiple moving parts in the state condition features based on the topological relationships of robot motion nodes. This enables the generative motion control model to perceive the spatial coupling relationships between different moving parts of the robot, thereby improving the rationality of the coordinated changes of joints, torso, feet, and center of mass in future motion trajectories. By temporally encoding the historical state feature sequence based on the control cycle sequence corresponding to the historical state feature sequence, the generative motion control model can extract the robot's historical motion trends, thereby improving the continuity between the robot's dynamic changes. By determining the generation condition features based on spatial coupling features, temporal motion features, task condition features, and motion condition features, spatial relationships, temporal relationships, task objectives, visible motion constraints, and target-completed motion trajectory components can be used as modeling conditions for trajectory probability distribution information, thereby improving the ability of trajectory probability distribution information to express complex robot motion scenarios. This embodiment addresses the problem that related robot motion control methods struggle to simultaneously consider spatial coupling, historical motion trends, task objectives, and only provide partial motion constraints. It adopts the concept of spatial encoding, temporal encoding, and the fusion of multiple types of condition features to form generation condition features, thereby improving the coordination, continuity, and adaptability to complex tasks in the generation of future robot motion trajectories.

[0063] In one example, the robot is a humanoid robot with 32 degrees of freedom, a control cycle of 20 milliseconds, and a target future time domain of 0.5 seconds, corresponding to 25 future time steps. State condition features include the target state features corresponding to the target control cycle and the sequence of historical state features corresponding to the 10 previous historical control cycles. The robot's motion node topology includes the connections between the torso node, left hip joint node, left knee joint node, left ankle joint node, right hip joint node, right knee joint node, right ankle joint node, left shoulder joint node, left elbow joint node, right shoulder joint node, and right elbow joint node. The generative motion control model, based on the robot's motion node topology, spatially encodes multiple motion parts corresponding to the target state features and historical state feature sequences to obtain spatial coupling features; and based on the sequential relationship of the control cycles corresponding to the 10 historical state features, temporally encodes the historical state feature sequences to obtain temporal motion features. The task condition features characterize the target's forward velocity (0.6 m / s) and turning angular velocity (0.15 radians / s). The motion condition features characterize the torso pitch angle range of -5 to 5 degrees, the left foot support state, and the right leg swing trajectory and the coordinated arm swing trajectory that need to be completed within the next 25 time steps. The generative motion control model determines the generative condition features based on spatial coupling features, temporal motion features, task condition features, and motion condition features, and then determines the trajectory probability distribution information of the robot within the next 0.5 seconds based on the generative condition features.

[0064] In some embodiments, the step of temporally encoding the historical state feature sequence based on the control cycle sequence corresponding to the historical state feature sequence to obtain temporal motion features may include: generating first temporal causal constraint information based on the control cycle sequence corresponding to multiple historical state features in the historical state feature sequence; and performing attention calculation on multiple historical state features in the historical state feature sequence based on the first temporal causal constraint information to obtain the temporal motion features.

[0065] In some embodiments, determining the generation condition feature based on the spatial coupling feature, the temporal motion feature, the task condition feature, and the motion condition feature may include: determining the future time domain query feature corresponding to each time step in the target future time domain based on the spatial coupling feature, the temporal motion feature, the task condition feature, and the motion condition feature; generating second temporal causal constraint information based on the sequential relationship between each time step in the target future time domain; performing attention update on the future time domain query feature based on the second temporal causal constraint information to obtain updated query feature; and determining the generation condition feature based on the updated query feature, the spatial coupling feature, the temporal motion feature, the task condition feature, and the motion condition feature.

[0066] The first temporal causal constraint information is used to constrain the interaction direction of attention calculation among multiple historical state features in the historical state feature sequence. This first temporal causal constraint information allows any historical state feature corresponding to a historical control period to reference that historical control period and historical state features preceding that period during attention calculation, while restricting historical state features following that period from participating in the attention calculation of the historical state feature corresponding to that period. The first temporal causal constraint information can take at least one form from causal masking matrix, temporal step visibility matrix, attention masking matrix, and temporal order constraint identifier.

[0067] Attention computation refers to a process that establishes association weights among multiple historical state features and then fuses these features. Attention computation can determine association weights based on the similarity, temporal distance, state change relationships, contact state change relationships, or motion trend relationships between historical state features, and then fuse multiple historical state features based on these association weights. Attention computation based on first-order causal constraint information allows generative motion control models to extract the robot's historical motion trends while maintaining temporal order.

[0068] Future time-domain query features are used to support trajectory generation requirements for corresponding future time steps. Future time-domain query features can be determined based on spatial coupling features, temporal motion features, task condition features, and motion condition features, and correspond to the first future time step, the second future time step, and so on up to the Nth future time step, respectively.

[0069] The second temporal causal constraint information is used to constrain the attention update direction among the future temporal query features corresponding to each time step within the target's future temporal domain. This second temporal causal constraint information allows any future temporal query feature corresponding to any time step within the target's future temporal domain to reference both that time step and future temporal query features preceding that time step during attention updates, while restricting future temporal query features following that time step from participating in the attention update of the corresponding future temporal query feature at that time step. The specific forms of the first and second temporal causal constraint information can be the same or different; both can adopt at least one of the following forms: causal mask matrix, time step visibility matrix, attention masking matrix, and temporal order constraint identifier. The difference lies in that the first temporal causal constraint information applies to the historical state feature sequence, while the second temporal causal constraint information applies to the future temporal query features.

[0070] Attention update refers to a processing method that updates future time-domain query features based on the relationships between them. Attention update can be performed under the constraint of second-order causal information, enabling the future time-domain query features corresponding to each time step within the target future time domain to gradually integrate information from previous time steps in chronological order.

[0071] Updating query features can refer to features obtained by updating future time-domain query features with attention based on second-series causal constraint information. Updating query features can reflect the temporal dependencies and trajectory evolution relationships between different time steps in the future time domain of the target. Updating query features can provide a continuous generation basis in the future time dimension for generating conditional features.

[0072] This embodiment generates first temporal causal constraint information based on the sequential relationship of control cycles corresponding to multiple historical state features in a historical state feature sequence. This ensures that the attention calculation of multiple historical state features conforms to the real time sequence, thereby improving the ability of temporal motion features to express the robot's historical motion trends. Specifically, by performing attention calculation on multiple historical state features in the historical state feature sequence based on the first temporal causal constraint information, motion inertia correlations can be established between different historical control cycles, thereby improving the ability of temporal motion features to represent processes such as acceleration, deceleration, landing, support switching, and disturbance recovery. By determining future temporal query features based on spatial coupling features, temporal motion features, task condition features, and motion condition features, the future temporal query features corresponding to each time step within the target future temporal domain can obtain initial generation conditions related to spatial structure, historical trends, task objectives, and motion conditions. By updating the attention of future temporal query features based on second temporal causal constraint information, the reverse influence of future temporal query features corresponding to subsequent time steps on future temporal query features corresponding to preceding time steps can be limited, thereby improving the temporal consistency and causal rationality of the robot's future motion trajectory generation process. By determining the generation condition features based on update query features, spatial coupling features, temporal motion features, task condition features, and motion condition features, the generation condition features can be determined. Specifically, historical time causal relationships, future time causal relationships, spatial coupling relationships, temporal motion trends corresponding to the robot's historical states, task conditions, and motion conditions can be used to model trajectory probability distribution information, thereby improving the continuity and stability of the robot's future motion trajectory generation.

[0073] In one example, the robot is a humanoid robot with 32 degrees of freedom, a control cycle of 20 milliseconds, and a historical state feature sequence including the historical state features corresponding to the 10 historical control cycles preceding the target control cycle. The target future time domain is 0.5 seconds into the future, corresponding to 25 future time steps. The generative motion control model generates first temporal causal constraint information based on the sequential relationship of the control cycles corresponding to the 10 historical state features. This first temporal causal constraint information can be a 10×10 causal mask matrix, allowing the 6th historical state feature to reference the 1st to 6th historical state features during attention calculation, without referencing the 7th to 10th. The generative motion control model performs attention calculation on the 10 historical state features based on the first temporal causal constraint information to obtain temporal motion features. Subsequently, the generative motion control model determines the future time domain query features corresponding to each of the 25 future time steps based on spatial coupling features, temporal motion features, task condition features, and motion condition features. The generative motion control model generates a second temporal causal constraint information based on the sequential relationships between 25 future time steps. This second temporal causal constraint information can be a 25×25 causal mask matrix, allowing the future temporal query feature corresponding to the 8th future time step to refer to the future temporal query features corresponding to the 1st to 8th future time steps during attention updates, instead of referring to the future temporal query features corresponding to the 9th to 25th future time steps. The generative motion control model updates the attention of the 25 future temporal query features based on the second temporal causal constraint information, obtaining updated query features. Based on the updated query features, spatial coupling features, temporal motion features, task condition features, and motion condition features, it determines the generation condition features.

[0074] In some embodiments, determining the trajectory probability distribution information based on the generation condition features may include: constructing a first trajectory variable corresponding to the target in the future time domain; performing conditional generation iteration on the first trajectory variable based on the generation condition features to obtain a trajectory generation result; and determining the trajectory probability distribution information based on the trajectory generation result.

[0075] The first trajectory variable can refer to the trajectory variable that corresponds to the target's future time domain and participates in the condition generation iteration. The first trajectory variable can include trajectory representations corresponding to multiple future time steps and multiple moving parts of the robot within the target's future time domain. The first trajectory variable can include one or more of the following: joint position variables, joint velocity variables, joint torque variables, trunk posture variables, center of mass motion variables, foot motion variables, and contact state variables. The first trajectory variable can be a trajectory variable obtained through random initialization, or a trajectory variable obtained based on the trajectory result of the previous control cycle, a coarse reference trajectory, or an initial planned trajectory.

[0076] Conditional generation iteration refers to the process of updating the first trajectory variable one or more times under the constraints of generation condition features to obtain the trajectory generation result. Conditional generation iteration can employ conditional denoising or flow field update processing. The number of conditional generation iterations can be set according to the trajectory generation accuracy, control real-time performance, and the future time domain length of the target, for example, it can be set to 4, 6, 8, or 12 times. Through conditional generation iteration, the first trajectory variable can gradually evolve into a trajectory generation result that matches the robot's state and task objective.

[0077] The trajectory generation results can include one or more of the following: joint position sequence, joint velocity sequence, joint torque sequence, trunk posture sequence, center of mass motion sequence, foot position sequence, and contact state sequence, corresponding to multiple moving parts of the target in the future time domain.

[0078] Determining trajectory probability distribution information based on trajectory generation results can refer to the probabilistic expression of future motion trajectories formed from the trajectory generation results. Specifically, the trajectory generation result can be used as the central trajectory of the trajectory probability distribution, or multiple trajectory generation results obtained from multiple samplings can be combined to determine the trajectory mean, trajectory variance, candidate trajectory probability, or trajectory score.

[0079] This embodiment constructs a first trajectory variable corresponding to the future time domain of the target, providing a unified variable carrier for the generation of trajectories for multiple time steps and multiple moving parts, thereby improving the robot's ability to express the future time domain of the target in the trajectory generation process. By performing conditional generation iteration on the first trajectory variable based on generation condition features, the first trajectory variable can be gradually updated under the combined effects of the spatial coupling relationship of multiple moving parts of the robot, the temporal motion trend corresponding to the robot's historical state, the robot's overall task objective, visible motion constraints, and the target-completed trajectory components, thereby improving the matching degree between the trajectory generation result and the current motion scene. By determining the trajectory probability distribution information based on the trajectory generation result, the trajectory result obtained from the conditional generation iteration can be converted into a probabilistic expression of the future motion trajectory, thus providing a basis for subsequent target motion trajectory selection.

[0080] In some embodiments, the step of performing conditional generation iteration on the first trajectory variable based on the generation condition features to obtain the trajectory generation result may include: performing at least one conditional denoising process on the first trajectory variable based on the generation condition features to obtain the trajectory generation result.

[0081] In some embodiments, the first trajectory variable may include at least one of a noisy initial trajectory variable, a coarse trajectory variable, and a randomly initialized trajectory variable.

[0082] Conditional denoising refers to the process of updating the first trajectory variable by denoising based on the generated conditional features. In each iteration, conditional denoising predicts the noise components, trajectory correction amounts, or denoised trajectory representations in the first trajectory variable and updates it accordingly. Through at least one conditional denoising process, the first trajectory variable can be gradually transformed into a trajectory generation result that satisfies the spatial coupling relationships of multiple moving parts of the robot, the temporal motion trends corresponding to the robot's historical states, the robot's overall task objective, visible motion constraints, and target completion trajectory components.

[0083] This embodiment uses conditional denoising on the first trajectory variable to generate a motion trajectory that meets the generation condition features by using a stepwise denoising method, thereby improving the continuity and smoothness of the robot's future motion trajectory.

[0084] In some embodiments, the step of performing conditional generation iteration on the first trajectory variable based on the generation condition features to obtain the trajectory generation result may include: determining a conditional flow field based on the generation condition features, and performing at least one flow field update process on the first trajectory variable based on the conditional flow field to obtain the trajectory generation result.

[0085] A conditional flow field can be defined as a trajectory generation field that, under the constraints of generative conditional features, indicates the update direction and magnitude of the first trajectory variable. It can be understood as a continuous update rule defined in the trajectory variable space. The conditional flow field can be determined by a generative motion control model based on generative conditional features and used to indicate how the first trajectory variable gradually evolves into a trajectory generation result that conforms to the task objective and the robot's state. For example, when the generative conditional features indicate that the robot needs to turn left and its right foot is about to land, the conditional flow field can instruct the trajectory variables corresponding to the hip, knee, ankle, trunk posture, and foot position to update in directions matching the left turn and foot landing.

[0086] Flow field update processing refers to the process of updating the first trajectory variable based on the conditional flow field. The flow field update processing can determine the update amount of the first trajectory variable at each time step and for each moving part within the target future time domain based on the conditional flow field, and adjust the first trajectory variable based on the update amount. Flow field update processing can be performed once or multiple times. Through at least one flow field update process, the first trajectory variable can gradually change along the conditional flow field, resulting in a trajectory generation result that matches the characteristics of the generation conditions.

[0087] This embodiment determines the conditional flow field based on the characteristics of the generation conditions and performs flow field update processing. It can guide the first trajectory variable to evolve step by step by continuously updating the direction, thereby improving the stability and controllability of the robot motion trajectory generation process.

[0088] The above embodiments address the problem that related robot motion control methods have difficulty in continuously generating future trajectories that conform to the task objectives and robot states. They adopt the concept of constructing a first trajectory variable and performing conditional denoising or flow field update processing based on the generation condition features, thereby improving the continuity, physical feasibility, and dynamic adaptability of robot future motion trajectory generation.

[0089] In one example, the robot is a humanoid robot with 32 degrees of freedom, a control cycle of 20 milliseconds, and a target future time domain of 0.5 seconds, corresponding to 25 future time steps. The robot motion control method constructs a first trajectory variable, which includes joint position and velocity variables corresponding to the 32 joints within the 25 future time steps. The generated conditional features characterize the robot's target forward speed as 0.6 m / s, target turning angular velocity as 0.15 radians / s, left foot in a supporting position, and torso pitch angle maintained within the range of -5 to 5 degrees. If conditional denoising is used, the robot motion control method performs 8 conditional denoising processes on the first trajectory variable based on the generated conditional features to obtain the trajectory generation result. If flow field update is used, the robot motion control method determines the conditional flow field based on the generated conditional features and performs 6 flow field update processes on the first trajectory variable. The joint position update amount corresponding to each flow field update process can be limited to within 0.03 radians to obtain the trajectory generation result. The robot motion control method determines the trajectory probability distribution information of 32 joints within the next 0.5 seconds based on the trajectory generation results.

[0090] In some embodiments, determining the target motion trajectory information based on the trajectory probability distribution information may include: generating multiple candidate motion trajectory information based on the trajectory probability distribution information; evaluating the motion feasibility of each of the multiple candidate motion trajectory information to obtain a trajectory evaluation result corresponding to each candidate motion trajectory information; and determining the target motion trajectory information from the multiple candidate motion trajectory information based on the trajectory evaluation result; wherein the motion feasibility evaluation includes at least one constraint evaluation based on joint motion restrictions, drive restrictions, contact state restrictions, stability conditions, and task matching conditions.

[0091] Candidate motion trajectory information refers to motion trajectory information generated based on trajectory probability distribution information, which can be further evaluated and selected. Candidate motion trajectory information can include at least one of the following: joint position trajectory, joint velocity trajectory, joint torque trajectory, trunk posture trajectory, center of mass motion trajectory, foot motion trajectory, and contact state trajectory, corresponding to multiple moving parts within the target's future time domain. Multiple candidate motion trajectory information can correspond to different movement strategies, such as adjusting the foot landing point, changing the upper body swing amplitude, reducing movement speed, increasing knee flexion, and adjusting the foot contact time.

[0092] Motion feasibility evaluation refers to the process of assessing whether multiple candidate motion trajectory information is suitable for actual robot execution. Motion feasibility evaluation does not simply select based on the probability of candidate motion trajectory information, but further determines whether the candidate motion trajectory information meets the robot's actual motion conditions, thereby reducing the possibility of high-probability but unsuitable candidate motion trajectory information being selected as the target motion trajectory information.

[0093] Trajectory evaluation results refer to the evaluation results obtained after evaluating the motion feasibility of candidate motion trajectory information. Trajectory evaluation results can include evaluation scores, ranking results, feasibility levels, pass / fail indicators, constraint violation degrees, and comprehensive cost values. For example, trajectory evaluation results can include joint restriction scores, drive restriction scores, contact state restriction scores, stability scores, and task matching scores; or multiple scores can be weighted to obtain a comprehensive evaluation score.

[0094] Target motion trajectory information refers to the motion trajectory information determined from multiple candidate motion trajectory information based on trajectory evaluation results, used to generate robot motion control data. Target motion trajectory information can be the candidate motion trajectory information with the highest comprehensive evaluation score, or it can be the candidate motion trajectory information that meets the constraints and has a high probability, or it can be the candidate motion trajectory information that achieves a good balance between task matching degree and motion feasibility. Target motion trajectory information can be used to guide the robot to perform corresponding movements in the target's future time domain.

[0095] Joint motion constraints refer to the limitations on the range of motion and changes in motion that each joint of a robot must meet during operation. These constraints may include upper and lower limits for joint angles, upper limits for joint velocity and acceleration, and limits on the amplitude of joint changes between adjacent control cycles. By evaluating constraints based on joint motion constraints, the risk of candidate motion trajectory information causing robot joints to approach or exceed their mechanical range of motion can be reduced.

[0096] Drive constraints refer to the limitations that a robot actuator must meet in terms of output capabilities. Drive constraints may include upper limits on joint torque, joint power, drive current, drive temperature, and torque change rate. By evaluating constraints based on drive constraints, the number of candidate motion trajectories that exceed the robot's drive capabilities can be reduced.

[0097] Contact state constraints refer to the limiting conditions that the robot's contact relationship with the external environment must meet. These constraints can include whether the foot is in contact with the ground, the timing of foot contact, whether the foot slips, whether the supporting foot is stable, whether the swinging foot is off the ground, and whether the hand maintains contact. By evaluating constraints based on contact state constraints, the matching degree between candidate motion trajectory information and the robot's support state can be improved.

[0098] Stability conditions refer to the conditions used to evaluate whether a robot can maintain balance or posture stability when executing candidate motion trajectory information. Stability conditions may include the projected position of the center of mass, the relationship of the supporting polygons, the range of torso tilt angles, the range of torso angular velocities, the plantar pressure distribution, the range of zero-moment points, and the ability to recover from sway. By evaluating constraints based on stability conditions, the balance reliability of the robot when executing target motion trajectory information can be improved.

[0099] Task matching conditions refer to the conditions used to evaluate the degree of matching between candidate motion trajectory information and the task target. Task matching conditions may include target movement speed error, target movement direction error, target position error, target posture error, target gait matching degree, target contact state matching degree, target hand position error, etc. By performing constraint evaluation based on task matching conditions, the degree to which the target motion trajectory information meets the task target can be improved.

[0100] This embodiment generates multiple candidate motion trajectory information based on trajectory probability distribution information, thereby improving the robot's motion selection space in scenarios involving disturbances, task switching, and multi-degree-of-freedom coupling. By evaluating the motion feasibility of each candidate motion trajectory, the candidate trajectory information can be filtered based on the robot's mechanical constraints, driving capabilities, contact states, stability requirements, and task objectives, thus improving the executability of the target motion trajectory information. By determining the target motion trajectory information from multiple candidate motion trajectory information based on the trajectory evaluation results, a comprehensive selection can be made between trajectory probability and motion constraints, thereby improving the stability of robot motion control and task completion performance. This embodiment addresses the problem that some robot motion control methods directly execute trajectories based solely on model output, which may lead to mismatches between the generated trajectory and joint constraints, driving capabilities, contact states, or stability requirements. It adopts a concept that combines candidate motion trajectory information generation and motion feasibility evaluation, thereby improving the physical feasibility, task adaptability, and dynamic robustness of the target motion trajectory information.

[0101] In one example, the robot is a humanoid robot with 32 degrees of freedom, and the target future time domain is 0.5 seconds into the future, corresponding to 25 future time steps. Trajectory probability distribution information is used to characterize the probability distribution of the motion trajectories corresponding to the 32 joints within the next 0.5 seconds. The robot motion control method generates 64 candidate motion trajectory information based on the trajectory probability distribution information. For each candidate motion trajectory information, the robot motion control method performs a motion feasibility evaluation: the joint angles need to be within the limits of each joint angle, the joint torque needs to be less than the maximum torque of the corresponding actuator (120 Nm), the foot slippage during the left foot support phase needs to be less than 2 cm, the center of mass projection needs to be within the support area, and the error between the average forward speed and the target speed of 0.6 m / s needs to be less than 0.08 m / s. The robot motion control method calculates the trajectory evaluation results for each of the 64 candidate motion trajectory information, and selects the candidate motion trajectory information with the highest task matching evaluation result from the candidate motion trajectory information that meets the joint motion constraints, actuator constraints, contact state constraints, and stability conditions as the target motion trajectory information.

[0102] In some embodiments, generating motion control data for the robot based on the target motion trajectory information to control the robot's motion may include: determining a reference motion state corresponding to a target control cycle from the target motion trajectory information; generating the motion control data based on the reference motion state and the robot's feedback motion state in the target control cycle; receiving updated state information of the robot after controlling the robot to execute the motion control data; and re-determining the updated trajectory probability distribution information using the generative motion control model based on the task target information and the updated state information to perform motion control on the robot in the next control cycle.

[0103] Reference motion state refers to the reference state extracted from the target motion trajectory information and corresponding to the target control cycle. The reference motion state can include joint target position, joint target velocity, joint target torque, torso target posture, torso target angular velocity, center of mass target position, center of mass target velocity, foot target position, and foot sole target contact state, etc., corresponding to the target control cycle. The reference motion state is used as a reference for generating robot motion control data.

[0104] Feedback motion state refers to the robot's actual motion state corresponding to the target control cycle. It can be obtained through robot sensors, joint actuator feedback data, inertial measurement data, foot contact data, or state estimation algorithms. Feedback motion state can include actual joint positions, actual joint velocities, actual joint torques, actual trunk posture, actual trunk angular velocity, actual center of mass position, actual center of mass velocity, actual foot position, and actual foot contact state. Feedback motion state reflects the robot's true execution state and is compared with a reference motion state.

[0105] Updating state information refers to the robot's state information that is re-obtained after the robot executes motion control data. This updated state information is used as input to redetermine the trajectory probability distribution information in the next control cycle.

[0106] This embodiment determines the reference motion state corresponding to the target control cycle from the target motion trajectory information, converting the robot's future time-domain motion trajectory information into an executable reference state for the current control cycle, thereby improving the correspondence between the target motion trajectory information and the actions executed in the actual control cycle. By generating motion control data based on the reference motion state and the robot's feedback motion state in the target control cycle, the control output can be corrected according to the robot's actual execution state, thereby improving the robot's ability to track the target motion trajectory information. By receiving the robot's updated state information after controlling the robot to execute motion control data, the robot's actual motion state after execution can be obtained, thus providing the latest state basis for the generation of the motion trajectory in the next control cycle. By redetermining the updated trajectory probability distribution information based on the task target information and the updated state information, the generative motion control model can continuously update the trajectory probability distribution information according to changes in the robot's state, thereby improving the continuity and dynamic adaptability of the robot's motion control process. This embodiment addresses the problem that related robot motion control methods are prone to mismatch between subsequent trajectories and actual states due to state deviations, external disturbances, or contact changes during execution. It adopts the concept of generating motion control data based on feedback motion states and continuously updating trajectory probability distribution information based on updated state information, thereby improving the closed-loop performance, stability, and state adaptability of robot motion control.

[0107] In some embodiments, the generative motion control model can be a model trained in the following manner: acquiring heterogeneous motion data, the heterogeneous motion data including at least two of human motion data, robot simulation motion data, and robot actual motion data; converting the heterogeneous motion data into training trajectory data matching the moving parts of the robot; constructing training samples based on the training trajectory data; wherein the training samples include sample task target information, sample robot state information, and sample future motion trajectory; determining the sample predicted trajectory probability distribution information through an initial generative motion control model based on the sample task target information and the sample robot state information, and determining the trajectory generation loss based on the sample predicted trajectory probability distribution information and the sample future motion trajectory; adjusting the model parameters of the initial generative motion control model based on the trajectory generation loss to obtain the generative motion control model.

[0108] An initial generative motion control model can refer to a basic model that has been pre-built before the task is carried out, but has not yet been trained or updated in terms of parameters.

[0109] Human motion data can be obtained through methods such as human motion capture, video pose estimation, inertial sensor acquisition, or manual action recording.

[0110] Robot simulation motion data refers to robot motion data generated in a physical simulation environment. It can include data obtained by the robot performing motion tasks under different mission objectives, terrain conditions, external force disturbances, friction conditions, and contact states.

[0111] Robot motion data refers to motion data collected during the actual operation of the robot's hardware. This data reflects the driving characteristics, sensing errors, contact changes, and execution deviations of the actual robot hardware.

[0112] Training trajectory data can include data such as joint position, joint velocity, joint torque, trunk posture, center of mass movement, foot position or contact state corresponding to different moving parts at multiple time steps.

[0113] Sample task target information can refer to the information in the training samples used to characterize the requirements of the robot's motion task.

[0114] Sample robot state information can refer to the information in the training samples used to characterize the robot's state.

[0115] The future trajectory of a sample can refer to the future trajectory data of the training samples used as the learning target.

[0116] Trajectory generation loss measures the degree of matching between the predicted trajectory probability distribution information and the future trajectory of a sample. The trajectory generation loss can be determined based on the probability of generating the future trajectory of the sample under the predicted trajectory probability distribution information. For example, when the probability of generating the future trajectory of the sample under the predicted trajectory probability distribution information is high, the trajectory generation loss can be small; when the probability of generating the future trajectory of the sample under the predicted trajectory probability distribution information is low, the trajectory generation loss can be large.

[0117] This embodiment expands the sources of training data by acquiring at least two types of heterogeneous motion data from human motion data, robot simulation motion data, and actual robot motion data, thereby improving the generative motion control model's learning ability for different action types, motion states, and scene conditions. By converting heterogeneous motion data into training trajectory data that matches the robot's moving parts, motion data from different sources can be unified into the robot's executable motion expression space, thus improving the matching degree between training data and robot motion control tasks. By constructing training samples based on training trajectory data, including sample task target information, sample robot state information, and sample future motion trajectories, the initial generative motion control model can learn the probability distribution of future motion trajectories under task and state conditions, thereby improving the relevance of model training. By determining the trajectory generation loss based on the sample predicted trajectory probability distribution information and the sample future motion trajectory, and adjusting the model parameters of the initial generative motion control model based on the trajectory generation loss, the generation probability of the sample future motion trajectory under the sample predicted trajectory probability distribution information can be improved. This embodiment addresses the problems of single-source training data for relevant robot motion control models, mismatch between training data and robot moving parts, and difficulty in learning the probability distribution of future motion trajectories. It adopts the concept of unified conversion of heterogeneous motion data, construction of conditional training samples, and training based on trajectory generation loss, thereby improving the generalization ability, motion feasibility modeling ability, and adaptability to complex scenarios of generative motion control models.

[0118] See Figure 2 , Figure 2 This is a structural block diagram of a robot motion control device provided in an embodiment of this application.

[0119] This application also provides a robot motion control device, including a model reasoning module, a trajectory determination module, and a motion control module.

[0120] The model inference module is used to determine the trajectory probability distribution information based on the robot's task target information and robot state information through a generative motion control model; wherein, the trajectory probability distribution information is used to characterize the probability distribution of the robot's motion trajectory in the target future time domain under the conditions corresponding to the task target information and the robot state information.

[0121] The trajectory determination module is used to determine the target motion trajectory information based on the trajectory probability distribution information.

[0122] The motion control module is used to generate motion control data for the robot based on the target motion trajectory information, so as to control the robot's movement.

[0123] See Figure 3 , Figure 3 This is a structural block diagram of a robot provided in an embodiment of this application.

[0124] This application also provides a robot, which includes a control module for performing the steps of any of the above methods.

[0125] This application also provides a computer-readable storage medium storing a computer program that, when executed by a processor, implements the steps of any of the above methods.

[0126] This application also provides a computer program product, which includes a computer program that, when executed by a processor, implements the steps of any of the above methods.

[0127] The computer program product may be a portable compact disc read-only memory (CD-ROM) and include program code, and may run on a terminal device, such as a personal computer. However, the computer program product of the embodiments of this application is not limited thereto, and the computer program product may be any combination of one or more computer-readable media.

[0128] See Figure 4 , Figure 4 This is a structural block diagram of a computer device provided in an embodiment of this application.

[0129] This application also provides a computer device, which includes a memory and a processor. The memory stores a computer program, and the processor executes the computer program to implement the steps of any of the above methods.

[0130] The embodiments of this application do not limit the computer device, which may be, for example, a local computer device, a cloud computer device, a distributed computer device, etc.

[0131] The computer device may include: a memory 110, a processor 120, and a communication interface 130. The memory 110, the processor 120, and the communication interface 130 are connected through internal connection paths.

[0132] The memory 110 is used to store computer programs, which in some implementations may include code for implementing the methods of the embodiments of this application.

[0133] The processor 120 executes the computer program stored in the memory 110 to control the communication interface 130 to receive input data and information, and output operation results and other data. In some implementations, when the solutions of the embodiments of this application are implemented by software or firmware, the computer program used to implement the solutions of the embodiments of this application can be stored in the processor 120 and executed by the processor 120.

[0134] The memory 110 may be volatile memory or non-volatile memory, or may include both volatile and non-volatile memory. The non-volatile memory may be read-only memory (ROM), programmable read-only memory (PROM), erasable programmable read-only memory (EPROM), electrically erasable programmable read-only memory (EEPROM), or flash memory. The volatile memory may be random access memory (RAM). It should be noted that the memory 110 described herein is intended to include, but is not limited to, any memory of these and other suitable types. As an example, the memory 110 includes random access memory (RAM), cache memory, and read-only memory (ROM). The memory 110 stores a computer program that can be executed by processor 120, causing processor 120 to implement the steps of any of the methods described above.

[0135] The processor 120 can be a central processing unit (CPU), or it can be other general-purpose processors, digital signal processors (DSPs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), or other programmable logic devices, discrete gate or transistor logic devices, discrete hardware components, etc. The general-purpose processor can be a microprocessor, or the processor 120 can be any conventional processor.

[0136] In implementation, each step of the above method can be completed by the integrated logic circuitry of the hardware in the processor 120 or by instructions in software form. The method disclosed in the embodiments of this application can be directly implemented by the hardware processor, or by a combination of hardware and software modules in the processor 120. The software modules can be located in mature storage media in the art, such as random access memory, flash memory, read-only memory, programmable read-only memory, electrically erasable programmable memory, or registers. This storage medium is located in the memory 110, and the processor 120 reads the information in the memory 110 and, in conjunction with its hardware, completes the steps of the above method. To avoid repetition, detailed descriptions are not provided here.

[0137] In some implementations, in addition to the hardware units described above, computer devices may also include software modules, such as operating systems, basic input / output systems (BIOS), and application software.

[0138] An operating system is used to manage the hardware and / or software resources of a computer device; it is the kernel and foundation of the computer. The operating system handles fundamental tasks such as managing and configuring memory, determining the priority of system resource allocation, controlling input and output devices, operating the network, and managing the file system. To facilitate user operation, most operating systems provide a user interface for interaction with the system.

[0139] The BIOS is used to perform hardware initialization during the power-on boot phase and to provide runtime services for the operating system and applications. In some implementations, the BIOS can also monitor and display processor temperature and execute temperature protection strategies.

[0140] Application software, also known as an application program, is software written for a specific user application purpose and is one of the main categories of computer software. For example, application software can be a program used to achieve purposes such as power control and temperature management.

[0141] It is understood that the specific examples in this application are only intended to help those skilled in the art better understand the implementation of this application, and are not intended to limit the scope of protection of this application.

[0142] It is understood that in the various embodiments of this application, the sequence number of each process does not imply the order of execution. The execution order of each process should be determined by its function and internal logic, and should not constitute any limitation on the implementation process of this application.

[0143] It is understood that the various implementation methods described in this application can be implemented individually or in combination, and this application does not limit them.

[0144] Unless otherwise stated, all technical and scientific terms used in this application have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this application is for the purpose of describing particular embodiments only and is not intended to limit the scope of this application. The singular forms “a,” “the,” and “the” used in this application and the appended claims are also intended to include the plural forms unless the context clearly indicates otherwise.

[0145] Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented in electronic hardware or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design intervention conditions of the technical solution. Those skilled in the art can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

[0146] Those skilled in the art will clearly understand that, for the sake of convenience and brevity, the specific working processes of the embodiments described above can be referred to the corresponding processes in other embodiments, and will not be repeated here.

[0147] In the embodiments provided in this application, it should be understood that the disclosed systems, apparatuses, and methods can be implemented in other ways. For example, the apparatus embodiments described above are merely illustrative; for instance, the division of units is only a logical functional division, and in actual implementation, there may be other division methods. For example, multiple units or components may be combined or integrated into another system, or some features may be ignored or not executed. Furthermore, the coupling or direct coupling or communication connection shown or discussed may be through some interfaces; the indirect coupling or communication connection between apparatuses or units may be electrical, mechanical, or other forms.

[0148] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; that is, they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of the technical solution in this application, depending on actual needs.

[0149] In addition, the functional units in the various embodiments of this application can be integrated into one processing unit, or each unit can exist physically separately, or two or more units can be integrated into one unit.

[0150] If a function is implemented as a software functional unit and sold or used as an independent product, it can be stored in a computer-readable storage medium. Based on this understanding, the technical solution of this application, essentially, or the part that contributes to related technologies, or part of the technical solution, can be embodied in the form of a software product. The computer software product is stored in a storage medium and includes several instructions to cause a computer device (which may be a personal computer, server, or network device, etc.) to execute all or part of the steps of the methods described in the various embodiments of this application. The aforementioned storage medium includes various media capable of storing program code, such as USB flash drives, portable hard drives, read-only memory (ROM), random access memory (RAM), magnetic disks, or optical disks.

[0151] The above are merely specific embodiments of this application, but the scope of protection of this application is not limited thereto. Any variations or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application. Therefore, the scope of protection of this application should be determined by the scope of the claims.

Claims

1. A robot motion control method, characterized in that, include: Based on the robot's task target information and robot state information, a generative motion control model is used to determine the trajectory probability distribution information; wherein, the trajectory probability distribution information is used to characterize the probability distribution of the robot's motion trajectory in the target future time domain under the conditions corresponding to the task target information and the robot state information; Based on the trajectory probability distribution information, the target motion trajectory information is determined; Motion control data for the robot is generated based on the target motion trajectory information to control the robot's movement; The determination of trajectory probability distribution information based on the robot's task target information and robot state information through a generative motion control model includes: The following processing is performed using the generative motion control model: The task objective information is encoded to obtain task condition features; The robot state information is encoded to obtain state condition features; wherein, the state condition features include target state features corresponding to the target control cycle and a sequence of historical state features corresponding to multiple historical control cycles prior to the target control cycle; Based on the task target information and the robot state information, visible motion constraints and target completion motion trajectory components are determined, and the visible motion constraints and target completion motion trajectory components are encoded to obtain motion condition features; Based on the state condition features, the task condition features, and the motion condition features, the trajectory probability distribution information is determined; The visible motion constraints refer to the motion constraints that have been given during the motion trajectory generation process and need to be consistent in the generation result; the target completion motion trajectory components refer to the future motion trajectory components that have not been fully specified and need to be completed by the generative motion control model under the conditions corresponding to the task target information and the robot state information.

2. The robot motion control method according to claim 1, characterized in that, The encoding of the visible motion constraints and the target completed motion trajectory components to obtain motion condition features includes: The visible motion constraints are conditionally encoded to obtain motion condition encoding information; Set occlusion markers for the completed motion trajectory components of the target; The motion condition features are determined based on the motion condition encoding information and the occlusion identifier.

3. The robot motion control method according to claim 1, characterized in that, The determination of the trajectory probability distribution information based on the state condition features, the task condition features, and the motion condition features includes: Based on the topological relationship of the robot's motion nodes, spatial encoding is performed on the features corresponding to multiple motion parts in the state condition features to obtain spatial coupling features; Based on the sequential relationship of control cycles corresponding to the historical state feature sequence, the historical state feature sequence is temporally encoded to obtain temporal motion features; Based on the spatial coupling characteristics, the temporal motion characteristics, the task condition characteristics, and the motion condition characteristics, the generation condition characteristics are determined; Based on the aforementioned generation condition features, the trajectory probability distribution information is determined.

4. The robot motion control method according to claim 3, characterized in that, The step of performing temporal encoding on the historical state feature sequence based on the control cycle sequence corresponding to the historical state feature sequence to obtain temporal motion features includes: generating first temporal causal constraint information based on the control cycle sequence corresponding to multiple historical state features in the historical state feature sequence; and performing attention calculation on multiple historical state features in the historical state feature sequence based on the first temporal causal constraint information to obtain the temporal motion features. The step of determining the generation condition feature based on the spatial coupling feature, the temporal motion feature, the task condition feature, and the motion condition feature includes: determining the future time domain query feature corresponding to each time step in the target future time domain based on the spatial coupling feature, the temporal motion feature, the task condition feature, and the motion condition feature; generating second temporal causal constraint information based on the sequential relationship between each time step in the target future time domain; performing attention update on the future time domain query feature based on the second temporal causal constraint information to obtain updated query feature; and determining the generation condition feature based on the updated query feature, the spatial coupling feature, the temporal motion feature, the task condition feature, and the motion condition feature.

5. The robot motion control method according to claim 3, characterized in that, The step of determining the trajectory probability distribution information based on the generation condition features includes: Construct the first trajectory variable corresponding to the future time domain of the target; Based on the aforementioned generation condition features, conditional generation iterations are performed on the first trajectory variable to obtain the trajectory generation result; Based on the trajectory generation results, the trajectory probability distribution information is determined.

6. The robot motion control method according to claim 5, characterized in that, The step of performing conditional generation iterations on the first trajectory variable based on the generation condition features to obtain the trajectory generation result includes: Based on the aforementioned generation condition features, the first trajectory variable undergoes at least one conditional denoising process to obtain the trajectory generation result; or... Based on the generated condition features, a conditional flow field is determined, and the first trajectory variable is updated at least once based on the conditional flow field to obtain the trajectory generation result.

7. The robot motion control method according to claim 1, characterized in that, Determining the target motion trajectory information based on the trajectory probability distribution information includes: Based on the trajectory probability distribution information, multiple candidate motion trajectory information are generated; The feasibility of motion is evaluated for each of the multiple candidate motion trajectory information to obtain the trajectory evaluation result corresponding to each candidate motion trajectory information. Based on the trajectory evaluation results, the target motion trajectory information is determined from the plurality of candidate motion trajectory information; The motion feasibility evaluation includes at least one constraint evaluation based on joint motion restrictions, drive restrictions, contact state restrictions, stability conditions, and task matching conditions.

8. The robot motion control method according to claim 1, characterized in that, The step of generating motion control data for the robot based on the target motion trajectory information to control the robot's movement includes: Determine the reference motion state corresponding to the target control cycle from the target motion trajectory information; The motion control data is generated based on the reference motion state and the robot's feedback motion state during the target control cycle. After controlling the robot to execute the motion control data, receive the robot's updated status information; Based on the task objective information and the updated state information, the updated trajectory probability distribution information is re-determined through the generative motion control model to perform motion control on the robot in the next control cycle.

9. The robot motion control method according to any one of claims 1 to 8, characterized in that, The generative motion control model is a model trained in the following way: Acquire heterogeneous motion data, which includes at least two of the following: human motion data, robot simulation motion data, and actual robot motion data; The heterogeneous motion data is converted into training trajectory data that matches the moving parts of the robot; Training samples are constructed based on the training trajectory data; wherein, the training samples include sample task target information, sample robot state information, and sample future motion trajectory; Based on the sample task target information and the sample robot state information, the sample predicted trajectory probability distribution information is determined through an initial generative motion control model, and the trajectory generation loss is determined based on the sample predicted trajectory probability distribution information and the sample future motion trajectory. The model parameters of the initial generative motion control model are adjusted based on the trajectory generation loss to obtain the generative motion control model.

10. A robot motion control device, characterized in that, include: The model inference module is used to determine the trajectory probability distribution information based on the robot's task target information and robot state information through a generative motion control model; wherein, the trajectory probability distribution information is used to characterize the probability distribution of the robot's motion trajectory in the target future time domain under the conditions corresponding to the task target information and the robot state information; The trajectory determination module is used to determine the target motion trajectory information based on the trajectory probability distribution information; A motion control module is used to generate motion control data for the robot based on the target motion trajectory information, so as to control the robot's movement; The determination of trajectory probability distribution information based on the robot's task target information and robot state information through a generative motion control model includes: The following processing is performed using the generative motion control model: The task objective information is encoded to obtain task condition features; The robot state information is encoded to obtain state condition features; wherein, the state condition features include target state features corresponding to the target control cycle and a sequence of historical state features corresponding to multiple historical control cycles prior to the target control cycle; Based on the task target information and the robot state information, visible motion constraints and target completion motion trajectory components are determined, and the visible motion constraints and target completion motion trajectory components are encoded to obtain motion condition features; Based on the state condition features, the task condition features, and the motion condition features, the trajectory probability distribution information is determined; The visible motion constraints refer to the motion constraints that have been given during the motion trajectory generation process and need to be consistent in the generation result; the target completion motion trajectory components refer to the future motion trajectory components that have not been fully specified and need to be completed by the generative motion control model under the conditions corresponding to the task target information and the robot state information.

11. A robot, characterized in that, The robot includes a control module for performing the steps of the method according to any one of claims 1 to 9.

12. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 9.

13. A computer program product, characterized in that, The computer program product includes a computer program that, when executed by a processor, implements the steps of the method according to any one of claims 1 to 9.

Citation Information

Patent Citations

  • Heavy-load robot motion trail method and system based on machine learning

    CN121315961A

  • Trajectory prediction method, trajectory prediction model training method and apparatus

    WO2026118764A1