Quadruped wheel-legged robot and control method thereof

Abstract

This application relates to a quadruped wheeled robot and its control method, comprising a body and four sets of wheel and leg assemblies respectively disposed around the body. The quadruped wheeled robot of this application eliminates the lateral swing joints commonly found in traditional quadruped robots, retaining only a single-degree-of-freedom lateral swing joint in each set of wheel and leg assemblies, allowing each set of wheel and leg assemblies to only perform outward flexion movements around its corresponding lateral swing axis. This structure significantly reduces the number of robot leg joints, transmission mechanisms, and drive motors, thereby reducing the overall structural complexity and manufacturing cost, and simplifying motion control logic. By rationally utilizing the changes in support relationships, center of gravity shifts, and contact between the body and environment caused by the outward flexion of the wheel and leg assemblies, the robot's body structure participates in the motion generation process. The robot can utilize its wheels, joint structures, and body itself to complete environmental adaptation and movement tasks, thus achieving strong terrain-crossing capabilities with fewer degrees of freedom.

Description

Technical Field

[0001] This application relates to the field of robotics, and in particular to a quadrupedal wheeled robot and its control method. Background Technology

[0002] In complex land environments (such as ditches, steps, and raised surfaces), quadruped robots face the dual challenges of terrain adaptability and motion stability. Their leg structures must simultaneously meet the technical requirements of "efficient obstacle crossing, adjustable posture, and controllable self-rescue" to cope with the risks of bumps, collisions, and tipping over. Currently, mainstream leg structure technologies can be mainly divided into the following two categories:

[0003] One approach is the wheel-legged conversion route. This route uses sensors to detect terrain features and the control system autonomously or remotely triggers the switch between wheeled and legged modes, relying on the collaborative work of multi-degree-of-freedom joint drive mechanisms and wheel transmission components. Its advantage lies in combining the high speed of wheeled vehicles with the strong adaptability of legged vehicles, enabling flexible movement in complex environments. However, this structure typically requires multi-joint drive mechanisms, resulting in numerous critical components, complex structural coupling, high overall manufacturing costs, and cumbersome assembly processes. The collaborative control of multiple joints relies on complex algorithms, making it difficult to guarantee motion synchronization and prone to execution deviations. Furthermore, after the robot tipps over, it often needs to switch back to legged mode for posture recovery, a time-consuming self-rescue process that is difficult to adapt to continuous operation or unattended scenarios.

[0004] Secondly, there are routes where the wheels are permanently attached to the end of the legs. This design permanently fixes the wheels to the legs, requiring no mode switching and adapting to the terrain solely through hip yaw, always moving in a wheeled rolling motion. However, its mobility is limited by the hip yaw angle and wheel size. In small-diameter structures, the yaw angle is generally no more than ±90°, making it prone to getting stuck when facing wide ditches or high steps, resulting in insufficient crossing ability; while large-diameter structures can increase the obstacle crossing height, the smooth wheel surface makes it easy to slip on the edge of ditches or loose ground, reducing reliability. In addition, this structure generally lacks the ability to autonomously recover after tipping over and to continue driving; once overturned, manual intervention is required, seriously affecting mission continuity and reliability.

[0005] In summary, how to reduce the overall structural complexity and manufacturing cost, and simplify the motion control logic, are technical problems that urgently need to be solved by those skilled in the art. Summary of the Invention

[0006] The purpose of this application is to provide a quadrupedal wheeled robot and its control method, so as to reduce the complexity of the overall structure and manufacturing cost, and simplify the motion control logic.

[0007] To achieve the above objectives, this application provides the following technical solution:

[0008] In a first aspect, this application provides a quadruped wheeled robot, including a body and four sets of wheeled leg assemblies respectively disposed around the body;

[0009] Each wheel-leg assembly includes a hip joint connector, a hip joint drive module, a wheel drive module, and a wheel.

[0010] One end of the hip joint connector is connected to the machine body, and the other end is connected to the hip joint drive module;

[0011] The hip joint drive module is a single-degree-of-freedom lateral joint, used to drive the corresponding wheel-leg assembly to fold outward relative to the body in the lateral direction.

[0012] The wheel drive module is connected to the wheel transmission and is used to drive the wheel to rotate continuously;

[0013] The four wheel-leg assemblies are arranged symmetrically relative to the body. The hip joint drive module has the range of motion to flip the wheel-leg assemblies outward and re-ground them after tipping over, so that the robot can travel by grounding the wheel body in both normal and tipped-over postures.

[0014] As can be seen from the above technical solution, the quadrupedal wheel-legged robot in this application eliminates the lateral swing joints commonly found in traditional quadrupedal robots, retaining only the single-degree-of-freedom lateral swing joint in each wheel-leg assembly. This allows each wheel-leg assembly to only perform outward turning movements around its corresponding lateral swing axis. This structure can significantly reduce the number of robot leg joints, transmission mechanisms, and drive motors, thereby reducing the overall structural complexity and manufacturing cost, and simplifying motion control logic.

[0015] In some embodiments, the yaw rate α of the hip joint drive module ranges from 0° to 180°.

[0016] In some embodiments, the hip joint connector is a hollow structure used to accommodate the cable connecting the body and the corresponding wheel leg assembly.

[0017] In some embodiments, the hip joint connector is made of carbon fiber reinforced composite material.

[0018] In some embodiments, the wheel is a tracked grooved wheel, and the tread of the wheel has a continuous groove structure.

[0019] In some embodiments, the continuous groove structure is a trapezoidal groove structure that is continuously distributed along the circumference of the wheel body.

[0020] In some embodiments, the wheels are symmetrical with no distinction between front and back, so that the wheels still have the same grounding characteristics after the robot tipps over.

[0021] In some embodiments, the wheel drive module is a 360° rotating brushless DC motor capable of continuously switching between forward and reverse directions.

[0022] In some embodiments, the four wheel leg assemblies are a front left wheel leg assembly, a front right wheel leg assembly, a rear left wheel leg assembly, and a rear right wheel leg assembly, and the hip joint lateral swing direction of the two front wheel leg assemblies is opposite to that of the two rear wheel leg assemblies.

[0023] In some embodiments, the robot includes a total of eight motors, including four hip joint yaw drive motors and four wheel drive modules.

[0024] In some embodiments, the body is a one-piece rectangular shell, and connecting flanges for connecting hip joint connectors are symmetrically arranged on the left and right sides of the body.

[0025] In some embodiments, the upper and lower surfaces of the robot have the same or symmetrical geometric and functional structures, so that the robot can continue to operate without changing the working mode of the main control system and power supply system after it has tipped over.

[0026] In some embodiments, a symmetrical sensor interface is provided on the outside of the body, which is used to connect a lidar and / or a depth camera.

[0027] In some embodiments, the main control module, power management module, and battery pack are installed inside the device;

[0028] The main control module is configured to determine the robot's current up and down orientation based on feedback from the attitude sensor, and adjust the rotation direction of the wheel drive module accordingly.

[0029] And / or, the main control module is configured to control the hip joint drive module to perform an eversion action when the robot flips over, so that the wheels re-contact the ground, and control the wheel drive module to output driving force to achieve autonomous movement after the flip;

[0030] And / or, the main control module is configured to control the four hip joint drive modules to rotate outward to 90°, so that the four wheel leg assemblies are in a horizontal position close to the body, so as to achieve a crouching and / or storage posture.

[0031] And / or, the main control module is configured to control each hip joint drive module to adjust the yaw angle according to the detected width of the ditch when crossing the ditch, so as to increase the wheel track, and control the wheel drive module to output torque to push the robot across the ditch;

[0032] And / or, the main control module is configured to control the four sets of wheel-leg components to alternately support the steps when climbing, so that the robot can gradually complete the climbing of the steps by transferring the center of gravity between the front and rear wheels;

[0033] And / or, the main control module is equipped with a policy network trained by reinforcement learning. The policy network is used to output control commands to control the four hip joint drive modules and the four wheel drive modules based on the robot's state variables.

[0034] In some embodiments, when a policy network trained through reinforcement learning is deployed within the main control module, the robot state variables include one or more of the following: IMU attitude information, yaw angle of each hip joint, wheel motor speed, center of mass height, obstacle height estimation, wheel contact state, motor current feedback, and action command from the previous control cycle.

[0035] In some embodiments, the policy network outputs eight-dimensional control commands, which correspond to the target values ​​of the hip joint yaw angle of the four wheel-leg assemblies and the target rotational speed of the four wheel drive modules.

[0036] In some embodiments, when a policy network trained through reinforcement learning is deployed in the main control module, the main control module performs amplitude limiting processing on the control commands output by the policy network and then sends them to the hip joint drive module and the wheel drive module.

[0037] Secondly, this application provides a tipping self-rescue control method for a quadrupedal wheeled robot, applicable to any of the quadrupedal wheeled robots described above, the method comprising:

[0038] Acquire robot posture information and determine whether the robot is in a tipped-over state;

[0039] When the robot is in a tipped-over state, the four sets of hip joint drive modules are controlled to drive the corresponding wheel and leg components to flip outward, so that the wheels change from an ungrounded state to a grounded state.

[0040] Adjust the rotation direction and speed of each wheel drive module according to the robot's current posture and wheel grounding status;

[0041] The robot outputs driving force through the wheel drive module, enabling it to recover its drivable state or continue moving after tipping over.

[0042] In some embodiments, controlling the four sets of hip joint drive modules to drive the corresponding wheel-leg assemblies to rotate outward includes controlling the wheel-leg assemblies to rotate outward to a position close to 180°.

[0043] In some embodiments, adjusting the rotation direction and speed of each wheel drive module includes: automatically determining the robot's current up-down orientation based on feedback from the attitude sensor, and dynamically switching the forward and reverse directions of the wheel drive modules.

[0044] In some embodiments, the method further includes: inputting attitude information, wheel contact state, hip joint angle, and wheel motor speed into a reinforcement learning policy network, and having the reinforcement learning policy network output the corresponding target value of hip joint yaw angle and target speed of wheel motor.

[0045] Thirdly, this application provides a step-climbing control method for a quadrupedal wheeled robot, applicable to any of the quadrupedal wheeled robots described above, the method comprising:

[0046] Acquire robot posture information, step height information, wheel contact status, and the current status of each wheel and leg assembly;

[0047] Control at least one front wheel leg assembly near the step to establish a support point;

[0048] Control the robot's center of gravity to shift towards the front wheel-leg assembly, which has already established a support point;

[0049] Control the remaining wheel leg components to adjust the sway angle sequentially and conform to the surface of the step;

[0050] The robot climbs stairs by using a single-leg alternating support method, driven by four wheel drive modules that output driving force.

[0051] In some embodiments, the yaw angle of each wheel leg assembly and the output torque of the wheel drive module are continuously updated within the control cycle.

[0052] In some embodiments, the stair climbing control method generates control commands through a reinforcement learning policy network. The reward constraints of the reinforcement learning policy network include one or more of the following: wheel contact with the stair surface, maintaining support stability, increasing front wheel contact force, reducing posture deviation, and constraining motion smoothness.

[0053] Fourthly, this application provides a ravine-crossing control method for a quadrupedal wheeled robot, applicable to any of the quadrupedal wheeled robots described above, the method comprising:

[0054] Detect the width of the ditch to be crossed;

[0055] The yaw angle of the corresponding wheel and leg assembly is adjusted by controlling the width of the groove to increase or adjust the robot's wheelbase.

[0056] Control the output driving force of each wheel drive module to make the wheel with a continuous groove structure form a gripping contact with the ground and / or the edge of the ditch;

[0057] Drive the robot across the ravine.

[0058] Fifthly, this application provides a storage and control method for a quadrupedal wheeled robot, applicable to any of the quadrupedal wheeled robots described above, the method comprising:

[0059] The four hip joint drive modules are controlled to drive the corresponding wheel and leg components to swing outward to the preset storage angle.

[0060] Position the four wheel and leg assemblies horizontally and close to the body;

[0061] The preset storage angle is 90° or close to 90° to reduce the overall space occupied by the robot.

[0062] Sixthly, this application provides a reinforcement learning control method for a quadrupedal wheeled robot, applied to any of the quadrupedal wheeled robots described above, the method comprising:

[0063] Collect the robot's state variables in the simulation environment. These state variables include one or more of the following: attitude, position, velocity, sensor data, and joint position.

[0064] The motion space is defined by the hip joint angle and the wheel rotation speed;

[0065] The policy network is trained based on the reward function;

[0066] Deploy the trained policy network to the robot's main control module;

[0067] The strategy network outputs target values ​​for the yaw angles of the four hip joints and target rotational speeds of the four wheel drive modules during the real-time control cycle to control the robot's movement.

[0068] In some embodiments, the motion space is an eight-dimensional motion vector, which includes four hip joint angle parameters and four wheel rotation speed parameters.

[0069] In some embodiments, the reward function includes a target error penalty term, a power consumption compensation term, and an action change penalty term.

[0070] In some embodiments, the reward function is:

[0071]

[0072] in, It is the reward value given state s and action a;

[0073] It is a penalty term for target error, including the penalty for tracking speed error;

[0074] It is a power consumption compensation, including energy consumption penalty, joint torque penalty, etc.;

[0075] It is a penalty for excessive movement, which is related to the difference between the current position of each joint and the position of the previous time step.

[0076] These are the weighting coefficients for each reward.

[0077] In some embodiments, the simulation environment includes multi-scene, multi-terrain, and multi-gait simulations, and simulates one or more of external disturbances, load changes, and terrain irregularities. Attached Figure Description

[0078] To more clearly illustrate the technical solutions in the embodiments of this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0079] Figure 1 This is a schematic diagram of the robot provided in an embodiment of this application;

[0080] Figure 2 This is a schematic diagram of the structure of the robot drive assembly provided in an embodiment of this application;

[0081] Figure 3 A schematic diagram of the structure of the part of the robot body used to connect the drive assembly, provided in an embodiment of this application;

[0082] Figure 4 This is a structural diagram of a robot climbing stairs according to an embodiment of this application;

[0083] Figure 5 This is a schematic diagram of the robot's structure when it finishes its task, as provided in an embodiment of this application.

[0084] Figure 6 A flowchart illustrating a self-rescue control method for tipping over provided in an embodiment of this application;

[0085] Figure 7 A flowchart illustrating a stair climbing control method provided in an embodiment of this application;

[0086] Figure 8 A flowchart illustrating a ravine crossing control method provided in an embodiment of this application;

[0087] Figure 9 A schematic flowchart illustrating a storage control method provided in an embodiment of this application;

[0088] Figure 10 A flowchart illustrating a reinforcement learning control method provided in an embodiment of this application;

[0089] The annotations in the attached figures are explained as follows:

[0090] 1. Body, 2. Wheel and leg assembly, 3. Steps;

[0091] 11 Connecting flange, 21 Hip joint connector, 22 Hip joint drive module, 23 Wheel drive module, 24 Wheel. Detailed Implementation

[0092] The technical solution of this application will be further described in detail below with reference to the accompanying drawings. It should be understood that the following embodiments are only used to illustrate this application and are not intended to limit the scope of protection of this application. Equivalent substitutions, modifications or improvements made by those skilled in the art based on the disclosure of this application without departing from the concept of this application shall fall within the scope of protection of this application.

[0093] As described in the background section, both wheel-leg structures relying on complex mode-changing mechanisms and simplified structures using fixed wheels have significant limitations when dealing with complex land-based scenarios. On the one hand, while wheel-leg conversion structures are more adaptable, they generally suffer from complex multi-joint coordination, high manufacturing costs, complex control algorithms, and lengthy self-rescue times after tipping over. On the other hand, while fixed-wheel structures offer mechanical simplification, their motion performance is limited by the joint range of motion and wheel diameter, and they generally lack effective self-recovery capabilities after tipping over, severely impacting the continuity and reliability of missions.

[0094] In order to reduce the complexity of the overall structure and manufacturing cost, and to simplify the technical problem of motion control logic, this application discloses a quadruped wheeled robot (hereinafter referred to as the robot).

[0095] like Figures 1 to 5 As shown, the robot includes a body 1 and four sets of wheel-leg assemblies 2 respectively located at the four corners or sides of the body 1. The four sets of wheel-leg assemblies 2 correspond to the four support positions of the robot's front left, front right, rear left, and rear right, and are arranged symmetrically with respect to the body 1 in terms of front-back and left-right orientation. Each set of wheel-leg assemblies 2 includes a hip joint connector 21, a hip joint drive module 22, a wheel drive module 23, and a wheel 24.

[0096] The quadruped wheeled robot of this application eliminates the lateral swing joints commonly found in traditional quadruped robots, retaining only a single-degree-of-freedom lateral swing joint in each wheel-leg assembly 2. This allows each wheel-leg assembly 2 to only perform outward flexion movements around its corresponding lateral swing axis. This structure significantly reduces the number of robot leg joints, transmission mechanisms, and drive motors, thereby reducing the overall structural complexity and manufacturing cost, and simplifying motion control logic. Unlike traditional designs that enhance motion capabilities by increasing joint degrees of freedom, this low-cost configuration utilizes the changes in support relationships, center of gravity shifts, and contact between the robot body and its environment resulting from the outward flexion of the wheel-leg assemblies to enable the robot's body structure to participate in the motion generation process. Through the synergistic cooperation of structural design and control strategies, the robot can utilize its wheels, joint structures, and body itself to complete environmental adaptation and motion tasks, thus achieving strong terrain-crossing capabilities with fewer degrees of freedom.

[0097] It should be noted that the above-mentioned four sides can be understood as the four corner positions of the body 1, that is, the four sets of wheel leg assemblies are arranged in a centrally symmetrical or approximately centrally symmetrical manner with the body as the center.

[0098] Specifically, please refer to Figure 1 and Figure 3 The body typically has a rectangular or near-rectangular outline. Each set of wheel leg assemblies 2 is connected to the four corners of the body outline or in the direction of extension via its hip joint connectors 21, thereby forming a stable and balanced rectangular support layout from a top view.

[0099] like Figure 1 As shown, the body 1 is the main structure of the robot, which preferably adopts a cuboid or near-cuboid shell structure. Body 1 can be made of aluminum alloy to balance structural strength, heat dissipation performance, and lightweight requirements. The body 1 may house a main control module, a power management module, a battery pack, an inertial measurement unit, a communication module, and a motor drive control module. The main control module receives sensor signals, runs motion control algorithms, and outputs control commands to the hip joint drive module 22 and the wheel drive module 23.

[0100] The robot body 1 has connecting flanges 11 on its left and right sides for mounting the hip joint connector 21. The connecting flanges 11 are preferably standardized flange structures with multiple bolt mounting holes to facilitate quick installation, disassembly, and replacement of the hip joint connector 21. This standardized connection method improves the versatility and ease of maintenance of robot components.

[0101] In some embodiments of this application, the upper and lower surfaces of the robot body 1 adopt a symmetrical structure with essentially the same geometry and functional interfaces. For example, sensor mounting interfaces, communication interfaces, or expansion interfaces can be provided on both the upper and lower surfaces, enabling the robot to maintain essentially the same environmental perception and functional expansion capabilities in both normal and overturned postures. Therefore, the robot can continue to perform its movement tasks without manual intervention after overturning.

[0102] The robot body 1 can also be equipped with mounting interfaces for external sensing devices such as LiDAR, depth cameras, ultrasonic sensors, collision sensors, or vision sensors. These sensors can be used to detect the width of ravines, the height of steps, the location of obstacles, the slope of the ground, and information about the robot's surrounding environment, providing input data for the robot's motion control.

[0103] like Figure 2 and Figure 3 As shown, each wheel-leg assembly 2 includes a hip joint connector 21, a hip joint drive module 22, a wheel drive module 23, and a wheel 24.

[0104] One end of the hip joint connector 21 is fixedly connected to the connecting flange 11 on the side wall of the machine body 1, and the other end is connected to the hip joint drive module 22. The hip joint connector 21 is preferably a straight rod structure, forming a rigid connecting rod without additional knee, ankle, or lateral joints. The hip joint connector 21 can be made of carbon fiber reinforced composite material, aluminum alloy, or other lightweight, high-strength materials. In a preferred embodiment, the hip joint connector 21 has a hollow structure, forming a wiring channel inside for arranging the power supply and signal cables of the hip joint drive module 22, wheel drive module 23, and related sensors, thereby reducing exposed cables and improving overall machine reliability and protection performance.

[0105] The hip joint drive module 22 constitutes a single-degree-of-freedom yaw joint for each wheel-leg assembly 2. The hip joint drive module 22 includes a yaw drive motor or servo, a reduction mechanism, an angle sensor, and a joint housing. The output end of the hip joint drive module 22 is connected to the hip joint connector 21 or the wheel drive part, enabling the wheel-leg assembly 2 to perform outward flexion relative to the body 1 around the yaw axis.

[0106] In some embodiments of this application, the hip joint drive module 22 employs a high-precision 180° servo motor, with an output angle α ranging from 0° to 180°. Here, 0° can be defined as the wheel-leg assembly 2 being in a normal or basically vertical support state; 90° can be defined as the wheel-leg assembly 2 being in an outwardly horizontally extended or downwardly retracted state; and 180° can be defined as the wheel-leg assembly 2 being in a maximum outward tilt state, used for self-rescue after tipping over or for reverse support. It should be noted that the above angle definitions are for illustrative purposes only, and those skilled in the art can adjust the zero-angle position according to the actual structural installation direction of the robot.

[0107] The hip joint drive module 22 in this application only allows the wheel-leg assembly 2 to rotate outward, without any inward retraction. By restricting the direction of joint movement, interference between the wheels or between the wheels and the body 1 can be avoided, while reducing the complexity of the joint mechanism and improving structural reliability.

[0108] In the four sets of wheel-leg assemblies 2 (front left, front right, rear left, and rear right), the rotation direction of the hip joints can be set according to the symmetry of the robot's structure. For example, the lateral swing direction of the two front wheel-leg assemblies can be defined as clockwise, and the lateral swing direction of the two rear wheel-leg assemblies can be defined as counterclockwise, or the opposite can be set according to the actual assembly method. Through this symmetrical outward-folding design, the robot can maintain better posture balance during deployment, storage, crossing ditches, and self-rescue after tipping over.

[0109] Wheel body 24 is located at the end of each wheel leg assembly 2 and is fixedly connected to the output end of wheel body drive module 23. Wheel body drive module 23 preferably uses a brushless DC motor capable of continuous 360° rotation, which can drive wheel body 24 to rotate continuously in either the forward or reverse direction. Wheel body drive module 23 and wheel body 24 can be fixedly connected via metal drive shaft, splined shaft, key connection, expansion sleeve, or flange connection to ensure stable and reliable power transmission.

[0110] In some embodiments of this application, the diameter of the wheel 24 is approximately 30 cm. Compared to smaller diameter wheels, this larger diameter wheel improves the robot's ability to cross ditches and step edges, reducing the risk of the wheel getting stuck in ditches or obstacles. It should be noted that 30 cm is a preferred size. While maintaining the same functionality, the wheel diameter can be appropriately adjusted according to the robot's overall size, load requirements, and application scenario, for example, it can be set to a range of 25 cm to 40 cm.

[0111] The tread of wheel 24 is provided with a continuous groove structure. Preferably, the continuous groove structure is a trapezoidal groove structure continuously distributed along the circumference of the wheel. The trapezoidal grooves can extend along the axial direction of the wheel or be arranged at a certain angle relative to the axial direction of the wheel. The continuous trapezoidal grooves create a non-smooth contact surface on the wheel surface, similar to a track, thereby increasing the friction and grip between the wheel and the ground. This structure is particularly suitable for complex contact environments such as the edges of ditches, loose ground, muddy ground, gravel roads, and the edges of steps, and can effectively reduce the risk of slippage of traditional smooth wheels.

[0112] The wheel 24 preferably adopts a symmetrical structure without a distinction between front and back sides, meaning that the shape, tread contact characteristics, and installation force characteristics of both sides of the wheel are basically the same. Therefore, whether the robot is in a normal posture or a tipped posture, the wheel 24 can participate in driving with the same or nearly the same grounding performance, without the need for a special front-to-back switching mechanism.

[0113] The drive system of this application includes four hip joint drive modules 22 and four wheel drive modules 23, therefore the whole machine preferably has a total of eight active drive units. The four hip joint drive modules 22 are used to control the yaw angle of the four sets of wheel leg assemblies 2 respectively; the four wheel drive modules 23 are used to control the rotation speed and steering of the four wheels 24 respectively.

[0114] The main control module inside the body 1 is communicatively connected to four hip joint drive modules 22, four wheel drive modules 23, and various sensors. The sensors include at least an inertial measurement unit for acquiring the robot's attitude angles, angular velocities, and accelerations, and may also include joint angle encoders, wheel speed sensors, motor current sensors, contact sensors, vision sensors, or distance sensors.

[0115] The main control module calculates the robot's current posture, wheel and leg positions, wheel speed, contact status, and environmental obstacle information in real time based on sensor feedback, and outputs the corresponding target values ​​for hip joint angle and wheel speed. Control commands are sent to each drive unit after being processed by amplitude limiting, filtering, or safety protection checks, thereby achieving stable robot movement.

[0116] In some embodiments of this application, the main control module deploys a policy network trained through reinforcement learning. The input state variables of the policy network may include robot posture information, position, velocity, yaw angles of each hip joint, wheel motor speeds, center of mass height, obstacle height estimation, groove width estimation, wheel contact state, motor current feedback, and action commands from the previous control cycle. The output of the policy network is an eight-dimensional motion vector, which corresponds to the target yaw angles of the four hip joint drive modules 22 and the target speeds of the four wheel drive modules 23, respectively.

[0117] Reinforcement learning policy networks can be trained in a simulation environment. The training environment can include scenarios such as flat ground, slopes, ditches, steps, gravel, loose ground, and random perturbations. During training, the robot agent outputs actions based on its current state and receives rewards or penalties according to a reward function, thereby continuously optimizing the policy network parameters.

[0118] The reward function can be expressed as:

[0119]

[0120] in, It is the reward value given state s and action a;

[0121] The penalty term representing the target error may include linear velocity tracking error, angular velocity tracking error, or attitude error;

[0122] Compensation for power consumption can include energy consumption penalty, joint torque penalty, or motor current penalty;

[0123] This indicates a penalty for changes in motion, which can be used to constrain excessive changes in joint position or wheel speed within adjacent control cycles.

[0124] These are the weighting coefficients for each reward.

[0125] Through the aforementioned reinforcement learning training method, the policy network can generate continuous, smooth, and stable control actions under different terrain and posture abnormalities, enabling the robot to complete driving, obstacle crossing, and self-rescue without relying on manually preset fixed action sequences. Furthermore, this application does not enhance environmental adaptability by increasing joint degrees of freedom, but rather through the collaborative design of the robot's body structure and control strategy, enabling the robot body to actively participate in motion generation and task execution. Specifically, the outward flipping of the wheel-leg assembly can cause changes in wheel track, support domain, and center of gravity distribution; in complex terrain environments, the wheels, hip joint connection structure, joint shell, and body shell can all act as supports, participating in force and environmental interaction, thus forming multiple support modes such as wheel support, joint support, and body support.

[0126] During reinforcement learning training, the policy network not only learns the control laws of wheel-driven and joint-driven motion, but also learns to autonomously select and utilize different body parts to participate in movement based on environmental conditions. For example, when crossing ditches, it uses wheelbase expansion to create a larger support domain; when climbing stairs, it uses weight transfer and joint support to achieve stable obstacle crossing; when falling over and needing to recover, it uses the body, wheels, and joint structures to re-establish support relationships; and in storage or special working conditions, it uses the contact between the body and the ground to adjust its posture. Through these methods, reinforcement learning strategies can automatically tap into the motion potential inherent in the robot's body structure under low-degree-of-freedom configurations, enabling some environmental adaptability to be achieved through the physical interaction between the body configuration and the environment, rather than relying entirely on complex mechanical structures and high-degree-of-freedom control. This reduces manufacturing costs and control complexity while improving the robot's adaptability to complex environments.

[0127] When the robot detects a ditch ahead during its movement, it can estimate the ditch's width and edge position using visual sensors, depth cameras, LiDAR, or other distance sensors. Based on the ditch width, the robot's current speed, and the wheel contact state, the main control module controls the four hip joint drive modules 22 to rotate outward at a certain angle to increase the left-right wheelbase or the front-rear support range.

[0128] During the crossing of gullies, the wheel 24 continuously outputs driving force at a set torque or set speed. Because the wheel 24 has a large diameter and a continuous trapezoidal groove tread, it can form a more reliable grip contact at the edge of the gully, thereby reducing the probability of the wheel slipping, spinning freely, or getting stuck. Simultaneously, the hip joint drive module 22 can make fine adjustments based on the robot's posture feedback, ensuring that the wheel-leg assembly 2 maintains a reasonable support angle and preventing excessive shift of the robot's center of gravity.

[0129] In one embodiment, when the width of the ditch is detected to be close to the upper limit of the robot's default wheelbase, the main control module controls the left and right wheel leg assemblies 2 to flip outwards respectively, so that the four wheels 24 form a wider support layout. The robot then outputs continuous rolling power through the wheel drive module 23, so that the front wheel crosses the ditch first, and then the rear wheel crosses the ditch under the traction of the front wheel and its own driving force. Throughout the process, the main control module continuously monitors the IMU attitude changes and wheel contact status, and dynamically adjusts the wheel speed and hip joint angle. IMU attitude specifically refers to the spatial angular state of the robot body relative to the gravity reference frame, which is calculated in real time by the Inertial Measurement Unit (IMU). The core parameters are roll angle and pitch angle. It directly describes whether the robot body is in a normal horizontal posture or has tilted laterally, pitched forward or backward, or even completely overturned. It is the most direct and critical sensory input for the main control module to determine motion stability, coordinate obstacle crossing actions, and trigger overturn self-rescue and recovery strategies.

[0130] like Figure 4 As shown, during the climbing of step 3, the robot of this application completes the climbing action by alternating single-leg support or dog-like step-by-step support. Unlike traditional foot climbing that relies on multi-degree-of-freedom leg mechanisms, this application achieves the contact between the wheel and the surface of step 3, the transfer of support points, and the shift of the overall center of gravity of the robot only through the cooperation of a single-degree-of-freedom lateral joint and the wheel drive module 23.

[0131] As the robot approaches step 3, sensors detect the height and edge position of step 3. Based on the height of step 3, the main control module controls the hip joint drive module 22 of the front wheel-leg assembly 2 to either flex outwards or retract, causing the front wheel 24 to contact the vertical surface or upper surface of step 3. The wheel drive module 23 then outputs driving force, causing the front wheel to roll along the edge of step 3 and gradually ascend it.

[0132] During the climbing process, the robot can first establish a support point with the front wheel closest to step 3, and then gradually shift the center of gravity of the whole machine to the front support point by adjusting the yaw angle of the remaining wheel leg assembly 2. After one or two sets of front wheels enter the upper surface of step 3, the rear wheel leg assembly 2 continues to provide thrust, and with the adjustment of the yaw angle of the hip joint, the rear wheel 24 fits against the edge of step 3 and completes the step-up movement.

[0133] In implementations deploying reinforcement learning policy networks, the stair-climbing motion does not rely on a pre-written fixed sequence of actions, but is generated by the policy network based on real-time state loops. In each control cycle, the policy network outputs four target hip joint angles and four target wheel rotation speeds based on inputs such as IMU attitude, wheel contact state, stair height estimation, hip joint angle, and wheel rotation speed. By incorporating constraints into the training rewards, such as wheel contact with the stair surface, maintaining stable support, reducing posture deviation, minimizing abrupt changes in motion, and reducing energy consumption, the policy network can generate stable climbing motions under different stair heights and surface conditions.

[0134] The robot described in this application has a structure without a fixed up-down orientation. When the robot overturns due to terrain bumps, obstacle collisions, or external forces, the inertial measurement unit inside the robot body 1 detects that the robot is in an inverted or abnormal posture. The main control module initiates a self-rescue control mechanism based on the posture detection results.

[0135] During the self-rescue process after tipping over, the main control module controls the four hip joint drive modules 22 to flip the corresponding wheel-leg components 2 outward to a near 180° extreme outward flip position, causing the originally upward-facing wheels 24 to turn towards the ground and re-establish contact. Subsequently, the main control module adjusts the rotation direction and speed of the four wheel drive modules 23 according to the current posture and wheel contact status, so that the robot can still obtain effective rolling driving force in the inverted posture.

[0136] Because the upper and lower surfaces of the robot body 1 are symmetrical or nearly symmetrical, and the wheels 24 have no distinct front and back sides, the robot can maintain normal mobility even after tipping over and touching the ground. In other words, the robot does not need to return to the preset "face-up" posture, but can continue moving in the new orientation. If necessary, the main control module can further restore the robot to a more stable support posture through slight adjustments to the hip joint angle and the combined force of the wheel drive.

[0137] In the implementation using a reinforcement learning policy network, the inverted posture is input into the policy network as part of the state input. The policy network outputs control commands to flip the four sets of wheel leg assemblies 2 outward and ground the wheels, while dynamically adjusting the wheel rotation direction based on the contact state and posture changes. Through the combined effects of posture balance rewards, motion smoothing penalties, and energy consumption penalties, the robot can complete the overturning self-rescue process with lower impact and higher stability.

[0138] like Figure 5 As shown, the robot of this application can also achieve a prone and stowed posture through the hip joint drive module 22. During the stowage process, the main control module controls the four sets of hip joint drive modules 22 to rotate synchronously or stepwise to a position of approximately 90°, so that the four sets of wheel leg assemblies 2 are roughly horizontal and close to the body 1. At this time, the overall height and width of the robot are significantly reduced, which is beneficial for transportation, storage and deployment in confined spaces.

[0139] Once the robot is in its stowed position, the four wheels 24 are positioned compactly on either side or around the body 1, reducing the overall volume by more than 40% compared to its extended driving position. The main control module can also disable some motor drive outputs in the stowed position, retaining only communication, status monitoring, or low-power standby functions to improve battery life.

[0140] When the robot needs to resume its work, the main control module controls the four hip joint drive modules 22 to return from the 90° position to the driving support position, so that the four wheels 24 are grounded again and the robot can resume normal driving.

[0141] To improve the safety and reliability of robot operation, this application may also set up a variety of safety protection strategies.

[0142] For example, the hip joint drive module 22 can be equipped with mechanical limiters or electronic limit programs to prevent the wheel-leg assembly 2 from exceeding a preset angle range and causing structural interference. The wheel drive module 23 can be equipped with overcurrent protection, overtemperature protection, and stall protection. When it is detected that the wheel is spinning idly for a long time, the motor current is abnormally high, or the robot's posture changes drastically, the main control module can reduce the wheel output torque or enter a protection mode.

[0143] After the reinforcement learning policy network outputs control commands, the main control module can perform amplitude limiting, smoothing filtering, and collision detection on the output actions. For example, it can limit the rate of change of the hip joint target angle to avoid the robot becoming unstable due to rapid swinging of the wheel legs; and limit the target rotation speed of the wheels to avoid excessive impact forces during tipping over for self-rescue or climbing.

[0144] Furthermore, before actual deployment, the reinforcement learning policy network can be thoroughly validated in a simulation environment. Different terrain friction coefficients, robot loads, sensor noise, and motor response delays can be simulated using domain randomization to improve the network's generalization ability when migrating from the simulation environment to a real robot. For high-risk conditions, a manual remote control takeover mode or an emergency stop switch can be configured to ensure safety during testing and use.

[0145] In summary, the quadrupedal wheeled robot provided in this application, while reducing the number of lateral joints and multi-degree-of-freedom leg mechanisms, still achieves functions such as traversing ditches, climbing stairs, self-rescue after tipping over, and adjusting its posture for storage. This robot combines advantages such as simple structure, low cost, good obstacle-crossing stability, strong self-rescue capabilities, and convenient transportation and storage, making it suitable for various application scenarios such as inspection, warehousing, field surveying, and transportation in complex terrain.

[0146] The main mechanical structure of the quadrupedal wheeled robot provided in this application embodiment has been described above. The following section, in conjunction with the specific control process, explains the robot's self-rescue from tipping over, stair climbing, ravine crossing, storage, and reinforcement learning control methods.

[0147] Please see Figure 4 , Figure 4 This is a schematic diagram illustrating a stair climbing process provided in an embodiment of this application. Please refer to [link / reference]. Figure 5 , Figure 5 This is a schematic diagram of a crouching and storage posture provided in an embodiment of this application. For ease of explanation, the four wheel-leg assemblies 2 will be referred to as the front left wheel-leg assembly, the front right wheel-leg assembly, the rear left wheel-leg assembly, and the rear right wheel-leg assembly, respectively. However, this naming only corresponds to one reference direction of motion and does not limit the robot's travel direction in actual applications.

[0148] This application provides a tipping self-rescue control method for a quadruped wheeled robot. This method can be executed by the main control module within the robot body 1, or it can be executed collaboratively by an external controller that is communicatively connected to the robot.

[0149] Please see Figure 6 , Figure 6 This is a flowchart illustrating a self-rescue control method for tipping over, provided as an embodiment of this application. The method may include the following steps:

[0150] Step S101: Obtain robot posture information and determine whether the robot is in a tipped-over state.

[0151] The robot's attitude information can be acquired by an inertial measurement unit (IMU). Attitude information may include one or more of the following: roll angle, pitch angle, yaw angle, angular velocity, acceleration, and an estimate of the robot's gravitational direction.

[0152] The main control module can determine whether the robot is overturned based on whether the roll angle or pitch angle exceeds a preset threshold. For example, when the angle between the robot's normal direction and the direction of gravity exceeds a preset threshold, or when the original upper surface of robot body 1 faces the ground for a period of time exceeding a preset time, the robot can be determined to be overturned. Furthermore, the main control module can also combine wheel contact status, motor current changes, and wheel speed feedback to determine whether the wheels still have ground contact conditions after overturning.

[0153] Step S102: When the robot is in a flipped state, control the four sets of hip joint drive modules to drive the corresponding wheel leg components to flip outward, so that the wheels change from a non-grounded state to a grounded state.

[0154] When the robot tip over, some or all of the wheels 24 may be out of ground. At this time, the main control module sends an outward control command to the four hip joint drive modules 22, causing the four wheel leg assemblies 2 to rotate outward around their respective yaw axes.

[0155] In a preferred embodiment, the main control module controls the four sets of wheel-leg assemblies 2 to flip outward to a position close to 180°. Through this extreme outward flipping motion of nearly 180°, the wheels 24, which were originally located above or to the side of the robot, can be guided to the vicinity of the ground and establish supporting contact with it. Since the four sets of wheel-leg assemblies 2 are symmetrically arranged relative to the body 1, the four wheels 24 after flipping outward can form a new support layout, enabling the robot to still have a basis for movement even in a flipped-over posture.

[0156] Step S103: Adjust the rotation direction and speed of each wheel drive module according to the robot's current posture and wheel grounding status.

[0157] After wheel 24 is grounded, the main control module automatically determines the robot's current vertical orientation and direction of travel based on the current attitude information fed back by the inertial measurement unit, the grounding status fed back by the wheel contact sensor, and the wheel speed feedback from each wheel drive module 23. Subsequently, the main control module dynamically switches the forward and reverse directions of the wheel drive modules 23 and adjusts the target rotational speed of each wheel drive module 23.

[0158] For example, when the robot is inverted, the steering of the wheels that were originally used for forward movement may cause the robot to move in the opposite direction. The main control module can adjust the steering mapping relationship of each wheel drive module 23 based on the current posture matrix or gravity direction estimation results, so that the robot can still output driving force in the desired direction even in an inverted posture. As a result, the robot can continue to move without relying on a fixed "forward / reverse" mode switching mechanism or human intervention.

[0159] Step S104: Output driving force through the wheel drive module to enable the robot to recover its drivable state or continue moving from its overturned posture.

[0160] After each wheel 24 establishes ground contact and completes the wheel drive direction adjustment, the main control module controls the four wheel drive modules 23 to output driving force. The robot can restore to a stable and drivable posture through the continuous rolling driving force of the wheel drive modules 23 and the small angle adjustment of the hip joint drive module 22, or move continuously in an inverted posture.

[0161] In one implementation, the robot does not need to return to the factory-defined "face up" position, but can continue to perform tasks in the new orientation after tipping over. Because the upper and lower surfaces of the body 1 are functionally symmetrical and the wheels 24 have no front or back distinction, the robot can achieve continuous driving capability without a fixed up or down orientation.

[0162] As a further implementation, a reinforcement learning strategy network can be introduced into the rollover self-rescue control process. The main control module inputs posture information, wheel contact state, hip joint angle, and wheel motor speed into the reinforcement learning strategy network, which then outputs the corresponding target values ​​for the hip joint yaw angle and the wheel motor speed. Through the reinforcement learning strategy network, the robot can automatically generate smoother and more stable self-rescue actions based on different rollover angles, ground friction coefficients, and contact states.

[0163] This application also provides a method for controlling stair climbing in a quadruped wheeled robot. This method is applicable to robot climbing on steps, curbs, platform edges, or similar terrain with abrupt changes in height.

[0164] Please see Figure 7 , Figure 7 This is a flowchart illustrating a stair climbing control method provided in an embodiment of this application. The method may include the following steps:

[0165] Step S201: Obtain robot posture information, step height information, wheel contact status, and the current status of each wheel and leg assembly.

[0166] Robot posture information can be acquired by an inertial measurement unit. Step height information can be acquired by LiDAR, depth camera, vision sensor, ultrasonic sensor, or other distance sensor. Wheel contact state can be obtained by wheel contact sensor, motor current feedback, wheel speed change, or estimation algorithm. The current state of each wheel leg assembly can include the current yaw angle of each hip joint drive module 22, the current rotational speed of each wheel drive module 23, the current output torque, or the current motor current.

[0167] Step S202: Control at least one front wheel leg assembly near the step side to establish a support point.

[0168] When the robot approaches the step, the main control module can control one or two front wheel leg assemblies near the step to adjust their yaw angle, so that the wheel 24 makes contact with the step's vertical surface, edge, or upper surface. Because the wheel 24 has a continuous trapezoidal groove tread, it can form a strong gripping contact at the step edge, thereby improving the stability of the initial support point.

[0169] As an example, when the robot approaches the step in a forward direction, the front left wheel leg assembly can first establish a support point, followed by the front right wheel leg assembly establishing a second support point; alternatively, depending on the shape of the step and the robot's posture, both front wheel leg assemblies can establish support points simultaneously. This application does not limit the specific order in which the support points are established.

[0170] Step S203: Control the robot's center of gravity to shift towards the front wheel leg assembly with the established support point.

[0171] After the front wheel-leg assembly establishes a support point, the main control module controls the coordinated movements of each hip joint drive module 22 and wheel drive module 23, causing the robot's center of gravity to gradually shift towards the front wheel-leg assembly that has established a support point. During the center of gravity shift, the rear wheel-leg assembly can provide thrust, the front wheel-leg assembly can provide traction through the wheel drive module 23, and the hip joint drive module 22 can adjust the support angle.

[0172] During the center of gravity transfer, the main control module continuously monitors the robot's roll angle, pitch angle, and angular velocity changes. When an excessive attitude deviation is detected, the output torque of the wheel drive module 23 can be reduced, or the yaw angle of the corresponding wheel leg assembly 2 can be adjusted to prevent the robot from tipping over or tilting backward at the edge of the step.

[0173] Step S204: Control the remaining wheel leg components to adjust the sway angle in sequence and fit the step surface.

[0174] Once the front wheel-leg assembly provides stable support and pulls the robot over the step, the main control module controls the remaining wheel-leg assemblies to adjust their yaw angles sequentially, ensuring that their wheels 24 fit against the edge or surface of the step. The rear wheel-leg assembly, while the wheel drive module 23 outputs thrust, can adjust its hip joint yaw angle to reduce impact with the step edge and improve the success rate of overcoming the step.

[0175] Step S205: The four wheel drive modules output driving force, enabling the robot to climb the steps by alternating support with one leg.

[0176] After the four sets of wheel-leg components 2 establish support points in sequence and complete the center of gravity transfer, the four wheel drive modules 23 continuously output driving force, enabling the robot to complete the overall step-climbing action. This climbing process is similar to the alternating support process of a canine's single leg, but this application does not rely on a multi-degree-of-freedom leg structure, but achieves it through the cooperation of a single-degree-of-freedom lateral joint and the wheel drive module 23.

[0177] During the stair climbing process, the yaw angle of each wheel leg assembly 2 and the output torque of each wheel drive module 23 are continuously updated within the control cycle. The control cycle can be in the millisecond range or can be set according to the computing power of the main control module and the motor control frequency. By continuously updating the control quantities, it can adapt to different stair heights, stair surface friction conditions, and external disturbances.

[0178] As a further implementation, the stair climbing control method can generate control commands through a reinforcement learning policy network. The reward constraints of the reinforcement learning policy network can include one or more of the following: wheel contact with the stair surface, maintaining support stability, increasing front wheel contact force, reducing posture deviation, and constraining motion smoothness. In this approach, the stair climbing action does not rely on a fixed preset action sequence, but rather the policy network generates continuous control actions based on real-time status, thereby improving the robot's adaptability in unknown or irregular stair environments.

[0179] This application also provides a method for controlling the crossing of ditches by a quadruped wheeled robot. This method is applicable to robots crossing ditches, cracks, potholes, road breaks, or other gap-type obstacles.

[0180] Please see Figure 8 , Figure 8 This is a flowchart illustrating a ravine crossing control method provided in an embodiment of this application. The method may include the following steps:

[0181] Step S301: Detect the width of the ditch to be crossed.

[0182] The main control module can detect ditches ahead using LiDAR, depth cameras, vision sensors, structured light sensors, or ultrasonic sensors, and estimate the width, edge position, and ground height difference of the ditches to be crossed. The ditches can be the distance between the two edges of the ditches, or the maximum gap distance that the robot needs to cross in its direction of travel.

[0183] Step S302: Control each hip joint drive module to adjust the yaw angle of the corresponding wheel leg assembly according to the width of the groove, so as to increase or adjust the robot wheel track.

[0184] When a large ditch is detected, the main control module controls the four hip joint drive modules 22 to fold outward at a certain angle to increase the wheelbase or adjust the support layout of the four wheels. By adjusting the yaw angle of the wheel-leg assembly 2, the robot's support area can be matched with the ditch width, reducing the risk of the wheels getting stuck in the ditch or the robot losing its center of gravity during crossing.

[0185] As an example, the main control module can determine the target valgus angle based on the relationship between the groove width and the current wheelbase. When the groove width is small, the robot can maintain a normal driving posture and pass through directly; when the groove width is close to the upper limit of the normal wheelbase crossing, the main control module increases the hip joint valgus angle to widen the support spacing.

[0186] Step S303: Control the output driving force of each wheel drive module to make the wheel with continuous groove structure form gripping contact with the ground and / or the edge of the ditch.

[0187] During the crossing of the ditch, the main control module controls the four wheel drive modules 23 to output driving force. The continuous trapezoidal groove structure of the wheel 24 can form a strong grip contact with the edge of the ditch or loose ground, thereby reducing the probability of the wheel spinning or slipping at the edge of the ditch.

[0188] The wheel drive module 23 can use a constant torque output method to propel the robot across ditches; or it can use a method of dynamically adjusting the torque based on wheel speed feedback and motor current feedback to avoid a wheel from continuously spinning due to suspension or slippage.

[0189] Step S304: Drive the robot across the ditch.

[0190] Once the robot's front wheels cross the ditch and establish contact on the other side, the rear wheels continue to output thrust, enabling the entire robot to complete the crossing. During the crossing process, the main control module continuously adjusts the yaw angle of each hip joint and the rotational speed of the wheel drive module 23 based on posture information, wheel speed information, and contact status to ensure a smooth crossing process.

[0191] Through the above-described method, this embodiment utilizes the combination of a large-diameter wheel 24 and a single-degree-of-freedom yaw joint to improve the robot's gripping ability and support stability on the edge of a ditch, reducing the problem of traditional fixed-wheeled robots easily getting stuck or slipping in ditch scenarios. In this process, the robot participates in the crossing process through wheelbase expansion caused by the outward rotation of the wheel-leg components, support domain reconstruction, and adjustment of the robot's center of gravity, enabling the robot's body structure to work synergistically with the environment, thereby improving its ditch-crossing ability.

[0192] This application also provides a method for the storage and control of a quadrupedal wheeled robot. This method is used for posture adjustment of the robot before transportation, storage, or entry into confined spaces.

[0193] Please see Figure 9 , Figure 9 This is a flowchart illustrating a storage control method provided in an embodiment of this application. The method may include the following steps:

[0194] Step S401: Control the four sets of hip joint drive modules to drive the corresponding wheel leg components to swing outward to the preset storage angle.

[0195] When the robot receives the storage command, the main control module controls the four hip joint drive modules 22 to synchronously or stepwise drive the corresponding wheel and leg components 2 to swing outward laterally. The preset storage angle can be 90° or close to 90°. At this angle, the four wheel and leg components 2 change from a vertical or inclined support posture to an approximately horizontal unfolded posture.

[0196] Step S402: Position the four wheel leg assemblies horizontally and close to the machine body.

[0197] When the four sets of wheel and leg components 2 are flipped outward to the preset storage angle, the wheel and leg components 2 are in a horizontal position and close to the body 1, thereby reducing the overall height and width of the robot. The overall volume occupied by the robot in the storage posture can be reduced by more than 40% compared with the unfolded driving posture.

[0198] In the stowed state, the main control module can shut down the high-power output of the wheel drive module 23, retaining only low-power standby, communication, and attitude monitoring functions. When the robot needs to be redeployed, the main control module controls the four hip joint drive modules 22 to return from the stowed angle to the driving support angle, and controls the wheels 24 to re-ground, thereby restoring driving capability.

[0199] This application also provides a reinforcement learning control method for a quadrupedal wheeled robot. This method is used to train and deploy the robot's motion strategy, enabling the robot to autonomously generate control actions in complex terrain and abnormal posture conditions. In this application embodiment, reinforcement learning is used to improve the adaptability and flexibility of the quadrupedal wheeled robot. Through continuous interaction with the environment, the robot can autonomously learn and optimize its motion strategy. This process achieves more efficient decision-making by having the agent receive environmental states, take actions, and optimize based on feedback rewards.

[0200] In motion control, the state consists of the robot's posture, position, velocity, and sensor data, such as the positions of various joints. These states are used as input to the algorithm to guide the robot in making optimal motion choices under different terrain conditions. The motion space mainly consists of an 8-dimensional motion vector, composed of hip joint angles and wheel speeds. Optimizing these parameters allows the robot to flexibly adapt to different ground features.

[0201] In the reward mechanism, the robot earns rewards for successfully overcoming obstacles, maintaining stable movement, consistently tracking linear speed, and quickly rescuing itself, while being penalized for actions that hinder performance enhancement, such as instability, oscillation, and shaking. Through this process of continuous trial and error, the robot gradually discovers the most effective strategy.

[0202] The application of reinforcement learning enables robots to adaptively adjust to changes in their environment. For example, when faced with unknown obstacles, a robot can adjust its posture and speed using learned strategies to achieve optimal crossing or detour capabilities. Furthermore, reinforcement learning provides more diverse and efficient solutions for robot self-rescue strategies, enabling them to quickly regain mobility after a fall and reducing human intervention.

[0203] The validated and evaluated policy network, after being trained using reinforcement learning, is deployed to the robot control unit, forming a real-time closed-loop control mechanism. Through this specially designed reward function, the system can autonomously adjust its actions to adapt to changes in target orientation and deviation without manual parameter tuning. The introduction of the custom reward function not only enhances the learning ability of the policy network but also helps maintain superior performance in complex real-world environments.

[0204] Reinforcement learning training is conducted in a simulation environment. Through multi-scene, multi-terrain, and multi-gait simulations, the policy network is trained to adjust its joint positions under different conditions to achieve continuous and stable tracking of linear velocity. During training, the simulation environment simulates external disturbances, load changes, and terrain inequalities, giving the policy network good generalization ability in real-world environments. After training, the policy network is deployed to the robot control unit to achieve real-time closed-loop control, enabling it to autonomously adjust joint positions and motor movements based on environmental inputs without manual parameter tuning. For details, please refer to [link to details]. Figure 10 , Figure 10 This is a flowchart illustrating a reinforcement learning control method provided in an embodiment of this application. The method may include the following steps:

[0205] Step S501: Collect the state variables of the robot in the simulation environment.

[0206] During the training phase, a dynamic simulation model of the quadrupedal wheeled robot can be constructed, and robot state variables can be collected in the simulation environment. State variables can include one or more of the following: robot posture, position, velocity, sensor data, and joint positions.

[0207] State variables may include IMU attitude information, robot center of mass position, robot linear velocity, angular velocity, yaw angle of the four hip joints, rotational speed of the four wheel motors, wheel contact state, motor current feedback, obstacle height estimation, ditch width estimation, and action commands from the previous control cycle.

[0208] The simulation environment can include multiple scenarios, terrains, and gait simulations. For example, it can be set up to simulate flat ground, slopes, steps, gullies, gravel surfaces, loose soil, slippery surfaces, and randomly raised surfaces. Furthermore, the simulation environment can also simulate one or more of the following: external disturbances, load changes, uneven terrain, sensor noise, and motor response delay, to improve the generalization ability of the trained policy network in real-world environments.

[0209] Step S502: The motion space is formed by the hip joint angle and the wheel rotation speed.

[0210] The motion space can be composed of four hip joint yaw angles and four wheel rotation speeds. As a preferred embodiment, the motion space is an eight-dimensional motion vector, which includes four hip joint angle parameters and four wheel rotation speed parameters.

[0211] Step S503: Train the policy network according to the reward function.

[0212] During training, the agent outputs actions based on its current state, and the simulation environment updates the robot's state based on these actions and calculates the reward value. The reward function may include a target error penalty term, a power consumption compensation term, and an action change penalty term.

[0213] The reward function can be expressed as:

[0214]

[0215] in, It is the reward value given state s and action a;

[0216] The penalty term representing the target error may include linear velocity tracking error, angular velocity tracking error, or attitude error;

[0217] Compensation for power consumption can include energy consumption penalty, joint torque penalty, or motor current penalty;

[0218] This indicates a penalty for changes in motion, which can be used to constrain excessive changes in joint position or wheel speed within adjacent control cycles.

[0219] These are the weighting coefficients for each reward.

[0220] In different tasks, the reward function can also add task-related rewards. For example, in the ravine crossing task, rewards can be added for successfully crossing the ravine, stable wheel contact, and avoiding slippage; in the stair climbing task, rewards can be added for wheel contact with the stair surface, maintaining support stability, increasing front wheel contact force, reducing posture deviation, and smooth movement; in the fall-and-recovery task, rewards can be added for quickly restoring drivability, re-grounding the wheels, and low-impact posture recovery.

[0221] Step S504: Deploy the trained policy network to the robot's main control module.

[0222] Once the training performance of the policy network in the simulation environment meets the preset requirements, the policy network parameters can be deployed to the main control module within the machine body 1. During actual operation, the main control module receives sensor input and invokes the policy network for real-time inference.

[0223] To enhance safety, after the policy network outputs control commands, the main control module can perform amplitude limiting, filtering, and safety verification on the output actions. For example, it can limit the target hip joint angle to no more than 0° to 180°, limit the target rotational speed of the wheel drive module 23 to no more than the safe rotational speed, and limit the rate of change of hip joint angle and wheel rotational speed within adjacent control cycles.

[0224] Step S505: The strategy network outputs the target values ​​of the four hip joint yaw angles and the target rotation speeds of the four wheel drive modules during the real-time control cycle to control the robot's movement.

[0225] In actual operation, the strategy network outputs eight-dimensional control commands in each real-time control cycle, corresponding to the target values ​​of the four hip joint yaw angles and the target rotation speeds of the four wheel drive modules. The main control module sends the above control commands to the hip joint drive module 22 and the wheel drive module 23, thereby realizing real-time closed-loop control of the robot during flat ground driving, ditch crossing, stair climbing, self-rescue after tipping over, and deployment and retrieval.

[0226] By employing reinforcement learning control methods, robots can avoid relying entirely on manually designed fixed action sequences. Instead, they can automatically adjust their motion strategies based on the current environmental state and their own posture, improving their adaptability to unknown terrain, complex obstacles, and abnormal postures. For example, during stair climbing, the policy network does not depend on a preset action sequence but generates continuous optimal actions under the constraints of the environmental state and the reward function. During the training phase, by defining indicators such as "wheel contact with the stair surface," "maintaining triangular support stability," "increasing front wheel contact force," "reducing posture deviation," and "constraining action smoothness" in the reward function, the strategy autonomously forms climbing behavior patterns during large-scale simulation iterations. For instance, under reward constraints, the network automatically learns to first establish a stable support point with the front left wheel leg closest to the stair. Subsequently, under the combined effect of posture error penalties and energy consumption penalties, it forms a strategy to gradually transfer the center of gravity to the front wheel leg, thereby driving the front right wheel leg to sway to a new support position. The lifting, swaying, and contacting actions of the rear wheel leg are all automatically generated by the network in closed-loop control based on IMU posture changes, contact state changes, and action smoothness constraints. Throughout the climbing process, the yaw angle of each wheel leg and the output torque of the wheel body are continuously updated within millisecond cycles to adapt to different step heights, uneven surfaces and external disturbances, so as to achieve continuous and stable climbing movements.

[0227] During the self-rescue process after tipping over, the reinforcement learning policy network is also responsible for generating the action sequence. When the attitude sensor detects that the robot is in an inverted state, this abnormal posture is directly used as part of the network input. Within the control cycle, the policy network outputs a command to make the four wheel leg units flip outward to approximately 180°, so that the grooved wheels can establish a stable contact surface with the ground. Subsequently, the policy network adjusts the rotational direction and amplitude of the four wheel motors based on real-time attitude data and contact information, enabling the robot to obtain continuous rolling driving force in the inverted state without relying on preset forward and reverse mode switching. During the process of restoring to a drivable posture, the attitude balance term, motion smoothing term, and energy consumption penalty work together to enable the network to generate a rollover action with lower acceleration and higher stability. Finally, through small adjustments to the hip joint yaw angle and continuous output of wheel driving force, the robot completes a reliable self-rescue process under structural conditions without a fixed up-down direction.

[0228] Through the above mechanism, the robot's various actions during stair climbing and self-rescue after falling over are all derived from the autonomous decision-making of the reinforcement learning policy network in the real-time control loop, rather than relying on a fixed process designed by humans. This significantly improves the robot's adaptability and continuous operation performance in unknown terrain and abnormal posture conditions.

[0229] Based on the above embodiments, the robot can also be equipped with a variety of safety protection mechanisms.

[0230] The hip joint drive module 22 can be configured with mechanical limit structures and electronic limit programs. When the hip joint angle is detected to be close to its limit range, the main control module reduces the hip joint drive speed or stops further eversion to prevent structural collision or motor overload.

[0231] The wheel drive module 23 can be configured with overcurrent protection, overtemperature protection, stall protection, and idling detection mechanisms. When it is detected that the wheel is idling for a long time, the motor current is abnormally high, or the motor temperature exceeds the threshold, the main control module can reduce the wheel output torque, adjust the hip joint angle, or enter protection mode.

[0232] The robot supports both automatic control and manual remote control takeover modes. In automatic control mode, the main control module executes motion control based on sensor information and policy network outputs. In manual remote control takeover mode, operators can send speed, direction, retraction, or self-rescue commands to the robot via a remote control terminal. The manual remote control takeover mode can be used in the testing phase or in high-risk operational scenarios.

[0233] The body 1 and the wheel-leg assembly 2 can adopt a modular design. When a wheel-leg assembly 2 fails, the corresponding hip joint connector 21, hip joint drive module 22, wheel drive module 23, or wheel 24 can be disassembled for repair or replacement. Since the four wheel-leg assemblies 2 have the same or basically the same structure, it is beneficial to reduce spare parts costs and maintenance difficulty.

[0234] In summary, the quadrupedal wheel-legged robot provided in this application embodiment comprises a simplified integrated wheel-leg structure consisting of a body 1, a hip joint connector 21, a single-degree-of-freedom hip joint drive module 22, a wheel drive module 23, and a wheel 24. This structure eliminates the lateral yaw joint and complex wheel-leg switching mechanism, achieving terrain adaptation solely through the single-degree-of-freedom yaw joint and wheel rotation drive. The large-diameter, continuously grooved wheels (approximately 30cm) enhance grip on ditch and step edges; the hip joint's 0° to 180° eversion capability and the wheel drive module's 360° continuous rotation capability enable autonomous grounding, autonomous steering adjustment, and continuous movement after tipping over; the near-90° folding posture of the hip joint reduces the space occupied during robot transportation and storage; and the reinforcement learning policy network enables the robot to autonomously generate stable and continuous control actions under complex terrain and abnormal postures.

[0235] As indicated in this application and claims, unless the context clearly indicates otherwise, the words "a," "an," "a," and / or "the" are not specifically singular and may include the plural. Generally, the terms "comprising" and "including" only indicate the inclusion of expressly identified steps and elements, which do not constitute an exclusive list, and the method or apparatus may also include other steps or elements. An element defined by the phrase "comprising an..." does not exclude the presence of other identical elements in the process, method, product, or apparatus that includes the element.

[0236] In the description of this application, unless otherwise expressly defined, terms such as "setup," "installation," and "connection" should be interpreted broadly, and those skilled in the art can reasonably determine the specific meaning of the above terms in this application in conjunction with the specific content of the technical solution.

[0237] It should be noted that the various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. The same or similar parts between the various embodiments can be referred to each other.

[0238] This document uses specific examples to illustrate the principles and implementation methods of this application. The descriptions of the embodiments above are only for the purpose of helping to understand the core ideas of this application. It should be noted that those skilled in the art can make several improvements and modifications to this application without departing from the principles of this application, and these improvements and modifications also fall within the protection scope of the claims of this application.

Claims

1. A quadrupedal wheeled robot, characterized in that, It includes a body (1) and four sets of wheel leg assemblies (2) respectively disposed around the body (1); Each wheel-leg assembly (2) includes a hip joint connector (21), a hip joint drive module (22), a wheel drive module (23), and a wheel (24). One end of the hip joint connector (21) is connected to the body (1), and the other end is connected to the hip joint drive module (22); The hip joint drive module (22) is a single-degree-of-freedom lateral joint, used to drive the corresponding wheel leg assembly to fold outward relative to the body (1) in the lateral direction; The wheel drive module (23) is connected to the wheel (24) for driving the wheel (24) to rotate continuously; The four sets of wheel-leg assemblies (2) are arranged symmetrically relative to the body (1). The hip joint drive module (22) has a range of motion that allows the wheel-leg assemblies (2) to flip outward and re-ground after tipping over, so that the robot can ground and travel through the wheels (24) in both normal and tipped-over postures.

2. The quadrupedal wheeled robot as described in claim 1, characterized in that, The lateral valgus angle α of the hip joint drive module (22) ranges from 0° to 180°.

3. The quadrupedal wheeled robot as described in claim 1, characterized in that, The wheel body (24) is a tracked grooved wheel, and the tread of the wheel body (24) is provided with a continuous groove structure. The continuous groove structure is a trapezoidal groove structure continuously distributed along the circumference of the wheel body (24).

4. The quadrupedal wheeled robot as described in claim 1, characterized in that, The wheel (24) is a symmetrical wheel without a front or back, so that the wheel (24) still has the same grounding characteristics after the robot falls over.

5. The quadrupedal wheeled robot as described in claim 1, characterized in that, The four sets of wheel leg assemblies (2) are the front left wheel leg assembly, the front right wheel leg assembly, the rear left wheel leg assembly and the rear right wheel leg assembly, respectively. The lateral swing direction of the hip joint of the two sets of wheel leg assemblies on the front side is opposite to the lateral swing direction of the hip joint of the two sets of wheel leg assemblies on the rear side.

6. The quadrupedal wheeled robot as described in claim 1, characterized in that, The upper and lower surfaces of the body (1) have the same or symmetrical geometric and functional structures so that the robot can continue to operate without changing the working mode of the main control system and power supply system after it is overturned.

7. The quadrupedal wheeled robot as described in claim 1, characterized in that, The main control module, power management module and battery pack are installed inside the body (1); The main control module is configured to determine the robot's current up and down orientation based on feedback from the attitude sensor, and adjust the rotation direction of the wheel drive module (23) according to the up and down orientation; And / or, the main control module is configured to control the hip joint drive module (22) to perform an outward tumbling action when the robot flips over, so that the wheel (24) re-contacts the ground, and control the wheel drive module (23) to output driving force to achieve autonomous movement after flipping over; And / or, the main control module is configured to control the four sets of hip joint drive modules (22) to be uniformly turned outward to 90°, so that the four sets of wheel leg assemblies are in a horizontal state close to the body (1) to achieve a crouching and / or storage posture; And / or, the main control module is configured to control each hip joint drive module (22) to adjust the yaw angle according to the detected ditch width when crossing the ditch, so as to increase the wheel track, and control the wheel drive module (23) to output torque to push the robot across the ditch; And / or, the main control module is configured to control the four sets of wheel leg assemblies to perform single-leg alternating support when climbing the steps (3), so that the robot can gradually complete the climbing of the steps (3) by transferring the center of gravity between the front and rear wheels; And / or, the main control module is equipped with a policy network trained by reinforcement learning, which is used to output control commands to control the four hip joint drive modules (22) and the four wheel drive modules (23) according to the robot state variables.

8. The quadrupedal wheeled robot as described in claim 7, characterized in that, When a policy network trained by reinforcement learning is deployed in the main control module, the robot state variables include one or more of the following: IMU attitude information, yaw angle of each hip joint, rotational speed of the wheel (24) motor, center of mass height, obstacle height estimation, wheel (24) contact state, motor current feedback, and action command of the previous control cycle.

9. The quadrupedal wheeled robot as described in claim 7, characterized in that, The strategy network outputs eight-dimensional control commands, which correspond to the target values ​​of the hip joint yaw angle of the four wheel-leg components and the target rotation speed of the four wheel drive modules (23).

10. The quadrupedal wheeled robot as described in claim 7, characterized in that, When a policy network trained by reinforcement learning is deployed in the main control module, the main control module performs amplitude limiting processing on the control commands output by the policy network and sends them to the hip joint drive module (22) and the wheel drive module (23).

11. A method for overturning self-rescue control of a quadrupedal wheeled robot, characterized in that, The method, applied to the quadrupedal wheeled robot according to any one of claims 1 to 10, comprises: Acquire robot posture information and determine whether the robot is in a tipped-over state; When the robot is in a tipped-over state, the four sets of hip joint drive modules are controlled to drive the corresponding wheel and leg components to flip outward, so that the wheels change from an ungrounded state to a grounded state. Adjust the rotation direction and speed of each wheel drive module according to the robot's current posture and wheel grounding status; The robot outputs driving force through the wheel drive module, enabling it to recover its drivable state or continue moving after tipping over.

12. The overturning self-rescue control method as described in claim 11, characterized in that, The control of the four sets of hip joint drive modules drives the corresponding wheel and leg assemblies to rotate outward, including controlling the wheel and leg assemblies to rotate outward to a position close to 180°.

13. The overturning self-rescue control method as described in claim 11, characterized in that, The adjustment of the rotation direction and speed of each wheel drive module includes: automatically determining the robot's current up-down orientation based on feedback from the attitude sensor, and dynamically switching the forward and reverse directions of the wheel drive modules.

14. The overturning self-rescue control method as described in claim 11, characterized in that, The method further includes: inputting attitude information, wheel contact state, hip joint angle, and wheel motor speed into a reinforcement learning policy network, and having the reinforcement learning policy network output the corresponding target value of hip joint yaw angle and target speed of wheel motor.

15. A method for controlling stair climbing in a quadruped wheeled robot, characterized in that, Applied to the quadrupedal wheeled robot according to any one of claims 1 to 10, the method comprises: Acquire robot posture information, step height information, wheel contact status, and the current status of each wheel and leg assembly; Control at least one front wheel leg assembly near the step to establish a support point; Control the robot's center of gravity to shift towards the front wheel-leg assembly, which has already established a support point; Control the remaining wheel leg components to adjust the sway angle sequentially and conform to the surface of the step; The robot climbs stairs by using a single-leg alternating support method, driven by four wheel drive modules that output driving force.

16. The stair climbing control method as described in claim 15, characterized in that, The yaw angle of each wheel leg assembly and the output torque of the wheel drive module are continuously updated within the control cycle.

17. The stair climbing control method as described in claim 15, characterized in that, The stair climbing control method generates control commands through a reinforcement learning policy network. The reward constraints of the reinforcement learning policy network include one or more of the following: wheel contact with the stair surface, maintaining support stability, increasing front wheel contact force, reducing posture deviation, and constraining motion smoothness.

18. A method for controlling ravine crossing in a quadruped wheeled robot, characterized in that, Applied to the quadrupedal wheeled robot according to any one of claims 1 to 10, the method comprises: Detect the width of the ditch to be crossed; The width of the groove is used to control the yaw angle of each hip joint drive module to adjust the corresponding wheel leg assembly, thereby increasing or adjusting the robot's wheelbase. Control the output driving force of each wheel drive module to make the wheel with a continuous groove structure form a gripping contact with the ground and / or the edge of the ditch; Drive the robot across the ravine.

19. A method for storing and controlling a quadrupedal wheeled robot, characterized in that, Applied to the quadrupedal wheeled robot according to any one of claims 1 to 10, the method comprises: The four hip joint drive modules are controlled to drive the corresponding wheel and leg components to swing outward to the preset storage angle. Position the four wheel and leg assemblies horizontally and close to the body; The preset storage angle is 90° or close to 90° to reduce the overall space occupied by the robot.

20. A reinforcement learning control method for a quadrupedal wheeled robot, characterized in that, Applied to the quadrupedal wheeled robot according to any one of claims 1 to 10, the method comprises: Collect the robot's state variables in a simulation environment, including one or more of attitude, position, velocity, sensor data, and joint position; The motion space is defined by the hip joint angle and the wheel rotation speed; The policy network is trained based on the reward function; Deploy the trained policy network to the robot's main control module; The strategy network outputs target values ​​for the yaw angles of four hip joints and target rotational speeds of four wheel drive modules during the real-time control cycle to control the robot's movement.

21. The reinforcement learning control method as described in claim 20, characterized in that, The motion space is an eight-dimensional motion vector, which includes four hip joint angle parameters and four wheel rotation speed parameters.

22. The reinforcement learning control method as described in claim 20, characterized in that, The reward function includes a target error penalty term, a power consumption compensation term, and an action change penalty term.

23. The reinforcement learning control method as described in claim 22, characterized in that, The reward function is: in, It is the reward value given state s and action a; It is a penalty term for target error, including the penalty for tracking speed error; It is a power consumption compensation, including energy consumption penalty and joint torque penalty; It is a penalty for excessive movement, which is related to the difference between the current position of each joint and the position of the previous time step. These are the weighting coefficients for each reward.

24. The reinforcement learning control method as described in claim 20, characterized in that, The simulation environment includes multi-scene, multi-terrain, and multi-gait simulations, and simulates one or more of external disturbances, load changes, and uneven terrain.

Images