A visual perception-based control method for obstacle-crossing motion of a quadruped robot

Abstract

This invention discloses a visual perception-based motion control method for quadruped robots to cross negative obstacles, comprising the following steps: S1, acquiring visual information including RGB images and depth maps, as well as the quadruped robot's body state information; S2, extracting geometric and semantic information of the negative obstacle from the visual information, fusing the body state information, and establishing a negative obstacle environment model for motion control. This invention uses visual perception to complete negative obstacle identification and crossing preparation. After the forefoot reaches the edge, the crossing is triggered only after comprehensive confirmation based on multi-source information such as foot position, contact state, and joint feedback. This two-stage crossing triggering mechanism of visual prediction and foot confirmation effectively avoids problems such as premature or delayed crossing caused by visual ranging errors, perception delays, positioning errors, and speed changes, fundamentally improving the safety and reliability of negative obstacle crossing.

Description

Technical Field

[0001] This invention belongs to the fields of robot motion control, embodied intelligence and artificial intelligence, specifically referring to a method for controlling the motion of a quadruped robot to overcome obstacles based on visual perception. Background Technology

[0002] Quadruped robots, due to their excellent terrain adaptability, can be widely used in patrol, transportation, rescue, and exploration tasks. In complex terrain environments, quadruped robots not only need to face positive obstacles such as steps and protrusions, but also negative obstacles such as ditches, potholes, and collapsed road surfaces. Overcoming negative obstacles places extremely high demands on the robot's environmental perception capabilities, motion control precision, and decision-making timing.

[0003] Existing quadruped robots face the following main technical challenges when performing negative obstacle crossing tasks:

[0004] First, relying on visual distance to directly trigger crossing is insufficient in terms of anti-interference capability. In existing technologies, quadruped robots mostly rely on visual distance detection to directly trigger the crossing action when crossing negative obstacles. However, visual ranging itself has inherent errors, image acquisition, transmission, and processing involve perception delays, robot positioning has accumulated errors, and robot speed changes in real time—all these factors combined make it difficult to accurately control the timing of crossing based on visual distance. Visual ranging errors may cause crossing to be triggered too early or too late: if triggered too early, the robot's forelegs have not yet reached the edge of the ditch before starting to cross, potentially leading to a fall into the ditch; if triggered too late, the robot's forelegs have already crossed the edge of the ditch before initiating the crossing, potentially leading to a fall.

[0005] Second, the lack of an adaptive approach mechanism makes it difficult to achieve smooth gait transitions. Existing methods typically use a fixed distance threshold to trigger traversal control, without considering factors such as the geometric differences of negative obstacles, the robot's current speed, and gait cycle. When the robot approaches negative obstacles of different sizes at different speeds, the fixed threshold cannot adapt to the changes, resulting in abrupt approach processes and jarring gait transitions, affecting the continuity and stability of traversal.

[0006] Third, the crossing trigger mechanism is singular and lacks multi-source information verification. Existing technologies rely solely on visual perception results to trigger crossing, failing to fully utilize proprioceptive information such as foot position, contact state, and joint feedback for comprehensive judgment. When visual perception is compromised due to occlusion, changes in lighting, or sensor noise, the single trigger mechanism lacks effective verification and fault tolerance measures.

[0007] Fourth, the control strategy lacks robustness to perception errors. In real-world environments, visual sensors are subject to noise, perception information is delayed, and there are uncertainties in the ground friction coefficient and robot dynamic parameters. Existing control strategies are typically trained or designed under ideal conditions, failing to adequately consider these uncertainties, leading to a significant performance degradation when migrating from simulation to real-world environments.

[0008] In summary, existing technologies lack a negative obstacle crossing motion control method that can integrate visual perception and proprioception, possess adaptive approach capabilities, employ a multi-source information verification triggering mechanism, and be robust to perception errors. As a result, it is difficult to achieve stable and safe autonomous negative obstacle crossing in complex real-world environments. Summary of the Invention

[0009] The purpose of this invention is to overcome the technical shortcomings of current quadruped robots, which lack adaptive approach capabilities and are difficult to achieve stable and safe autonomous obstacle crossing in complex real-world environments, and to provide a motion control method for quadruped robots to cross obstacles based on visual perception.

[0010] To achieve the above objectives, the present invention employs the following technical solution: a method for controlling the movement of a quadruped robot traversing negative obstacles based on visual perception, comprising the following steps:

[0011] S1. Collect visual information including RGB images and depth maps, as well as the body state information of the quadruped robot;

[0012] S2. Extract the geometric and semantic information of negative obstacles from visual information and integrate the ontological state information to establish a negative obstacle environment model for motion control.

[0013] S3. Based on the negative obstacle spatial attributes in the negative obstacle environment model and the real-time motion state of the quadruped robot, construct the area to be crossed. After the quadruped robot enters the area to be crossed, adaptively adjust the movement speed, step frequency, body posture and center of gravity position to achieve a smooth transition from the normal walking state to the negative obstacle crossing state.

[0014] S4. After the quadruped robot enters the area to be crossed, it switches to the negative obstacle crossing control mode and continuously uses visual information to correct the approach process.

[0015] S5. When the quadruped robot’s forelegs reach the edge of the negative obstacle, the edge is confirmed based on the foot position information, foot contact state information and joint feedback information. After confirming that the edge has been reached, the crossing strategy is triggered.

[0016] S6. Perform the negative obstacle crossing action, and resume normal walking after completing the crossing.

[0017] The geometric information in step S2 includes at least one of the width, depth, edge position, and edge direction of the negative obstacle; the semantic information includes the category information of the negative obstacle, and the category includes at least one of the ditch, pothole, step, and collapsed area.

[0018] The distance threshold of the area to be traversed in step S3 is determined jointly by the geometry of the negative obstacle, the quadruped robot's movement speed, gait period, and current motion state. The formula for calculating the distance threshold of the area to be traversed is: ,in, , Negative obstacle depth This represents the current speed of the quadruped robot. This represents the current gait cycle, and the state is the current motion state of the quadruped robot.

[0019] The negative obstacle crossing control mode mentioned in step S4 is a dedicated gait mode for negative obstacle crossing. The foot trajectory of the dedicated gait mode has a higher leg lift height than that of the normal walking gait, and the body posture is more forward-leaning than that of the normal walking gait.

[0020] The edge confirmation mentioned in step S5 specifically refers to: when the distance between the position coordinates of the forefoot and the negative obstacle edge is less than a preset threshold, the contact state of the forefoot meets the preset contact conditions, and the joint feedback information of the forefoot is within a preset range, it is confirmed that the forefoot has reached the negative obstacle edge and the crossing strategy is triggered; if any of the above conditions are not met, the crossing strategy is not triggered and the approach process continues to be adjusted.

[0021] The negative obstacle crossing action described in step S6 has a reinforcement learning strategy. This reinforcement learning strategy refers to introducing at least one random perturbation among visual noise, perception delay, control delay, friction coefficient change, robot mass change, and center of mass position change during the training process to perform domain randomization training.

[0022] The "adaptive adjustment of movement speed, step frequency, body posture, and center of gravity position" mentioned in step S3 triggers adaptive proximity control. The triggering condition for this adaptive proximity control is that the horizontal distance d between the quadruped robot and the edge of the negative obstacle satisfies d ≤ The This represents the distance threshold of the area to be traversed.

[0023] Compared with the prior art, the beneficial effects of the present invention are:

[0024] (1) This invention uses visual perception to complete the identification and preparation for crossing negative obstacles. After the forefoot reaches the edge, the crossing is triggered only after comprehensive confirmation based on multi-source information such as foot position, contact state and joint feedback. This two-stage crossing triggering mechanism of visual prediction and foot confirmation can effectively avoid the problems of premature or delayed crossing caused by visual ranging error, perception delay, positioning error and speed change, and fundamentally improve the safety and reliability of crossing negative obstacles.

[0025] (2) This invention comprehensively extracts the geometric and semantic attributes of negative obstacles and establishes a unified environment model for motion control, providing a more complete environmental description for crossing decisions, gait switching and timing control, which can significantly improve the accuracy of negative obstacle recognition and motion planning in complex scenarios.

[0026] (3) This invention adaptively calculates the distance threshold of the area to be crossed by taking into account factors such as the geometric dimensions of the negative obstacle, the robot's movement speed, gait cycle, and movement state, rather than using a fixed distance threshold trigger control. When the robot enters the area to be crossed, it dynamically adjusts its movement speed, step frequency, body posture, and center of gravity position to achieve a gradual and smooth gait transition, which significantly improves the continuity and stability of the crossing process.

[0027] (4) In the reinforcement learning training process of the negative obstacle crossing action, this invention introduces multi-dimensional random perturbations such as visual noise, perception delay, control delay, friction coefficient, and robot mass, so that the training strategy can cover different perception errors and dynamic change conditions. This method effectively narrows the gap between the simulation environment and the real environment and improves the generalization ability and robustness of the crossing strategy in real complex environments.

[0028] (5) The present invention divides the negative obstacle crossing process into multiple control states such as normal walking, negative obstacle recognition, adaptive approach, gait switching, foot edge confirmation, crossing execution and recovery walking, and completes state adaptive switching based on the robot's motion state and environmental perception information, which can improve the stability and interpretability of motion control in complex environments and realize phased fine control of the entire crossing process. Attached Figure Description

[0029] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0030] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0031] Example

[0032] like Figure 1 As shown in the figure, the visual perception-based quadruped robot obstacle crossing motion control method described in this embodiment includes steps S1 to S6. The following is a detailed description of each step:

[0033] S1. Collect visual information including RGB images and depth maps, as well as the body state information of the quadruped robot.

[0034] This step involves the quadruped robot's visual perception and information acquisition. The quadruped robot of this invention has the following built-in functional modules, specifically:

[0035] The visual perception module is used to acquire visual information, including RGB images and depth maps. In this embodiment, the visual perception module includes at least: an RGB camera for acquiring color images and identifying appearance information such as the color and texture of obstacles; a depth camera for acquiring depth maps or point cloud data and measuring the distance, shape, and spatial position of obstacles; and a lidar for acquiring high-precision 3D environmental point cloud data through laser scanning to construct a high-precision map.

[0036] The body state perception module is used to collect the body state information of the quadruped robot and to perceive its own posture and motion. This body state information includes at least the quadruped robot's joint angles, angular velocities, foot contact forces, and body posture. In this embodiment, the body state perception module is an inertial measurement unit (IMU).

[0037] The environment modeling module is used to extract the geometric and semantic information of negative obstacles from the visual information and fuse it with the ontological state information to establish a negative obstacle environment model. The negative obstacle refers to a terrain structure in the quadruped robot's movement environment that has a significant downward depression, break, or gap relative to the robot's current standing plane or the surrounding ground height. The characteristic of a negative obstacle is that its edges (ditch edges) and internal depressions together constitute a constraint on the quadruped robot's movement, requiring the quadruped robot to safely pass through by crossing, detouring, or avoiding it. In this embodiment, the negative obstacle is at least one of ditches, potholes, trenches, and collapsed road surfaces.

[0038] An adaptive proximity control module is used to construct a region to be crossed based on the negative obstacle spatial attributes in the negative obstacle environment model and the real-time motion state of the quadruped robot, and adaptively adjust motion parameters after entering the region to be crossed.

[0039] The gait switching module is used to switch the quadruped robot to a negative obstacle crossing control mode after entering the area to be crossed.

[0040] The edge confirmation module is used to confirm the negative obstacle edge based on foot position information, foot contact status information, and joint feedback information.

[0041] The obstacle crossing control module is used to perform a negative obstacle crossing action after confirming that the edge has been reached.

[0042] The RGB image is a standard color image, which is composed of three color channels: red (R), green (G), and blue (B). It records the appearance information of the scene, such as color, texture, and brightness.

[0043] The depth map is a distance image where each pixel value represents the physical distance of that point relative to the camera. It records the geometric information (spatial location, shape) of the scene but does not include color information. It is usually a grayscale image, where brighter (whiter) colors indicate closer distances and darker (blacker) colors indicate farther distances.

[0044] S2. Extract the geometric and semantic information of negative obstacles from visual information and integrate the ontological state information to establish a negative obstacle environment model for motion control.

[0045] This step involves modeling the negative obstacle environment. The quadruped robot extracts the geometric and semantic information of the negative obstacles from the acquired RGB images and depth maps. The geometric information includes at least one of the following: width, depth, edge position, and edge direction of the negative obstacle. For example, the width (e.g., 0.8 meters), depth (e.g., 0.5 meters), left edge position, right edge position, and edge direction of a ditch.

[0046] The semantic information includes category information for negative obstacles, which includes at least one of ditches, potholes, steps, and collapsed areas.

[0047] The quadruped robot integrates the aforementioned geometric and semantic information with its current motion speed, body posture, and other ontological state information to establish a negative obstacle environment model. This negative obstacle environment model refers to storing the spatial and categorical attributes of the negative obstacles, as well as the relative motion relationship between the robot and the obstacles, in structured data form, providing a unified environmental description for subsequent control.

[0048] Spatial attributes refer to the geometric and positional information of the negative obstacle in physical space, used to quantitatively describe the spatial form of the obstacle and its relative relationship with the robot. Category attributes refer to the type label and semantic meaning of the negative obstacle, used to describe what the obstacle is and its impact on the robot's motion, i.e., the category information mentioned above.

[0049] S3. Based on the negative obstacle spatial attributes in the negative obstacle environment model and the real-time motion state of the quadruped robot, construct the area to be crossed. After the quadruped robot enters the area to be crossed, adaptively adjust the movement speed, step frequency, body posture and center of gravity position to achieve a smooth transition from the normal walking state to the negative obstacle crossing state.

[0050] The constructed traversable region refers to a dynamically calculated virtual space trigger zone. When the quadruped robot enters this region, the system triggers a state switch from normal walking to traversal preparation and adaptively adjusts the motion parameters. The distance threshold of the traversable region is determined by the geometry of the negative obstacle, the quadruped robot's speed, gait cycle, and current motion state. In other words, multiple state variables at the current moment are input into a function f to calculate the boundary distance of the traversable region in real time.

[0051] The function f is a multivariate decision function constructed based on robot kinematic constraints, or a nonlinear mapping relationship implicitly defined by a reinforcement learning policy network. Its input variables include at least the negative obstacle width, robot movement speed, and gait cycle, and the output variable is the boundary distance threshold of the area to be crossed.

[0052] The formula for this distance threshold is expressed as: ,in, , Negative obstacle depth This represents the current speed of the quadruped robot. This represents the current gait cycle, and the state is the current motion state of the quadruped robot.

[0053] Once the quadruped robot enters the area to be traversed, it adaptively reduces its speed (e.g., from 0.5 m / s to 0.3 m / s), adjusts its stride frequency, and lowers its center of gravity, achieving a smooth transition from normal walking to obstacle-crossing. This gradual adjustment avoids abruptness and instability during gait transitions.

[0054] The aforementioned "adaptive adjustment of movement speed, step frequency, body posture, and center of gravity position after the quadruped robot enters the area to be crossed" triggers adaptive proximity control. The trigger condition for this adaptive proximity control is: the horizontal distance d between the quadruped robot and the edge of the negative obstacle satisfies d ≤ The horizontal distance d is a real-time state variable that is jointly determined and directly measured and calculated by the body state perception module and the visual perception module.

[0055] When d> When d ≤ , it is determined that the quadruped robot has not yet entered the area to be traversed, and the current state is maintained; when d ≤ When this occurs, adaptive proximity control is triggered.

[0056] The method for obtaining the horizontal distance d is as follows:

[0057] First, the quadruped robot uses IMU and SLAM algorithms to determine its precise position on the map in real time. This step determines the quadruped robot's position. The SLAM algorithm is a simultaneous localization and mapping algorithm. The origin of the quadruped robot's coordinates in the world coordinate system is the world origin O(0,0,0) when the quadruped robot starts or when the map is built. The coordinates are ,in, Let x be the position of the quadruped robot's center of mass in the x-direction of the world coordinate system. Let be the position of the quadruped robot's center of mass in the y-direction of the world coordinate system. Let be the position of the quadruped robot's center of mass in the z-direction of the world coordinate system. At this moment, the quadruped robot's position in the world coordinate system is: .

[0058] Secondly, the visual perception module processes environmental data to obtain the precise location of the negative obstacle edge on the map. The location of the negative obstacle edge is determined by the visual perception module through analysis of the depth map, and its coordinate origin is typically the camera itself. The quadruped robot then invokes the coordinate transformation matrix (rotation matrix R and translation vector). The coordinates of the negative obstacle edges, originally in the "camera coordinate system," were forcibly transformed to the same world coordinate system as the quadruped robot. The transformed coordinates of the negative obstacle edges are: At this point, the quadruped robot and the edge of the negative obstacle are on the same map, sharing a common origin. The coordinates are ,in, This represents the x-direction position of the negative obstacle leading edge in the world coordinate system. This represents the position of the negative obstacle leading edge in the y-direction within the world coordinate system. The position of the negative obstacle leading edge in the z-direction in the world coordinate system.

[0059] Finally, in a unified world coordinate system, the Euclidean distance from the current position of the quadruped robot to the edge of the negative obstacle is calculated, and only its component on the horizontal plane is taken to obtain the horizontal distance d.

[0060] Since negative obstacles are terrain features in three-dimensional space with a height z-axis, determining whether the quadruped robot has reached the edge of a negative obstacle is independent of its ground clearance z-axis; therefore, only its horizontal component needs to be considered. This is relevant when calculating the quadruped robot's position. With edge position When the vector difference is calculated, the program forcibly sets the z-axis coordinates of both to zero (i.e., discards the vertical height data). At this point, the three-dimensional coordinates are compressed into two-dimensional planar coordinates, i.e., the horizontal position of the quadruped robot: Horizontal position of negative obstacle edge: The system calculates the horizontal difference vector Δd from the quadruped robot to the edge of the negative obstacle based on the above positions, i.e. Finally, the modulus of the obtained horizontal difference vector Δd is calculated, i.e., the Euclidean distance is calculated, to obtain the final horizontal distance d, which is expressed as: .

[0061] S4. After the quadruped robot enters the area to be crossed, it switches to the negative obstacle crossing control mode and continuously uses visual information to correct the approach process.

[0062] After the quadruped robot enters the area to be traversed, the control system switches the quadruped robot from its normal walking gait to a negative obstacle traversing control mode. This negative obstacle traversing control mode is a dedicated gait mode for negative obstacle traversing. This dedicated gait mode means that, compared to the normal walking gait, its foot trajectory has a higher leg lift (e.g., increasing from 0.05m to 0.15m) and its body posture is more forward-leaning, in order to adapt to the requirements of traversing negative obstacles.

[0063] During the approach, the visual perception module continuously collects new images and depth data, constantly updates the location information of the negative obstacle, corrects the approach trajectory, and ensures that the quadruped robot approaches the negative obstacle in the correct posture and direction.

[0064] S5. When the quadruped robot's forelegs reach the edge of a negative obstacle, the edge is confirmed based on the foot position information, foot contact state information, and joint feedback information. After confirming that the edge has been reached, the crossing strategy is executed.

[0065] This step involves the quadruped robot simultaneously verifying the foot position information, foot contact state information, and joint feedback information for edge confirmation. Crossing is only triggered when all three edge confirmations are met.

[0066] The edge confirmation specifically refers to the following: when the distance between the position coordinates of the forefoot and the edge of the negative obstacle is less than a preset threshold, the contact state of the forefoot meets the preset contact conditions, and the joint feedback information of the forefoot is within a preset range, the forefoot is confirmed to have reached the edge of the negative obstacle and the crossing strategy is triggered; if any of the above conditions are not met, the crossing strategy is not triggered, and the approach process continues to be adjusted.

[0067] Specifically, determining that the distance between the position coordinates of the forefoot and the edge of the negative obstacle is less than a preset threshold means comparing the three-dimensional coordinates of the current forefoot of the quadruped robot in the world coordinate system with the extracted coordinates of the negative obstacle edge. If the horizontal distance between the coordinates of the current foot and the coordinates of the negative obstacle edge is less than the preset threshold (the preset threshold ranges from 0.02 to 0.05 meters), then the distance between the position coordinates of the forefoot and the edge of the negative obstacle is determined to be less than the preset threshold.

[0068] Determining whether the foreleg contact state meets the preset contact conditions means determining whether the foot contact force is within the normal support range (neither suspended in mid-air nor experiencing a violent impact). The purpose is to prevent visual misjudgment from causing the quadruped robot to accidentally trigger a crossing on flat ground. The preset contact conditions are: foot contact force... satisfy ≤ ≤ ,in This is the critical value (close to 0) at which the foot loses contact with the ground. It is 120% of the rated load at the foot end, used to prevent false triggering due to visual misjudgment.

[0069] The joint feedback information of the forefoot being within the preset range means that the angle, angular velocity, and driving current (torque) of the forefoot joint are read, and the joint angle does not exceed the physical limit and the driving current is stable (without abnormal impact or uncontrolled swaying). When the joint feedback information of the forefoot is within the preset range, it proves that the forefoot is in a stable support phase.

[0070] The triggering of the crossing strategy means that when the above three conditions are met simultaneously, the control system confirms that "the foreleg has truly touched the edge of the groove," immediately generates a trigger signal, and formally starts the crossing action sequence in step S6 (i.e., significantly raising the leg, moving the body forward, and landing on the opposite side). If any condition is not met, the system determines that the state is unsafe, refuses to trigger the crossing, and reverts to step S4 (continuously correcting the approach process), fine-tuning the quadruped robot's position until all three conditions are met.

[0071] S6. Perform the negative obstacle crossing action, and resume normal walking after completing the crossing.

[0072] After triggering the crossing strategy, the quadruped robot executes a pre-planned sequence of crossing actions: lifting its forelegs, moving its body forward, following with its hind legs, and landing on the opposite safe area. After completing the crossing, the quadruped robot confirms stable landing using foot contact force sensors, then switches back to normal walking mode and resumes normal walking.

[0073] The entire process of the negative obstacle crossing action is divided into the following control states: normal walking state, negative obstacle recognition state, adaptive approach state, gait switching state, foot edge confirmation state, crossing execution state, and walking recovery state. The quadruped robot adaptively switches between these control states based on its motion state and environmental perception information.

[0074] To better adapt to different environments, the negative obstacle crossing action in this embodiment employs a reinforcement learning strategy. This reinforcement learning strategy involves introducing at least one random perturbation from visual noise, perception delay, control delay, friction coefficient variation, robot mass variation, and center of mass position variation during training for domain randomization training. To better implement this invention, the visual noise is Gaussian noise with a mean of 0 and a standard deviation of 0.02–0.05 meters added to the depth map; the perception delay is a random delay of 0.05–0.15 seconds; the control delay is a random delay of 0.01–0.03 seconds; the friction coefficient varies from 0.4 to 1.2; the robot mass variation is ±10% of the nominal value; and the center of mass position variation is a random offset of ±0.02 meters.

[0075] As described above, the present invention can be well implemented.

Claims

1. A method for controlling the motion of a quadruped robot traversing negative obstacles based on visual perception, characterized in that, Includes the following steps: S1. Collect visual information including RGB images and depth maps, as well as the body state information of the quadruped robot; S2. Extract the geometric and semantic information of negative obstacles from visual information and integrate the ontological state information to establish a negative obstacle environment model for motion control. S3. Based on the negative obstacle spatial attributes in the negative obstacle environment model and the real-time motion state of the quadruped robot, construct the area to be crossed. After the quadruped robot enters the area to be crossed, adaptively adjust the movement speed, step frequency, body posture and center of gravity position to achieve a smooth transition from the normal walking state to the negative obstacle crossing state. S4. After the quadruped robot enters the area to be crossed, it switches to the negative obstacle crossing control mode and continuously uses visual information to correct the approach process. S5. When the quadruped robot’s forelegs reach the edge of the negative obstacle, the edge is confirmed based on the foot position information, foot contact state information and joint feedback information. After confirming that the edge has been reached, the crossing strategy is triggered. S6. Perform the negative obstacle crossing action, and resume normal walking after completing the crossing.

2. The method for controlling the negative obstacle crossing motion of a quadruped robot based on visual perception according to claim 1, characterized in that, The geometric information in step S2 includes at least one of the width, depth, edge position, and edge direction of the negative obstacle; the semantic information includes the category information of the negative obstacle, and the category includes at least one of the ditch, pothole, step, and collapsed area.

3. The method for controlling the negative obstacle crossing motion of a quadruped robot based on visual perception according to claim 1, characterized in that, The distance threshold of the area to be crossed in step S3 is determined by the geometry of the negative obstacle, the movement speed of the quadruped robot, the gait cycle, and the current movement state.

4. The method for controlling the negative obstacle crossing motion of a quadruped robot based on visual perception according to claim 1, characterized in that, The negative obstacle crossing control mode mentioned in step S4 is a dedicated gait mode for negative obstacle crossing. The foot trajectory of the dedicated gait mode has a higher leg lift height than that of the normal walking gait, and the body posture is more forward-leaning than that of the normal walking gait.

5. The method for controlling the negative obstacle crossing motion of a quadruped robot based on visual perception according to claim 1, characterized in that, The edge confirmation mentioned in step S5 specifically refers to: when the distance between the position coordinates of the forefoot and the negative obstacle edge is less than a preset threshold, the contact state of the forefoot meets the preset contact conditions, and the joint feedback information of the forefoot is within a preset range, the forefoot is confirmed to have reached the negative obstacle edge and the crossing strategy is triggered; if any of the above conditions are not met, the crossing strategy is not triggered, and the approach process continues to be adjusted.

6. The method for controlling the negative obstacle crossing motion of a quadruped robot based on visual perception according to claim 1, characterized in that, The negative obstacle crossing action described in step S6 has a reinforcement learning strategy. This reinforcement learning strategy refers to introducing at least one random perturbation among visual noise, perception delay, control delay, friction coefficient change, robot mass change, and center of mass position change during the training process to perform domain randomization training.

7. The method for controlling the negative obstacle crossing motion of a quadruped robot based on visual perception according to claim 3, characterized in that, The formula for calculating the distance threshold of the area to be traversed in step S3 is as follows: ,in, , Negative obstacle depth This represents the current speed of the quadruped robot. This represents the current gait cycle, and the state is the current motion state of the quadruped robot.

8. The method for controlling the negative obstacle crossing motion of a quadruped robot based on visual perception according to claim 7, characterized in that, The "adaptive adjustment of movement speed, step frequency, body posture, and center of gravity position" mentioned in step S3 triggers adaptive proximity control. The triggering condition for this adaptive proximity control is that the horizontal distance d between the quadruped robot and the edge of the negative obstacle satisfies d ≤ The This represents the distance threshold of the area to be traversed.

Images