A method for continuous memory of a quadruped robot through an obstacle mission critical environment

Abstract

The application discloses a kind of four-legged robot through obstacle task critical environment's continuous memory method, including S1, collection RGB image, depth map or point cloud, IMU data, ontology state information and environment perception information;S2, from the information data collected in step S1 Identify the key environmental information that influences the robot to pass through the task of obstacle, extract geometric information, motion information, interaction information and semantic information four categories of key environmental information and the like steps.The application establishes key environmental continuous memory model, so that key information such as obstacle edge, landing area can still be stably maintained after leaving visual field, effectively overcomes the memory loss problem caused by limited visual field, perception delay and occlusion in traditional method, significantly improves the stability of robot in visual degradation environment.

Description

Technical Field

[0001] This invention belongs to the fields of robot motion control, embodied intelligence and artificial intelligence technology, specifically referring to a method for a quadruped robot to continuously memorize the key environment of an obstacle course. Background Technology

[0002] Quadruped robots, with their excellent terrain adaptability, can be widely used in patrol, transportation, rescue, and exploration tasks, maintaining dynamic balance and stable walking on complex terrain. In obstacle course tasks, quadruped robots need to face various complex environments such as positive obstacles, negative obstacles, suspended obstacles, and narrow passages, which places extremely high demands on the robot's environmental perception, motion decision-making ability, and control robustness.

[0003] Existing quadruped robots primarily rely on current visual observation for environmental perception when performing obstacle-crossing tasks. As the robot approaches an obstacle, obstacle edges, passable areas, footholds, and other crucial environmental information gradually disappear from the camera's field of view. Furthermore, due to delays in image acquisition, communication, and model inference, environmental information exhibits a certain degree of lag. This means that when the robot most needs critical environmental information to guide precise maneuvers such as crossing, climbing, or navigating around obstacles, this information may no longer be within its current field of view or may have become outdated.

[0004] Currently, there are also some studies involving robot environmental memory and world models. For example, some schemes propose robot navigation methods based on external memory, which couple external memory resources with neural networks through a neural Turing machine structure to achieve the memorization of historical pose sequences. Although this method introduces an external memory mechanism, its memory object is the robot's historical pose sequence rather than the environmental information itself, and it is mainly aimed at general navigation tasks, without specifically processing the key environmental information in obstacle-crossing tasks. Another scheme proposes a legged robot motion control method based on a world model, which predicts the current state information by inputting the previous moment's state information, encoded information, and historical actions into the world model. However, the world model in this method is mainly used for state prediction, lacking a dedicated extraction and persistent memory mechanism for key environmental information in obstacle-crossing tasks. In addition, some studies have proposed robot behavior decision-making methods based on memory sequence playback mechanisms, and multi-scenario rapid motion planning methods for robotic arms that integrate memory mechanisms, but none of these schemes have established a dedicated key environmental persistent memory model specifically for the characteristics of quadruped robots crossing obstacles.

[0005] In general, existing world models, reinforcement learning models, and visual-motor models typically build environmental representations for the overall scene, lacking a mechanism for sustained memory of key environments for obstacle crossing tasks. This makes it difficult to maintain stable and consistent environmental cognition after key environmental information leaves the current field of vision, thus affecting the reliability and robustness of subsequent motion imagination prediction, future risk assessment, and motion control. Summary of the Invention

[0006] The purpose of this invention is to overcome the current shortcomings of quadruped robots, which lack a continuous memory mechanism for key environments in obstacle-crossing tasks and have difficulty maintaining stable and consistent environmental cognition after key environmental information leaves the current field of vision. This invention provides a method for continuous memory of key environments in obstacle-crossing tasks for quadruped robots.

[0007] To achieve the above objectives, the present invention employs the following technical solution: a method for a quadruped robot to continuously memorize the key environment of an obstacle course, comprising the following steps: S1. Acquire RGB images, depth maps or point clouds, IMU data, body state information, and environmental perception information; S2. Identify key environmental information that affects the robot's obstacle crossing task from the information data collected in step S1, and extract four types of key environmental information: geometric information, motion information, interaction information, and semantic information. S3. Unify the encoding of the four types of key environmental information and establish a key environmental continuous memory model; S4. Combining the quadruped robot's motion state and new environmental observations, dynamically update the key environment's continuous memory; when the key environment leaves the current field of view, continue to maintain a stable and consistent environmental representation based on the key environment's continuous memory; S5 continuously outputs a unified environmental representation to the motion imagination prediction module, future risk assessment module, and motion control module, providing environmental information support for the robot to complete obstacle crossing tasks.

[0008] Furthermore, the geometric information mentioned in step S2 includes at least one of the following: obstacle edge, obstacle size, spatial structure, passable area boundary, slope, height, width, and depth.

[0009] The motion information mentioned in step S2 includes at least one of the following: relative distance between the robot and the obstacle, relative posture, motion speed, motion direction, and approach state.

[0010] The interaction information in step S2 includes at least one of the expected contact area, expected landing area, expected passage area, and expected crossing area, as well as the interaction relationship between the robot and the obstacle.

[0011] The semantic information mentioned in step S2 includes at least one of the following: passable area, impassable area, danger area, safe area, and obstacle category.

[0012] The obstacles mentioned in steps S2 and S4 include at least one of positive obstacles, negative obstacles, suspended obstacles, boundary obstacles, and narrow passages.

[0013] The key environmental information mentioned in step S2 refers to the environmental information that affects the robot's decision-making in the obstacle-crossing task.

[0014] The dynamic updating of the key environment continuous memory mentioned in step S4 refers to the fusion and updating of the key environment continuous memory in combination with the continuously updated environmental observations during the continuous movement of the robot, so as to realize the continuous evolution of the environment representation.

[0015] The unified environment representation mentioned in step S5 refers to a unified data structure composed of four types of key environmental information. The motion imagination prediction module, future risk assessment module, and motion control module all use this unified environment representation as input.

[0016] Step S3 involves uniformly encoding the four types of key environmental information to establish a key environmental persistent memory model, which includes the following steps: S31. Spatiotemporal alignment and multimodal feature embedding are performed on geometric information, motion information, interaction information and semantic information respectively to generate four types of high-dimensional feature vectors; S32. Calculate the dynamic fusion weights of various feature vectors through a cross-modal attention mechanism, and then weight and fuse the four types of high-dimensional feature vectors into a global feature vector based on the weights. S33. Associate the global feature vector with the current obstacle index key, instantiate it into a structured memory tensor, and calculate and attach the confidence score to complete the establishment of the critical environment persistent memory model.

[0017] Compared with the prior art, the beneficial effects of the present invention are: (1) By establishing a key environment continuous memory model, the present invention can stably maintain key information such as obstacle edges and landing areas after leaving the visual field of vision, effectively overcoming the memory loss problem caused by limited field of vision, perception delay and occlusion in traditional methods, and significantly improving the robot's stability in visual degradation environment.

[0018] (2) The present invention encodes four types of key environmental information into a structured memory tensor and uses it as a common input for the motion imagination prediction, risk assessment and motion control modules, thereby eliminating information conflicts and inconsistencies caused by each module using different environmental representations independently, and fundamentally improving the synergy and reliability of the decision chain.

[0019] (3) In the encoding process, the present invention introduces a cross-modal attention mechanism, which enables the fusion weight to be automatically adjusted according to the robot's movement stage (approaching, crossing, leaving), combined with Bayesian confidence update, to achieve continuous and accurate evolution of environmental representation, taking into account both the stability of memory and the sensitivity of response to new observations.

[0020] (4) This invention differs from the traditional method of modeling the entire scene indiscriminately. It extracts only the key environmental information that affects the decision to pass through obstacles for continuous memory. Under the premise of ensuring the integrity of the environmental information required for decision-making, it greatly compresses the data scale of the memory model and improves the engineering feasibility under the condition of limited onboard computing power of quadruped robots. Attached Figure Description

[0021] Figure 1 This is a schematic diagram of the overall process of the present invention. Detailed Implementation

[0022] 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.

[0023] Example

[0024] like Figure 1 As shown, the method for continuous memory of key environments in obstacle courses by a quadruped robot according to the present invention includes steps S1 to S5, and the content of each step is described in detail below.

[0025] S1. Acquire RGB images, depth maps or point clouds, IMU data, body state information, and environmental perception information.

[0026] This step involves the quadruped robot collecting information and data through its built-in functional modules. The quadruped robot internally includes at least a perception module, a key environment recognition module, an encoding module, a memory maintenance module, an output module, a motion prediction module, a future risk assessment module, and a motion control module.

[0027] The perception module is used to acquire RGB images, depth maps or point clouds, IMU data, body state information, and environmental perception information. The environmental perception information refers to multi-source data, in addition to RGB images, depth maps or point clouds, and IMU data, collected by other external sensors mounted on the quadruped robot, used to characterize the physical properties and dynamic changes of the external environment.

[0028] The environmental perception information includes, but is not limited to: external torque sensing information, which is the external contact force and torque when the sole of the foot comes into contact with an obstacle (or the ground) as sensed by foot or joint force sensors.

[0029] The body state information includes at least one of the following: robot joint angles, joint angular velocities, foot contact forces, and body posture angles.

[0030] Auxiliary ranging information consists of discrete distance data acquired through ultrasonic sensors, infrared ranging sensors, or lidar. This information is primarily used to compensate for the perception deficiencies of RGB-D cameras when facing transparent obstacles (such as glass), highly reflective surfaces, or blind spots along the lower edge of suspended obstacles, thereby improving the accuracy of spatial depth in geometric information.

[0031] Environmental physical quantity information refers to ambient air pressure / altitude change data obtained through barometers or temperature and humidity sensors. When traversing negative obstacles (potholes), the minute altitude changes provided by the barometer can serve as a redundancy check for depth map data, helping to determine whether the depth change is caused by terrain or sensor noise.

[0032] The key environment identification module is used to identify key environmental information that affects the robot's obstacle crossing task from the collected environmental perception information, and extract four types of key environmental information: geometric information, motion information, interaction information, and semantic information.

[0033] The encoding module is used to uniformly encode the four types of key environmental information and establish a key environmental continuous memory model.

[0034] The memory maintenance module is used to dynamically update the continuous memory of the key environment by combining the robot's motion state and new environmental observations, and to continue to maintain a stable and consistent environmental representation based on the continuous memory of the key environment after the key environment leaves the current field of vision.

[0035] The output module is used to continuously output a unified environmental representation to the motion imagination prediction module, the future risk assessment module, and the motion control module.

[0036] The motion imagination prediction module generates multiple candidate foot trajectories and body posture sequences in the action space based on the current environment representation and the robot's body state. It also uses the built-in dynamics model of the World Model to deduce the state changes of each candidate trajectory in the future several control cycles. The module outputs the predicted state sequence (such as the body posture change curve in the next 0.5 seconds) for subsequent evaluation.

[0037] The future risk assessment module is used to calculate risk indicators for each candidate trajectory, including but not limited to collision risk (minimum distance between the torso / feet and the obstacle), stability margin (whether the zero-moment point ZMP falls within the supporting polygon), and slip / overturning risk (estimated based on the friction coefficient of the contact area in the interactive information). Trajectories exceeding the safety threshold are marked or eliminated, and a risk score or risk level is output for each candidate trajectory.

[0038] The motion control module is used to convert the planned foot trajectory and body posture sequence into the desired angle, angular velocity and foot force commands of each joint; at the same time, it drives the actuators of each joint of the robot through the underlying servo control (such as PID control, model predictive control MPC) to achieve stable dynamic walking, crossing or climbing, and provides real-time feedback on execution errors and fine-tunes the control parameters to adapt to the slight undulations of the ground.

[0039] S2. Identify key environmental information affecting the robot's obstacle crossing task from the information data collected in step S1, and extract four types of key environmental information: geometric information, motion information, interaction information, and semantic information.

[0040] This step involves identifying key environmental information from the collected environmental perception information. This key environmental information is the environmental information that affects the robot's decision-making in the obstacle crossing task.

[0041] The geometric information includes at least one of the following: obstacle edge, obstacle size, spatial structure, passable area boundary, slope, height, width, and depth. For example, information such as the edge position, height, depth, and width of the step can be extracted from an RGB image and a depth map.

[0042] The motion information includes at least one of the following: relative distance between the robot and the obstacle, relative posture, motion speed, motion direction, and approach state. For example, the relative distance and relative posture information between the robot and the step are calculated for the relative position of the robot and the step.

[0043] The interaction information includes at least one of the expected contact area, expected landing area, expected passage area, and expected crossing area, such as the expected landing area and expected crossing area based on the robot's current motion state and the geometric features of the steps, as well as the interaction relationship between the robot and the obstacle.

[0044] The semantic information includes at least one of passable areas, impassable areas, dangerous areas, safe areas, and obstacle categories. For example, to determine whether a step is passable based on its height and depth, information is generated indicating whether it is passable or impassable.

[0045] The aforementioned obstacles include at least one of the following: positive obstacles, negative obstacles, suspended obstacles, boundary obstacles, and narrow passages.

[0046] S3. Unify the encoding of the four types of key environmental information and establish a key environmental persistent memory model, which includes the following steps: S31. Spatiotemporal alignment and multimodal feature embedding are performed on geometric information, motion information, interaction information and semantic information respectively to generate four types of high-dimensional feature vectors.

[0047] This step involves spatiotemporal feature alignment and normalization. Since the four types of key environmental information mentioned above come from different sensors (vision, lidar, IMU, joint encoder) and have different timestamps and coordinate systems, spatiotemporal alignment is required.

[0048] Spatiotemporal alignment includes both time alignment and spatial alignment. Time alignment refers to synchronizing asynchronous data collected by various sensors to the same time period using linear interpolation, with the robot's control cycle as the base timestamp. The calculation formula is as follows: ,in, , These are sensor measurements taken at adjacent time points. , For the corresponding timestamp, t is the base timestamp used for unified synchronization.

[0049] Spatial alignment refers to transforming all the obstacle edge point clouds in geometric information, the expected landing area coordinates in interaction information, and the relative poses in motion information into a world coordinate system with the robot's center of mass as the origin through rigid body transformation, thus forming a spatially consistent environmental representation basis.

[0050] The expression for the world coordinate system is: ,in, R is the position vector in the sensor coordinate system, and R is the rotation matrix. It is a translation vector.

[0051] Geometric information is spatiotemporally aligned to generate a fixed-length pre-geometric feature vector; motion information is spatiotemporally aligned to generate a pre-motion feature vector; interaction information is spatiotemporally aligned to generate a pre-interaction feature vector; and semantic information is spatiotemporally aligned to generate a pre-semantic feature vector.

[0052] After generating the four types of pre-feature vectors mentioned above, multimodal feature embedding is required to generate the final four types of high-dimensional feature vectors. The so-called high-dimensional feature vectors refer to the encoding of the four types of pre-feature vectors after spatiotemporal alignment into high-dimensional feature vectors, so that they can be processed in the same semantic space.

[0053] The multimodal feature embedding of the aforementioned pre-geometric feature vector specifically refers to: using a 3D sparse convolutional neural network (SCNN) to downsample and extract features from the point cloud data of the obstacle, generating a geometric feature vector G of dimension D1. This vector explicitly encodes the obstacle edge curvature, plane normal vector, and spatial occupancy features. The expression for this geometric feature vector G is: ,in, For sparse convolutional neural networks, This is a three-dimensional point cloud coordinate matrix.

[0054] The multimodal feature embedding of the pre-motion feature vector specifically refers to combining the normalized relative distance, relative attitude angle, and approach velocity scalars into a one-dimensional motion state vector, and then mapping it to a D2-dimensional motion feature vector M through a multilayer perceptron (MLP). The expression for this motion feature vector M is: ,in, For multilayer perceptron, This is the original motion state vector.

[0055] The multimodal feature embedding of the pre-interaction feature vector specifically refers to: converting the position coordinates of the expected contact area and the expected landing area into high-dimensional spatial coordinate features through position encoding, concatenating them with the robot's foot trajectory features, and compressing them through a fully connected layer into an interaction feature vector I of dimension D3. The expression for this interaction feature vector I is: ,in, It is a fully connected layer. This is the original interactive space coordinate vector.

[0056] The multimodal feature embedding of the pre-semantic feature vector specifically refers to mapping the semantic segmentation result into a dense semantic feature vector S with dimension D4 through a learnable embedding matrix. The expression for this semantic feature vector S is: ,in, For learnable semantic embedding matrices, This is a one-hot encoding for semantic tags.

[0057] S32. Calculate the dynamic fusion weights of various feature vectors through a cross-modal attention mechanism, and then weight and fuse the four types of high-dimensional feature vectors into a global feature vector based on the weights.

[0058] To achieve adaptive fusion (i.e., dynamically adjusting the weights of the four types of information at different motion stages), this embodiment specifically introduces a cross-modal attention fusion module. This module is used to treat the four feature vectors G, M, I, and S generated above as multimodal inputs and perform the following operations: Constructing a query-key-value (QKV) matrix involves concatenating the four feature vectors and performing a linear transformation to generate the global query matrix Q; then, each feature vector is transformed independently into its own key matrix. Sum matrix The expression for the global query matrix Q is: Key matrix The expression is: Value matrix The expression is: , where i represents G, M, I, and S.

[0059] Calculate the cross-modal attention weights, which involves calculating the global query matrix Q and the weights of each key matrix. The dot product similarity is used to obtain the attention weight coefficients after Softmax normalization. The attention weight coefficient The expression is: ,in, Let be the dimension of the key vector, and + + + =1. Geometric information weights when the robot is in the preparation phase. Automatic lifting; motion information weights when the robot is in the approach phase. It rises automatically.

[0060] Weighted fusion, that is, the value matrix of each modality According to its corresponding attention weight coefficient We perform a weighted summation to obtain the fused global feature vector. Its formula is expressed as = This global feature vector This is the core feature vector that will be subsequently stored in the memory model.

[0061] S33. Associate the global feature vector with the current obstacle index key, instantiate it into a structured memory tensor, and calculate and attach the confidence score to complete the establishment of the critical environment persistent memory model.

[0062] To prevent erroneous information from being retained in memory due to sensor noise or dynamic occlusion, a confidence assessment mechanism is needed. This mechanism specifically involves: calculating the confidence score *c* of the current encoding, which is determined by the current observation quality and the output entropy of the feature extraction network; when a new environmental observation conflicts with an existing memory state, a Bayesian update formula is used to decay or increase the confidence score; if the confidence score is below a preset threshold (set to 0.3), the memory entry is marked as pending verification, and the new observation will be prioritized for coverage in subsequent update steps.

[0063] The expression for the Bayesian update formula is: .in, For posterior confidence level, Let z be the prior confidence level, z be the current environmental observation, x be the existing memory state, ¬x be the memory failure hypothesis, and p be the prior confidence level. To determine the likelihood probability of observing the current data z given that the existing memory state x is accurate, Let z be the likelihood probability of observing the current data z under the memory failure hypothesis.

[0064] The instantiation into a structured memory tensor means that the fused global features are then directed to... This is instantiated as the core data structure of the Critical Environment Persistent Memory Model (PKEM). The PKEM employs a key-value storage structure, where the "key" refers to the memory key, which is the raster index coordinates or unique identifier ID of the current obstacle in the global map; and the "value" refers to the memory value, which is the global feature vector generated from the above steps. The data tuple consists of the confidence score c of the information (initially set to 1.0).

[0065] The specific threshold determination method for "if the confidence level is lower than the preset threshold, the memory entry is marked as to be verified and new observations are used to cover it in subsequent updates" is as follows: (1) When a new frame of current environmental observation z is obtained, the observation features are compared with the memory features of the corresponding key in PKEM, and the feature differences are calculated; (2) The observation likelihood ratio is calculated based on the differences. (3) Substitute the above Bayesian update formula to obtain the updated result. (4) Execute the decision, if If the value is ≥0.7, then the existing memory will be maintained, and the updated value will be used. Save back to PKEM; if 0.3 ≤ If the confidence level is less than 0.7, the memory entry is marked as "to be verified," and the memory value is not updated temporarily, waiting for confirmation in the next frame of observation; if the confidence level is less than 0.3 (preset threshold), it is determined that there is a serious conflict with the existing memory, and the old memory value of the entry is immediately overwritten with the feature vector generated by the current new observation, and the confidence level is reset to 0.3. =0.8 (to avoid failure due to minor disturbances immediately after coverage). The aforementioned This represents the confidence level after the reset.

[0066] The above observation likelihood ratio The specific expression is: ,in, The Euclidean distance between the currently observed feature vector and the existing memory feature vector. This is the preset standard deviation of sensor observation noise.

[0067] S4. Combining the quadruped robot's motion state and new environmental observations, dynamically update the key environment's continuous memory; when the key environment leaves the current field of view, continue to maintain a stable and consistent environmental representation based on the key environment's continuous memory.

[0068] The dynamic updating of the critical environment persistent memory refers to the fusion and updating of the critical environment persistent memory by combining the continuously updated environmental observations during the robot's continuous movement, thereby achieving the continuous evolution of the environmental representation. For example, as the robot continues to move towards the steps, the steps gradually approach the robot and may eventually leave the camera's field of view. During this process, the robot continuously acquires new environmental observations and fuses and updates the critical environment information extracted from the new observations with the existing information in the critical environment persistent memory model. When the steps are completely out of the camera's field of view, the critical environment persistent memory model still retains key information such as the height, depth, and edge position of the steps, allowing the robot to know the existence and specific parameters of the steps without relying on the current vision.

[0069] S5 continuously outputs a unified environmental representation to the motion imagination prediction module, future risk assessment module, and motion control module, providing environmental information support for the robot to complete obstacle crossing tasks.

[0070] The unified environment representation refers to a unified data structure composed of four types of key environmental information. The motion imagination prediction module, future risk assessment module, and motion control module all use this unified environment representation as input.

[0071] The critical environment persistent memory model continuously outputs geometric and interactive information of the steps to the motion imagination prediction module, enabling the module to predict the trajectory of the robot's feet when crossing the steps; it outputs step height and depth information to the future risk assessment module, enabling the module to assess the risk of falling; and it outputs a unified environmental representation to the motion control module, enabling the control module to generate precise commands for leg lift height and crossing distance.

[0072] The three modules mentioned above work collaboratively based on the same environmental representation, forming a real-time loop of prediction → evaluation → execution → re-prediction. This process is repeated in each control cycle (e.g., 500Hz). This invention provides a unified and continuous environmental representation to the three modules through a critical environmental persistent memory model, ensuring that each link in the closed loop makes decisions based on the same environmental cognition. This fundamentally eliminates decision conflicts caused by outdated or inconsistent environmental information used by different modules in traditional methods.

[0073] As described above, the present invention can be well implemented.

Claims

1. A method for the continuous memorization of critical environments in an obstacle-crossing task for a quadruped robot, characterized in that, Includes the following steps: S1. Acquire RGB images, depth maps or point clouds, IMU data, body state information, and environmental perception information; S2. Identify key environmental information that affects the robot's obstacle crossing task from the information data collected in step S1, and extract four types of key environmental information: geometric information, motion information, interaction information, and semantic information. S3. Unify the encoding of the four types of key environmental information and establish a key environmental continuous memory model; S4. Combining the quadruped robot's motion state and new environmental observations, dynamically update the key environment's continuous memory; when the key environment leaves the current field of view, continue to maintain a stable and consistent environmental representation based on the key environment's continuous memory; S5 continuously outputs a unified environmental representation to the motion imagination prediction module, future risk assessment module, and motion control module, providing environmental information support for the robot to complete obstacle crossing tasks.

2. The method for continuous memory of the critical environment of a quadruped robot navigating an obstacle course as described in claim 1, characterized in that, The geometric information mentioned in step S2 includes at least one of the following: obstacle edge, obstacle size, spatial structure, passable area boundary, slope, height, width, and depth.

3. The method for continuous memory of the critical environment of a quadruped robot navigating an obstacle course as described in claim 2, characterized in that, The motion information mentioned in step S2 includes at least one of the following: relative distance between the robot and the obstacle, relative posture, motion speed, motion direction, and approach state.

4. The method for continuous memory of the critical environment of a quadruped robot navigating an obstacle course as described in claim 3, characterized in that, The interaction information in step S2 includes at least one of the expected contact area, expected landing area, expected passage area, and expected crossing area, as well as the interaction relationship between the robot and the obstacle.

5. The method for continuous memory of the critical environment of a quadruped robot navigating an obstacle course as described in claim 4, characterized in that, The semantic information mentioned in step S2 includes at least one of the following: passable area, impassable area, danger area, safe area, and obstacle category.

6. The method for continuous memory of the critical environment of a quadruped robot navigating an obstacle course as described in claim 5, characterized in that, The obstacles mentioned in steps S2 and S4 include at least one of positive obstacles, negative obstacles, suspended obstacles, boundary obstacles, and narrow passages.

7. The method for continuous memory of the critical environment of a quadruped robot navigating an obstacle course as described in claim 6, characterized in that, The key environmental information mentioned in step S2 refers to the environmental information that affects the robot's decision-making in the obstacle-crossing task.

8. The method for continuous memory of the critical environment of a quadruped robot navigating an obstacle course as described in claim 1, characterized in that, The dynamic updating of the key environment continuous memory mentioned in step S4 refers to the fusion and updating of the key environment continuous memory in combination with the continuously updated environmental observations during the continuous movement of the robot, so as to realize the continuous evolution of the environment representation.

9. A method for the continuous memorization of critical environments in an obstacle-crossing task for a quadruped robot according to claim 1, characterized in that, The unified environment representation mentioned in step S5 refers to a unified data structure composed of four types of key environmental information. The motion imagination prediction module, future risk assessment module, and motion control module all use this unified environment representation as input.

10. A method for the continuous memorization of critical environments for a quadruped robot to overcome obstacles, as described in any one of claims 1 to 9, characterized in that... Step S3 involves uniformly encoding the four types of key environmental information to establish a key environmental persistent memory model, which includes the following steps: S31. Spatiotemporal alignment and multimodal feature embedding are performed on geometric information, motion information, interaction information and semantic information respectively to generate four types of high-dimensional feature vectors; S32. Calculate the dynamic fusion weights of various feature vectors through a cross-modal attention mechanism, and then weight and fuse the four types of high-dimensional feature vectors into a global feature vector based on the weights. S33. Associate the global feature vector with the current obstacle index key, instantiate it into a structured memory tensor, and calculate and attach the confidence score to complete the establishment of the critical environment persistent memory model.

Images