A visual degeneration-oriented humanoid robot complex terrain control method and system

Abstract

The application relates to the technical field of robot control, and discloses a complex terrain control method and system for a visual degradation-oriented humanoid robot, which comprises the following steps: obtaining a depth map and performing invalid value replacement, distance clipping and normalization to obtain a pretreated depth map; in the training stage, calculating a return fluctuation from a round reward window, updating visual noise intensity and mapping the visual noise intensity into a Gaussian depth noise standard deviation and a structured depth failure trigger probability, applying Gaussian depth noise and continuous multi-frame structured depth failure to obtain a degraded depth map and write the degraded depth map into a depth history cache; encoding a body history observation to obtain channel scaling parameters and channel bias parameters; extracting convolution features from the depth map and performing scaling and bias to obtain visual latent features through gated recurrent temporal fusion; inputting the body observation, the body history latent features and the visual latent features into a strategy network to output joint target positions; and the application can improve the motion stability of the humanoid robot on complex terrain under visual degradation conditions.

Description

Technical Field

[0001] This invention relates to the field of robot control technology, specifically to a method and system for controlling complex terrain in humanoid robots with vision degradation. Background Technology

[0002] When humanoid robots move through complex terrains such as stairs, slopes, gaps, steps, and undulating ground, they need to simultaneously utilize both proprioceptive perception information and external environment perception information to generate joint control actions. Proprioceptive perception information can include body angular velocity, gravity direction projection, velocity commands, joint positions, joint velocities, historical actions, and gait phase; external environment perception information can be acquired by a depth camera to characterize the geometry of the terrain in front of the robot, the edges of steps, the landing area, and ground undulations.

[0003] Existing reinforcement learning control methods typically configure various terrains in a simulation environment and combine dynamic randomization, terrain learning, and observational perturbation to train control strategies. In actual deployment, depth cameras are affected by changes in viewpoint, reflective surfaces, occlusion, motion blur, ambient light, ranging errors, and abnormal data refresh, which may result in ranging noise, blocky holes, missing stripes, long-range contamination, continuous frame drops, and image freezing in the depth map.

[0004] When the edges of steps, footholds, or terrain heights in the depth map are missing, contaminated, or frozen, the control strategy may produce joint movements that are incompatible with the current terrain. Pixel-level random noise of fixed intensity is insufficient to characterize depth sensor failures that possess spatial continuity and temporal persistence. Applying strong visual degradation in the early stages of training may also affect the formation of basic motor abilities, while continuously using weak visual degradation in the later stages of training limits the adaptability of the control strategy to depth observation anomalies.

[0005] Furthermore, some visual control methods first extract features from the depth map independently, and then stitch them together with the robot's body state. The terrain information represented by the depth map is related to the robot's current pose, joint state, motion phase, and short-term motion trend. The control actions corresponding to the same depth map can be different in different body states.

[0006] Therefore, this application provides a method and system for controlling complex terrain for humanoid robots with visual degradation to solve the above-mentioned technical problems. Summary of the Invention

[0007] The purpose of this invention is to provide a method and system for controlling complex terrain in humanoid robots with visual degradation, in order to solve the technical problems in the prior art where humanoid robots have insufficient robustness in motion control in complex terrain under deep visual degradation conditions due to insufficient coverage of structured visual failure during the training phase, lack of association between visual encoding and body state, and fixed visual noise intensity.

[0008] To address the aforementioned technical problems, this invention provides a method for controlling complex terrain in humanoid robots with visual degradation, comprising:

[0009] Obtain the depth map, and perform invalid value replacement, distance clipping, and normalization on the depth map to obtain a preprocessed depth map; During the training phase, the reward fluctuation is calculated from the round reward window, the visual noise intensity is updated according to the reward fluctuation, the visual noise intensity is mapped to the standard deviation of Gaussian depth noise and the structured depth failure trigger probability, Gaussian depth noise and continuous multi-frame structured depth failure triggered according to the structured depth failure trigger probability are applied to the preprocessed depth map to obtain the degraded depth map. Write the degraded depth map to the depth history cache according to the depth map update time; Encode the ontology historical observations to obtain ontology historical latent features. Input the current ontology observations and ontology historical latent features into the modulation parameter generation network to obtain channel scaling parameters and channel bias parameters. Convolutional features are extracted from the depth map of each frame in the depth history cache. The convolutional features are modulated based on the channel scaling parameter and the channel bias parameter. Gated cyclic temporal fusion is performed on the modulated convolutional features of each frame to obtain the visual latent features. The current ontology observation, ontology historical latent features, and visual latent features are input into the control policy network, which outputs the joint target position and controls the humanoid robot to move in complex terrain based on the joint target position.

[0010] In some specific embodiments, writing the degraded depth map to the depth history cache according to the depth map update time further includes: At the moment the depth map is updated, the oldest depth map in the depth history cache is deleted, and the current degraded depth map is written to the end of the depth history cache; When the training environment is reset, the current degraded depth map is used to fill all cache locations of the depth history cache. When a valid depth map is first obtained during the deployment phase, a preprocessed depth map is used to fill all cache locations of the depth history cache. In subsequent depth map updates, the oldest preprocessed depth map is deleted and the newly obtained preprocessed depth map is written.

[0011] In some specific embodiments, the continuous multi-frame structured depth failure triggered by the structured depth failure trigger probability includes at least one of the following: Blocky holes, bottom occlusion, central foothold area occlusion, stripe loss, whole frame loss, consecutive multi-frame loss, image freeze, and far-distance depth pollution; For blocky holes, lower occlusion, central foothold area occlusion, stripe loss, whole frame loss and consecutive multi-frame loss, generate failure areas in the preprocessed depth map and set the pixel values ​​of the failure areas to zero. Freeze the frame, save the preprocessed depth map at the trigger moment, and replace it with a new depth map within the frame; For deep contamination at long distances, Gaussian perturbation is added to pixels outside a preset distance and the pixels are set to zero with a preset probability; The depth map that has been processed through a preset number of depth map update cycles is written to the depth history cache.

[0012] In some specific embodiments, calculating the reward fluctuation from the round reward window and updating the visual noise intensity according to the reward fluctuation further includes: Calculate the mean and standard deviation of the reward window for each round. Divide the standard deviation of the reward by the sum of the absolute value of the mean reward and a preset positive number to obtain the coefficient of variation of the reward. Based on a preset monotonically decreasing mapping relationship, the coefficient of variation of the reward is converted into the original visual noise intensity; The updated visual noise intensity is synthesized by weighting the previous visual noise intensity and the original visual noise intensity using an exponential moving average method. The standard deviation of Gaussian depth noise and the probability of structured depth failure are calculated linearly between a preset lower limit and a preset upper limit based on the updated visual noise intensity.

[0013] In some specific embodiments, encoding ontology history observations to obtain latent ontology history features further includes: The root node angular velocity, gravity direction projection in the fuselage coordinate system, velocity command, joint position, joint velocity, action of the previous control cycle, and sine and cosine values ​​of the left and right leg gait phases are used to form the body observation at each control moment. Encode the continuous preset number of ontology observation frames in chronological order and output the historical potential features of the ontology; The current ontology observations and historical potential features of the ontology are concatenated according to feature dimensions to generate modulation parameters.

[0014] In some specific embodiments, the convolutional features are modulated based on channel scaling parameters and channel bias parameters, further including: Each frame of depth map in the depth history cache is convolved sequentially to obtain the convolutional features of the depth map output by each convolutional layer. Generate channel scaling parameters and channel bias parameters for each convolutional layer; Based on the channel scaling parameters and channel bias parameters, the convolutional features of the depth map of each convolutional layer are scaled and biased, and the channel scaling parameters and channel bias parameters are shared among the depth maps of each frame in the depth history cache. The modulated convolutional features output from the final convolutional layer are then subjected to gated cyclic temporal fusion according to the depth map acquisition order.

[0015] In some specific embodiments, the current ontology observation, ontology historical latent features, and visual latent features are input into the control policy network, and the joint target position is output, further including: Output normalized joint motion, perform numerical scaling on the normalized joint motion based on a preset motion scaling ratio, and convert it into the joint target position; The ontological observations, rewards, and termination signals generated by the interaction of humanoid robots with complex terrain are combined into trajectory data; During the training phase, the control strategy is updated based on trajectory data, and the joint target position is output based on the updated control strategy.

[0016] In some specific embodiments, the central footing area is obstructed, further including: The rectangle size is determined based on a first preset ratio range for the depth map height and a second preset ratio range for the depth map width. The center of the rectangle is randomly determined within the third preset ratio range of the depth map height and the fourth preset ratio range of the depth map width; Crop the rectangular edges that extend beyond the image boundaries; Set the pixels within the rectangle to zero, and repeat the process of setting the pixels within the rectangle to zero on each newly acquired preprocessed depth map within a preset number of depth map update cycles.

[0017] In some specific embodiments, the updated visual noise intensity is synthesized by weighting the previous visual noise intensity and the original visual noise intensity using an exponential moving average method, further including: Calculate the mean return, standard deviation of return, and coefficient of variation of return within a reward window of a preset number of completed training rounds; When the absolute value of the average return is lower than a preset positive number, the absolute value of the average return is replaced with the preset positive number; The visual noise intensity of the previous round and the original visual noise intensity are weighted and summed using a preset exponential moving average coefficient. The weighted summation result is cropped to a preset visual noise intensity range and mapped to the Gaussian depth noise standard deviation and the structured depth failure trigger probability.

[0018] Based on the same concept, the present invention also provides a humanoid robot complex terrain control system for visual degradation, comprising: The depth map preprocessing module is configured to acquire a depth map and perform invalid value replacement, distance clipping, and normalization on the depth map to obtain a preprocessed depth map. The visual degradation training module is configured to calculate the reward fluctuation from the round reward window during the training phase, update the visual noise intensity according to the reward fluctuation, map the visual noise intensity to the standard deviation of Gaussian depth noise and the structured depth failure trigger probability, apply Gaussian depth noise and continuous multi-frame structured depth failure triggered according to the structured depth failure trigger probability to the preprocessed depth map, and obtain the degradation depth map. The depth history cache module is configured to write degraded depth maps to the depth history cache according to the depth map update time. The ontology history encoding and modulation parameter generation module is configured to encode ontology history observations to obtain ontology history latent features, and input the current ontology observations and ontology history latent features into the modulation parameter generation network to obtain channel scaling parameters and channel bias parameters. The visual feature extraction module is configured to extract convolutional features from the depth map of each frame in the depth history cache, modulate the convolutional features based on the channel scaling parameter and the channel bias parameter, and perform gated cyclic temporal fusion on the modulated convolutional features of each frame to obtain the visual latent features. The motion control module is configured to input the current ontology observation, ontology historical latent features and visual latent features into the control strategy network, output the joint target position, and control the humanoid robot to move in complex terrain based on the joint target position.

[0019] Compared with existing technologies, its advantages are as follows: This invention discloses a method and system for controlling complex terrain in humanoid robots to address visual degradation. During the training phase, Gaussian depth noise and structured depth failures that persist for multiple frames are simultaneously applied to the preprocessed depth map. This causes the training data to contain depth observation states such as blocky holes, missing frames, image freezes, and long-distance depth contamination. As a result, the control strategy is fully exposed to depth sensor anomalies with spatial continuity and temporal persistence during the training phase, thereby improving the adaptability to real depth observation degradation during the deployment phase.

[0020] The visual noise intensity is adaptively adjusted based on the reward fluctuation in the round reward window, and the updated visual noise intensity is synchronously mapped to the standard deviation of Gaussian depth noise and the structured depth failure trigger probability. This ensures that the visual degradation intensity matches the current training performance of the policy, avoids strong visual degradation in the early stage of training causing the policy to fail to converge, and ensures that the policy has sufficient visual robustness in the later stage of training.

[0021] By utilizing current ontology observations and historical latent features, channel scaling and bias parameters are generated, and the convolutional features of the depth map are scaled and biased to produce differentiated visual feature expressions for the same depth map under different body poses, joint states, and motion phases, thereby improving the matching degree between visual features and the current motion state.

[0022] Gated cyclic temporal fusion is performed on the depth maps of each frame in the depth history cache to form visual latent features containing time information. This enables the control strategy to maintain the continuity of control actions by utilizing historical effective depth information and the body's motion trend when a single frame depth map has local missing parts, a whole frame is lost, or the screen freezes.

[0023] By inputting the current ontology observation, historical ontology features, and visual features into the control strategy network, the control strategy can still output stable joint control actions based on the ontology state, historical depth information, and modulated visual features when the depth map is noisy, locally missing, continuously dropped, or frozen. This reduces the humanoid robot's dependence on ideal depth observation and improves its motion stability and mobility in complex terrain. Attached Figure Description

[0024] Other features, objects, and advantages of this application will become more apparent from the following detailed description of non-limiting embodiments with reference to the accompanying drawings: Figure 1 This is a flowchart illustrating some specific embodiments of the present invention, which describes a method for controlling complex terrain in humanoid robots with visual degradation. Figure 2 This is a schematic diagram of the structure of a humanoid robot complex terrain control system for visual degradation according to some specific embodiments of the present invention. Detailed Implementation

[0025] To make the objectives, technical solutions, and advantages of this application clearer, the application will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of this application, and not all embodiments. Based on the embodiments in this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0026] The terminology used in the embodiments of this application is for the purpose of describing particular embodiments only and is not intended to limit the application. The singular forms "a," "the," and "the" as used in the embodiments of this application are also intended to include the plural forms, unless the context clearly indicates otherwise, and "multiple" generally includes at least two.

[0027] It should be understood that the term "and / or" used in this article is merely a description of the relationship between related objects, indicating that three relationships can exist. For example, A and / or B can represent: A existing alone, A and B existing simultaneously, and B existing alone. Additionally, the character " / " in this article generally indicates that the preceding and following related objects have an "or" relationship.

[0028] It should be understood that although the terms first, second, third, etc., may be used in the embodiments of this application, these descriptions should not be limited to these terms. These terms are only used to distinguish the descriptions. For example, first may also be referred to as second without departing from the scope of the embodiments of this application, and similarly, second may also be referred to as first.

[0029] Depending on the context, the words “if” or “suppose” as used here can be interpreted as “when” or “in response to determination” or “in response to detection.” Similarly, depending on the context, the phrases “if determination” or “if detection (of the stated condition or event)” can be interpreted as “when determination” or “in response to determination” or “when detection (of the stated condition or event)” or “in response to detection (of the stated condition or event).”

[0030] It should also be noted that the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that an article or device that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such an article or device. Without further limitation, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the article or device that includes said element.

[0031] It should be noted that any symbols and / or numbers present in the specification that are not marked in the accompanying drawings are not reference numerals.

[0032] The humanoid robot in this embodiment includes a depth camera, body sensors, a control processor, and joint actuators. The depth camera is mounted on the robot's head and faces the direction of movement, used to acquire depth maps of the terrain in front, the ground near the front, and potential foothold areas. The body sensors are used to obtain body information such as root node angular velocity, body posture, joint positions, joint velocities, and gait phase. The control processor generates target positions for each joint based on the body information and depth maps, and the joint actuators drive the humanoid robot to move through gaps, slopes, steps, gravel surfaces, or undulating terrain according to the target joint positions.

[0033] The training platform is built upon a graphics processing unit (GPU) capable of ray tracing computation, Isaac Lab simulation software, and the RSL-RL reinforcement learning framework. During training, simulated depth maps are obtained through ray casting, while during deployment, depth maps are obtained using a real depth camera. The control policy operates at a frequency of 50 Hz, and the depth map is updated at a frequency of 10 Hz. The control processor updates the depth map every 5 control cycles, reusing the current depth history cache for the remaining control cycles.

[0034] Reference Figure 1A method for controlling complex terrain in humanoid robots with vision degradation, comprising: S101, Obtain the depth map, perform invalid value replacement, distance clipping and normalization on the depth map to obtain a preprocessed depth map; S102, During the training phase, the reward fluctuation is calculated from the round reward window, the visual noise intensity is updated according to the reward fluctuation, the visual noise intensity is mapped to the standard deviation of Gaussian depth noise and the structured depth failure trigger probability, Gaussian depth noise and continuous multi-frame structured depth failure triggered according to the structured depth failure trigger probability are applied to the preprocessed depth map to obtain the degraded depth map. S103, Write the degraded depth map into the depth history cache according to the depth map update time; S104, Encode the ontology historical observations to obtain ontology historical latent features, and input the current ontology observations and ontology historical latent features into the modulation parameter generation network to obtain channel scaling parameters and channel bias parameters. S105: Extract convolutional features from the depth map of each frame in the depth history cache, modulate the convolutional features based on the channel scaling parameter and the channel bias parameter, and perform gated cyclic temporal fusion on the modulated convolutional features of each frame to obtain visual latent features. S106 inputs the current ontology observation, ontology historical latent features, and visual latent features into the control policy network, outputs the joint target position, and controls the humanoid robot to move in complex terrain based on the joint target position.

[0035] To provide a clearer explanation, the steps in the embodiments of the present invention are described in detail below: S101, Obtain the depth map, perform invalid value replacement, distance clipping and normalization on the depth map to obtain a preprocessed depth map; In this embodiment, the original depth map at the current moment is obtained. The original depth map during the training phase is generated by the depth camera model through ray casting, while the original depth map during the deployment phase is acquired by a real depth camera.

[0036] In this embodiment, the original depth map is 36×64 pixels in size. The original depth map is cropped according to the depth camera's mounting height, field of view, and the position of the landing area in the image, removing edge regions with low relevance to motion control, resulting in a 18×32 pixel depth map to be processed.

[0037] Let the depth map to be processed be... , and These represent the row index and column index of the pixel, respectively. Indicates the maximum effective depth. When When the depth marker is an infinite value, non-numerical, null value, or an invalid depth marker output by the depth camera, use Replace the corresponding pixel; when When the depth value is valid, the corresponding pixel is retained, resulting in a depth map after invalid values ​​are replaced. : ; in, Coordinates in the depth map to be processed The original depth value at that location, The depth value is the result of replacing the invalid value. This represents the maximum effective depth.

[0038] set up Represents the minimum effective depth, for Perform distance clipping to obtain the clipping depth map in the physical depth domain. : ; in, This is the distance to the depth value after clipping. For the minimum effective depth, This is the maximum effective depth. Distance clipping will be less than... The depth value is set to , will be greater than The depth value is set to This prevents values ​​exceeding the effective measurement range of the depth camera from entering subsequent networks. In this embodiment, m, m.

[0039] right Perform normalization to obtain a preprocessed depth map. : ; in, This is the normalized depth value. The value is used as input to the neural network. Simultaneously, it serves as the physical depth record for the corresponding pixel, used for subsequent determination of the effective distance range and the set of distant pixels. By retaining both the normalized depth map and the physical depth record, dimensional confusion between the distant distance threshold and the noise standard deviation, which are in metric units, and the normalized pixel values ​​is avoided.

[0040] During the training phase, the preprocessed depth map undergoes visual degradation processing, which involves applying Gaussian depth noise and triggering continuous multi-frame structured depth failures based on the probability of structured depth failures, resulting in a degraded depth map. During the deployment phase, the preprocessed depth map is not artificially noiseed; it is directly written to the depth history cache according to the depth map update time. The image size and effective distance mentioned above are exemplary parameters and can be adjusted based on the depth camera resolution, installation pose, and robot movement speed.

[0041] S102, During the training phase, the reward fluctuation is calculated from the round reward window, the visual noise intensity is updated according to the reward fluctuation, the visual noise intensity is mapped to the standard deviation of Gaussian depth noise and the structured depth failure trigger probability, Gaussian depth noise and continuous multi-frame structured depth failure triggered according to the structured depth failure trigger probability are applied to the preprocessed depth map to obtain the degraded depth map. Furthermore, persistent multi-frame structured depth failures triggered by the structured depth failure trigger probability include at least one of the following: Blocky holes, bottom occlusion, central foothold area occlusion, stripe loss, whole frame loss, consecutive multi-frame loss, image freeze, and far-distance depth pollution; For blocky holes, lower occlusion, central foothold area occlusion, stripe loss, whole frame loss and consecutive multi-frame loss, generate failure areas in the preprocessed depth map and set the pixel values ​​of the failure areas to zero. Freeze the frame, save the preprocessed depth map at the trigger moment, and replace it with a new depth map within the frame; For deep contamination at long distances, Gaussian perturbation is added to pixels outside a preset distance and the pixels are set to zero with a preset probability; The depth map that has been processed through a preset number of depth map update cycles is written to the depth history cache.

[0042] Furthermore, the reward fluctuation is calculated from the round reward window, and the visual noise intensity is updated according to the reward fluctuation, including: Calculate the mean and standard deviation of the reward window for each round. Divide the standard deviation of the reward by the sum of the absolute value of the mean reward and a preset positive number to obtain the coefficient of variation of the reward. Based on a preset monotonically decreasing mapping relationship, the coefficient of variation of the reward is converted into the original visual noise intensity; The updated visual noise intensity is synthesized by weighting the previous visual noise intensity and the original visual noise intensity using an exponential moving average method. The Gaussian depth noise standard deviation and the structured depth failure trigger probability are calculated linearly between a preset lower limit and a preset upper limit based on the updated visual noise intensity.

[0043] Furthermore, central landing area obstruction includes: The rectangle size is determined based on a first preset ratio range for the depth map height and a second preset ratio range for the depth map width. The center of the rectangle is randomly determined within the third preset ratio range of the depth map height and the fourth preset ratio range of the depth map width; Crop the rectangular edges that extend beyond the image boundaries; Set the pixels within the rectangle to zero, and repeat the process of setting the pixels within the rectangle to zero on each newly acquired preprocessed depth map within a preset number of depth map update cycles.

[0044] Furthermore, the updated visual noise intensity is synthesized by weighting the previous round's visual noise intensity and the original visual noise intensity using an exponential moving average method, including: Calculate the mean return, standard deviation of return, and coefficient of variation of return within a reward window of a preset number of completed training rounds; When the absolute value of the average return is lower than a preset positive number, the absolute value of the average return is replaced with the preset positive number; The visual noise intensity of the previous round and the original visual noise intensity are weighted and summed using a preset exponential moving average coefficient. The weighted summation result is cropped to a preset visual noise intensity range and mapped to the Gaussian depth noise standard deviation and the structured depth failure trigger probability.

[0045] In this embodiment, the visual noise intensity is maintained during the training phase. , Indicates the first The visual noise course is updated. The visual noise intensity varies between 0 and 1, simultaneously controlling for the intensity of Gaussian depth noise and structured depth failure. Updates are based on the reward fluctuations after completed training rounds. Then Mapped to Gaussian depth noise standard deviation and structured deep failure trigger probability Therefore, the robot mainly learns basic motor skills in the early stages of training, and gradually encounters more advanced visual degradation after its training performance stabilizes.

[0046] Set the round reward window to include Rewards for each completed training round Calculate the average return and standard deviation of returns : ; ; in, For the first The reward for each completed round. The number of rounds in the reward window. For the average return, The standard deviation of the return.

[0047] Calculate the coefficient of variation of returns based on the mean return and the standard deviation of returns. : ; in, It is a preset positive number to prevent the denominator from becoming invalid when the average return is close to 0.

[0048] In this embodiment, Take 4096, Pick After each training round, the corresponding reward is written to the round reward window; when the window is full, the oldest reward is deleted and a new reward is written.

[0049] The mean reward reflects the overall task completion level, while the coefficient of variation of the reward reflects the consistency of training performance across multiple parallel environments. When some parallel environments stably traverse the terrain while others frequently fail, the mean reward may still be within an acceptable range, but the standard deviation and coefficient of variation of the reward will increase. Therefore, using the coefficient of variation of the reward can help to make the visual noise curriculum respond to fluctuations in training performance.

[0050] In this embodiment, a monotonically decreasing mapping relationship is constructed using the hyperbolic tangent function to calculate the original visual noise intensity. : ; in, The original visual noise intensity, For the coefficient of variation of returns, This means that the value is limited to between 0 and 1.

[0051] when When the value decreases, it indicates that the training performance in multiple parallel environments tends to be consistent. Increase; when An increase indicates significant fluctuations in training performance. Decrease.

[0052] In addition to the hyperbolic tangent function mentioned above, piecewise linear monotonically decreasing functions, monotonically decreasing exponential functions, or preset lookup tables can also be used. The input is the return fluctuation amount, and the output is the original visual noise intensity, while maintaining a monotonically decreasing relationship.

[0053] To avoid abrupt changes in visual degradation intensity between adjacent course updates, visual noise intensity is updated using an exponential moving average method: ; in, The intensity of visual noise in the previous round, This is the coefficient of the exponential moving average.

[0054] In this embodiment, That is, the weight of the previous round of visual noise intensity is 0.99, and the weight of the current original visual noise intensity is 0.01. Pruning is performed when the weighted result exceeds 0 to 1.

[0055] Let the lower and upper limits of the standard deviation of Gaussian depth noise be respectively... and The lower and upper bounds of the probability of structured deep failure are respectively and ,calculate: ; ; in, The current standard deviation of Gaussian depth noise. This represents the current probability of triggering a deep structured failure. This represents the updated visual noise intensity.

[0056] In this embodiment, , m, , When the visual noise intensity increases, the standard deviation of Gaussian depth noise and the probability of structured depth failure triggering increase simultaneously; when the visual noise intensity decreases, both decrease simultaneously, thus coordinating pixel-level measurement error and persistent structural failure with the same course variable.

[0057] Clipping depth map in the physical depth domain Apply Gaussian depth noise. Define the effective distance mask. : ; in, Indicates the current time Pixels Is it within the effective distance range?

[0058] Sample zero-mean Gaussian noise for each valid pixel. : ; The physical depth map after Gaussian degradation is as follows: ; in, For pixels The sampled Gaussian noise, The current standard deviation of Gaussian depth noise. This is the depth value after adding Gaussian noise.

[0059] Will Divide by This generates a Gaussian degradation depth map for use as input to the neural network. The above method only adds ranging noise to pixels within the effective distance range, avoiding the application of physically meaningless perturbations to regions where invalid values ​​are replaced.

[0060] At each depth map update time, according to Perform Bernoulli sampling: ; when When the current Gaussian degradation depth map does not trigger a new structured depth failure; when At that time, select one failure mode from block holes, lower occlusion, central landing area occlusion, stripe loss, whole frame loss, consecutive multi-frame loss, image freeze, and distant depth pollution, and sample the duration from 2 to 4 depth map update cycles. .

[0061] To ensure the temporal continuity of structured depth failures, the visual degradation training process maintains a failure state record for each parallel environment. This record includes at least the failure mode, remaining duration frames, failure region parameters, and a frozen image. Once a failure mode is triggered, it is not reselected until the remaining duration frames reach zero; the remaining duration frames are decremented by 1 for each new depth map update cycle. Thus, structured depth failures remain correlated both spatially and temporally.

[0062] Blocky Holes: Randomly generated axis-aligned rectangles within the entire depth map. The rectangle height occupies 35% to 75% of the depth map height, and the rectangle width occupies 35% to 80% of the depth map width. The top-left corner of the rectangle is randomly determined within a range that ensures the rectangle does not cross the boundaries, and the pixel values ​​within the rectangle are set to 0. For an 18×32 pixel depth map, the rectangle height is approximately 6 to 14 rows, and the rectangle width is approximately 11 to 26 columns.

[0063] Bottom occlusion: A continuous rectangle is generated from the bottom edge of the depth map upwards. The width of the rectangle is equal to the full width of the depth map, and the height of the rectangle accounts for 35% to 70% of the height of the depth map. The pixel values ​​within the rectangle are set to 0. Bottom occlusion mainly covers the ground near the robot's front and the nearby landing area, while preserving distant terrain information in the upper part of the depth map.

[0064] Central landing area occlusion: Determine the rectangle height as 35% to 70% of the depth map height, and the rectangle width as 35% to 75% of the depth map width; randomly determine the rectangle's center ordinate within 55% to 85% of the depth map height, and randomly determine the rectangle's center abscissa within 40% to 60% of the depth map width. Let the depth map height and width be... and The height and width of the rectangle are respectively and The center coordinates of the rectangles are respectively and ,but: ; ; according to , , and Calculate the rectangle boundary, perform cropping on the portion that exceeds the image boundary, set the pixel values ​​within the cropped rectangle to 0, and perform the same region's pixel setting to 0 on each newly acquired depth map in 2 to 4 depth map update cycles.

[0065] Stripe loss: Randomly generate 1 to 3 horizontal or vertical stripes. The position of the horizontal stripes is randomly determined along the height direction of the depth map, and the position of the vertical stripes is randomly determined along the width direction of the depth map. The stripe thickness accounts for 8% to 25% of the corresponding image size. Multiple stripes can overlap, and the pixel values ​​within the stripe area are set to 0.

[0066] Full-frame loss and consecutive multi-frame loss: The mask corresponding to the entire depth map is set as the effective invalid region, and all pixels are set to 0. Full-frame loss and consecutive multi-frame loss use the same full-map 0-setting operation; when the duration is multiple depth map update cycles, the full-map 0-setting is repeatedly performed on each newly acquired depth map within the duration window, thus forming consecutive multi-frame loss.

[0067] Screen freeze: When screen freeze is triggered, save the complete depth map at the moment of triggering. Within the continuous window, newly acquired depth maps are not used; the output is repeated. The image freeze is used to simulate a situation where the robot has moved but the depth camera data has not been updated.

[0068] Long-distance deep contamination: Determined based on physical depth records A set of distant pixels with a standard deviation of m is used. A zero-mean Gaussian perturbation with a standard deviation of 0.08 m is applied to distant pixels, and each distant pixel is set to 0 with a 35% probability, while nearby pixels remain unchanged. The Gaussian perturbation and 0-position are resampled each continuous frame, allowing the noise pattern of distant contamination to vary over time.

[0069] After processing Gaussian depth noise and structured depth failures, the physical depth values ​​are re-constrained within the effective distance range and normalized to obtain the degraded depth map. .

[0070] S103, Write the degraded depth map into the depth history cache according to the depth map update time; Furthermore, at the moment the depth map is updated, the oldest depth map in the depth history cache is deleted, and the current degraded depth map is written to the end of the depth history cache; When the training environment is reset, the current degraded depth map is used to fill all cache locations of the depth history cache. When a valid depth map is first obtained during the deployment phase, a preprocessed depth map is used to fill all cache locations of the depth history cache. In subsequent depth map updates, the oldest preprocessed depth map is deleted and the newly obtained preprocessed depth map is written.

[0071] In this embodiment, an independent deep history cache is maintained for each parallel training environment. Let the current control time be... Deep history cache includes Frame depth map, interval between adjacent buffer frames Each control cycle, then deep history cache for: ; in, for Deep historical cache at any moment This is the degradation depth map at the corresponding time point. For the number of buffered frames, This refers to the number of interval control cycles.

[0072] In this embodiment, , The control strategy operates at 50 Hz, and the depth map is updated at 10 Hz. Therefore, the time interval between adjacent buffer frames is 0.1 s, and there is a time span of about 0.3 s between the oldest and the newest depth maps. The 4-frame strafing buffer reduces redundancy between high-frequency adjacent images while retaining short-term temporal information about terrain changes relative to the robot.

[0073] At the moment a depth map is updated, the oldest depth map is deleted from the depth history cache, and the remaining depth maps are shifted forward one cache position in sequence. The current degraded depth map is then written to the end of the cache. Between two consecutive depth map updates, the control strategy reads the same depth history cache and avoids repeatedly collecting depth maps.

[0074] When the training environment is reset, the robot's initial state, terrain type, terrain difficulty, or initial position may change, and the historical depth map before the reset will no longer correspond to the robot and terrain after the reset. The current degraded depth map after the reset is used to fill all cache positions of the depth history cache. Subsequent new depth maps replace the repeatedly filled depth maps frame by frame according to the update time, thereby avoiding the image before the reset from polluting the visual history after the reset.

[0075] When a valid depth map is first obtained during the deployment phase, the current preprocessed depth map is used to fill all cache positions in the depth history cache. Subsequently, for each new valid depth map frame obtained, the oldest preprocessed depth map is deleted and the new preprocessed depth map is written to the end of the cache. Degenerate depth maps are written during the training phase, and real preprocessed depth maps are written during the deployment phase. Both phases use the same number of cached frames, the same sorting order, and the same update frequency to maintain consistency in the temporal input structure between the training and deployment phases.

[0076] S104, Encode the ontology historical observations to obtain ontology historical latent features, and input the current ontology observations and ontology historical latent features into the modulation parameter generation network to obtain channel scaling parameters and channel bias parameters. Furthermore, the ontological history observations are encoded to obtain latent ontological history features, including: The root node angular velocity, gravity direction projection in the fuselage coordinate system, velocity command, joint position, joint velocity, action of the previous control cycle, and sine and cosine values ​​of the left and right leg gait phases are used to form the body observation at each control moment. Encode the continuous preset number of ontology observation frames in chronological order and output the historical potential features of the ontology; The current ontology observations and historical potential features of the ontology are concatenated according to feature dimensions to generate modulation parameters.

[0077] In this embodiment, current body observations are collected at each control moment. And maintain the ontological historical observations for 10 consecutive control moments.

[0078] In this embodiment, the humanoid robot has 29 controlled joints, and the ontology observation at each control moment is 100-dimensional. The observation items are shown in the table below.

[0079] Table 1: Components of the ontological observation:

[0080] Combining 10 consecutive body observations in chronological order Input ontology historical encoding network To obtain the latent features of the ontology's history : ; in, As a potential feature of ontology history, This is a sequence of 10 consecutive body observation frames. The network parameters are used to encode the ontology's history.

[0081] In this embodiment, the ontology history encoding network includes two fully connected hidden layers with dimensions of 256 and 128 respectively. The hidden layers use a linear rectified activation function to output 64-dimensional latent features of ontology history. It is used to reflect the direction of joint movement, the trend of posture change, the continuity of movement and the evolution of gait over 10 consecutive control moments.

[0082] The current ontology observations are concatenated with the historical potential features of the ontology according to feature dimensions to form a joint modulation condition. : ; The modulation parameter generation network is used for the first modulation parameter generation network of the depth map encoder. Each convolutional layer generates channel scaling parameters. and channel offset parameters : ; in, and The first The channel scaling and channel bias parameters of each convolutional layer have the same dimension as the number of output channels of the corresponding convolutional layer.

[0083] Current ontology observations reflect the posture, joint state, and gait phase at the current control moment, while ontology history latent features reflect short-term motion processes prior to the current control moment. For two states with similar current postures but different previous motion directions or gait processes, joint modulation conditions are used. They can be different, thus generating different state-related visual features from the same or similar depth maps.

[0084] S105: Extract convolutional features from the depth map of each frame in the depth history cache, modulate the convolutional features based on the channel scaling parameter and the channel bias parameter, and perform gated cyclic temporal fusion on the modulated convolutional features of each frame to obtain visual latent features. Furthermore, the convolutional features are modulated based on channel scaling parameters and channel bias parameters, including: Each frame of depth map in the depth history cache is convolved sequentially to obtain the convolutional features of the depth map output by each convolutional layer. Generate channel scaling parameters and channel bias parameters for each convolutional layer; Based on the channel scaling parameters and channel bias parameters, the convolutional features of the depth map of each convolutional layer are scaled and biased, and the channel scaling parameters and channel bias parameters are shared among the depth maps of each frame in the depth history cache. The modulated convolutional features output from the final convolutional layer are then subjected to gated cyclic temporal fusion according to the depth map acquisition order.

[0085] In this embodiment, the depth history cache is read from the depth map acquisition time in ascending order. Frame depth map. For the first frame depth map... Frame depth map The depth map is input into a multi-layer convolutional network.

[0086] In this embodiment, the multi-layer convolutional network includes two convolutional layers. The number of output channels in the first and second layers are 8 and 16, respectively. The kernel size of both layers is 3×3, and the stride of both layers is 1. A linear rectified activation function is used after the convolutional layers.

[0087] Let the first Frame depth map in the The input features of each convolutional layer are The convolution kernel and bias are respectively and Convolutional features for: ; in, This represents a convolution operation. It reads the generated data. and Based on the channel scaling parameters and channel bias parameters, the depth map convolutional features of each convolutional layer are scaled and biased to obtain the modulated convolutional features. : ; in, This represents channel-level multiplication. and Broadcast in the spatial dimension of convolutional features.

[0088] The frames in the deep history buffer share the current control time determined by joint modulation conditions. The same set of channel scaling parameters and channel offset parameters are generated so that each historical frame can be visually encoded under the same robot motion state.

[0089] Flatten the features of the last layer of modulation convolution or form a temporal input vector through a fully connected mapping. Gated cyclic timing fusion is performed according to the depth map acquisition sequence.

[0090] In this embodiment, the gated cyclic timing fusion updates the hidden state according to the following relationship: ; ; ; ; in, To reset the door, To update the door, and These are the hidden states corresponding to the previous and current historical frames, respectively. After processing... After the frame depth map is generated, the hidden state of the last frame is used as the visual latent feature. : ; Gated cyclic temporal fusion enables the control strategy to not only use the current depth map but also incorporate historical terrain changes over a 0.3s time span. When the current depth map experiences local holes, frame loss, or image freeze, gated cyclic temporal fusion can generate latent visual features based on previously valid depth features and body motion trends, reducing the degree to which single-frame visual anomalies are directly transmitted to motion output.

[0091] S106 inputs the current ontology observation, ontology historical latent features, and visual latent features into the control policy network, outputs the joint target position, and controls the humanoid robot to move in complex terrain based on the joint target position.

[0092] Furthermore, the current ontology observations, historical latent features, and visual latent features are input into the control policy network, which outputs the joint target location, including: Output normalized joint motion, perform numerical scaling on the normalized joint motion based on a preset motion scaling ratio, and convert it into the joint target position; The ontological observations, rewards, and termination signals generated by the interaction of humanoid robots with complex terrain are combined into trajectory data; During the training phase, the control strategy is updated based on trajectory data, and the joint target position is output based on the updated control strategy.

[0093] In this embodiment, the current ontology observation Ontological historical potential characteristics and visual latent features By concatenating features along their dimensions, a comprehensive input to the control policy network is formed. : ; In this embodiment, the control strategy network includes three fully connected hidden layers with dimensions of 512, 256, and 128 respectively, outputting 29-dimensional normalized joint motion. Normalized joint movements are scaled according to the movement scaling ratio. Perform scaling and compare with the default joint position. Combine to obtain the target joint position. : ; in, .

[0094] The target joint positions are sent to the joint actuators, which then drive the 29 joints to move according to these positions. The motion scaling can be set separately for different joint groups, and trimming can be performed after output based on the soft position limits of each joint.

[0095] During the training phase, the humanoid robot interacts with complex terrain based on joint target positions. The environment returns the body observations, depth map, reward, and termination signal for the next control moment. The trajectory data includes at least the current body observations, body history observations, depth history cache, normalized joint actions, joint target positions, reward, termination signal, and observation data for the next control moment.

[0096] Training objectives include policy loss. Value function loss Entropy regularization term The total training loss is expressed as: ; in, and These are the weights of the value function loss and the entropy regularization term, respectively.

[0097] The training loss is calculated based on the sampled trajectory, and the parameters of the control policy network, ontology history encoding network, modulation parameter generation network, multi-layer convolutional network, and gated recurrent temporal fusion network are updated. After training, the trained network parameters are saved and used to output joint target positions during the deployment phase.

[0098] Total reward at each control moment The sum is calculated by weighting multiple reward items: ; in, Indicates the first One reward item, This indicates the corresponding weight. The reward items are shown in the table below.

[0099] Table 2: Reinforcement Learning Rewards

[0100] The parameters and weights in Table 2 represent a specific training configuration that can be calibrated based on robot mass, number of joints, terrain scale, and target motion speed. When adjusting the reward parameters, constraints such as velocity tracking, posture stability, joint smoothing, foot contact, and abnormal termination still work together to train the control strategy.

[0101] The following describes this embodiment in conjunction with application scenarios: The complete training process includes: Initialize the humanoid robot model, complex terrain, depth camera model, contact sensors, reward function, visual degradation state, round reward window, and network parameters. Read ontology observations and depth maps in multiple parallel simulation environments. Generate a preprocessed depth map by applying Gaussian depth noise and continuous multi-frame structured depth failures based on the current visual noise intensity. Maintain a depth history cache. Generate ontology history latent features, channel scaling parameters, and channel bias parameters. Generate visual latent features. Output joint target positions and control the robot's interaction with the complex terrain. Generate trajectory data based on the interaction results and update network parameters. Write the rewards for completed rounds into the round reward window and recalculate the visual noise intensity, Gaussian depth noise standard deviation, and structured depth failure trigger probability. Repeat the above process until the control strategy reaches the preset training termination condition.

[0102] A closed-loop relationship is formed between structured depth failure, visual noise curriculum, ontology-aware modulation, and temporal fusion. Structured depth failure provides visual anomalies with spatial continuity and temporal persistence; the visual noise curriculum adjusts the intensity of visual anomalies based on training performance; current ontology observations and latent features from ontology history directly participate in depth feature formation; gated cyclic temporal fusion of multi-frame visual features; and the rewards generated by control actions are then used for the next visual noise curriculum update.

[0103] The deployment process includes: After training, the parameters of the ontology history encoding network, modulation parameter generation network, multi-layer convolutional network, gated recurrent temporal fusion, and control policy network are saved. During deployment, Gaussian depth noise and structured depth invalidation from the training phase are not artificially introduced. The depth map thread performs depth map preprocessing and maintains the depth history cache at 10Hz, while the control thread generates latent ontology features and modulation parameters, extracts visual features, and outputs joint target positions at 50Hz. When a valid depth map is first obtained, all cache positions are filled with the current preprocessed depth map; subsequent cache updates are performed at 10Hz.

[0104] When the real depth map experiences ranging noise, local holes, occlusion, consecutive frame drops, or short-term freezes, abnormal or duplicate depth frames may appear in the depth history cache. Since the training phase already includes corresponding visual degradation, ontology-aware modulation and gated cyclic temporal fusion can combine the current ontology state, historical motion trends, and historical depth information to form the control strategy input and output the joint target position.

[0105] Ablation validation includes: Under stress testing conditions with a Gaussian depth noise standard deviation of 0.1 and a structured depth failure trigger probability of 1, ablation validation was performed on the complete method, the visual noise-free course, the gateless cyclic temporal fusion, and the proprioceptive modulation-free schemes. The test results are shown in the table below.

[0106] Table 3: Results of gap topography ablation:

[0107] Table 4: Results of slope topographic ablation:

[0108] Table 5: Results of step-like terrain ablation:

[0109] Higher SR and MXD values ​​indicate greater throughput, while lower BCR and FSSR values ​​indicate fewer collisions or tripping incidents. The data above show that removing visual noise lessons, gated cyclic temporal fusion, or proprioceptive modulation degrades success rate, maximum distance traveled, collision rate, or tripping rate in at least some terrains, demonstrating the synergistic effect of various technical elements under visual degradation conditions.

[0110] For the purpose of simplicity, the method steps disclosed in the above embodiments are described as a series of actions. However, those skilled in the art should understand that the embodiments of the present invention are not limited to the described order of actions, because according to the embodiments of the present invention, some steps can be performed in other orders or simultaneously. Furthermore, those skilled in the art should also understand that the embodiments described in the specification are all preferred embodiments, and the actions involved are not necessarily essential to the embodiments of the present invention.

[0111] like Figure 2 As shown, the present invention also provides a humanoid robot complex terrain control system for vision degradation, comprising: The depth map preprocessing module 201 is configured to acquire a depth map and perform invalid value replacement, distance clipping and normalization on the depth map to obtain a preprocessed depth map. The visual degradation training module 202 is configured to calculate the reward fluctuation from the round reward window during the training phase, update the visual noise intensity according to the reward fluctuation, map the visual noise intensity to the standard deviation of Gaussian depth noise and the structured depth failure trigger probability, apply Gaussian depth noise and continuous multi-frame structured depth failure triggered according to the structured depth failure trigger probability to the preprocessed depth map, and obtain the degradation depth map. The depth history cache module 203 is configured to write the degraded depth map into the depth history cache according to the depth map update time. The ontology history encoding and modulation parameter generation module 204 is configured to encode ontology history observations to obtain ontology history latent features, and input the current ontology observations and ontology history latent features into the modulation parameter generation network to obtain channel scaling parameters and channel bias parameters. The visual feature extraction module 205 is configured to extract convolutional features from the depth map of each frame in the depth history cache, modulate the convolutional features based on the channel scaling parameter and the channel bias parameter, and perform gated cyclic temporal fusion on the modulated convolutional features of each frame to obtain visual latent features. The motion control module 206 is configured to input the current body observation, the body's historical latent features and visual latent features into the control strategy network, output the joint target position, and control the humanoid robot to move in complex terrain based on the joint target position.

[0112] It is worth noting that although only some basic functional modules are disclosed in the embodiments of this invention, it does not mean that the composition of this system is limited to the above-mentioned basic functional modules. On the contrary, based on the above-mentioned basic functional modules, those skilled in the art can arbitrarily add one or more functional modules in combination with existing technology to form an infinite number of embodiments or technical solutions. That is to say, this system is open rather than closed. The fact that this embodiment only discloses a few basic functional modules should not be considered as the scope of protection of this invention being limited to the disclosed basic functional modules. At the same time, for the convenience of description, the above devices are described separately according to their functions as various units and modules. Of course, in implementing this invention, the functions of each unit and module can be implemented in one or more software and / or hardware.

[0113] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for controlling complex terrain in humanoid robots with visual degradation, characterized in that, include: Obtain the depth map, and perform invalid value replacement, distance clipping, and normalization on the depth map to obtain a preprocessed depth map; During the training phase, the reward fluctuation is calculated from the round reward window, the visual noise intensity is updated according to the reward fluctuation, the visual noise intensity is mapped to the standard deviation of Gaussian depth noise and the structured depth failure trigger probability, Gaussian depth noise and continuous multi-frame structured depth failure triggered according to the structured depth failure trigger probability are applied to the preprocessed depth map to obtain the degraded depth map. Write the degraded depth map to the depth history cache according to the depth map update time; Encode the ontology historical observations to obtain ontology historical latent features. Input the current ontology observations and ontology historical latent features into the modulation parameter generation network to obtain channel scaling parameters and channel bias parameters. Convolutional features are extracted from the depth map of each frame in the depth history cache. The convolutional features are modulated based on the channel scaling parameter and the channel bias parameter. Gated cyclic temporal fusion is performed on the modulated convolutional features of each frame to obtain the visual latent features. The current ontology observation, ontology historical latent features, and visual latent features are input into the control policy network, which outputs the joint target position and controls the humanoid robot to move in complex terrain based on the joint target position.

2. The method for controlling complex terrain in a humanoid robot with visual degradation according to claim 1, characterized in that, Write the degraded depth map to the depth history cache according to the depth map update time, and further include: At the moment the depth map is updated, the oldest depth map in the depth history cache is deleted, and the current degraded depth map is written to the end of the depth history cache; When the training environment is reset, the current degraded depth map is used to fill all cache locations of the depth history cache. When a valid depth map is first obtained during the deployment phase, a preprocessed depth map is used to fill all cache locations of the depth history cache. In subsequent depth map updates, the oldest preprocessed depth map is deleted and the newly obtained preprocessed depth map is written.

3. The method for controlling complex terrain in a humanoid robot with visual degradation according to claim 1, characterized in that, A sustained multi-frame structured depth failure triggered by a structured depth failure trigger probability includes at least one of the following: Blocky holes, bottom occlusion, central foothold area occlusion, stripe loss, whole frame loss, consecutive multi-frame loss, image freeze, and far-distance depth pollution; For blocky holes, lower occlusion, central foothold area occlusion, stripe loss, whole frame loss and consecutive multi-frame loss, generate failure areas in the preprocessed depth map and set the pixel values ​​of the failure areas to zero. Freeze the frame, save the preprocessed depth map at the trigger moment, and replace it with a new depth map within the frame; For deep contamination at long distances, Gaussian perturbation is added to pixels outside a preset distance and the pixels are set to zero with a preset probability; The depth map that has been processed through a preset number of depth map update cycles is written to the depth history cache.

4. The method for controlling complex terrain in a humanoid robot with visual degradation according to claim 1, characterized in that, The reward fluctuation is calculated from the round reward window, and the visual noise intensity is updated according to the reward fluctuation. Further steps include: Calculate the mean and standard deviation of the reward window for each round. Divide the standard deviation of the reward by the sum of the absolute value of the mean reward and a preset positive number to obtain the coefficient of variation of the reward. Based on a preset monotonically decreasing mapping relationship, the coefficient of variation of the reward is converted into the original visual noise intensity; The updated visual noise intensity is synthesized by weighting the previous visual noise intensity and the original visual noise intensity using an exponential moving average method. The standard deviation of Gaussian depth noise and the probability of structured depth failure are calculated linearly between a preset lower limit and a preset upper limit based on the updated visual noise intensity.

5. A method for controlling complex terrain in a humanoid robot with visual degradation according to claim 1, characterized in that, Encoding ontological history observations yields latent ontological history features, which further include: The root node angular velocity, gravity direction projection in the fuselage coordinate system, velocity command, joint position, joint velocity, action of the previous control cycle, and sine and cosine values ​​of the left and right leg gait phases are used to form the body observation at each control moment. Encode the continuous preset number of ontology observation frames in chronological order and output the ontology's historical potential features; The current ontology observations and historical potential features of the ontology are concatenated according to feature dimensions to generate modulation parameters.

6. A method for controlling complex terrain in a humanoid robot with visual degradation according to claim 1, characterized in that, Modulating convolutional features based on channel scaling parameters and channel bias parameters further includes: Each frame of depth map in the depth history cache is convolved sequentially to obtain the convolutional features of the depth map output by each convolutional layer. Generate channel scaling parameters and channel bias parameters for each convolutional layer; Based on the channel scaling parameters and channel bias parameters, the convolutional features of the depth map of each convolutional layer are scaled and biased, and the channel scaling parameters and channel bias parameters are shared among the depth maps of each frame in the depth history cache. The modulated convolutional features output from the final convolutional layer are then subjected to gated cyclic temporal fusion according to the depth map acquisition order.

7. A method for controlling complex terrain in a humanoid robot with visual degradation according to claim 1, characterized in that, The current ontology observation, historical latent features of the ontology, and visual latent features are input into the control policy network, which outputs the joint target location, further including: Output normalized joint motion, perform numerical scaling on the normalized joint motion based on a preset motion scaling ratio, and convert it into the joint target position; The ontological observations, rewards, and termination signals generated by the interaction of humanoid robots with complex terrain are combined into trajectory data; During the training phase, the control strategy is updated based on trajectory data, and the joint target position is output based on the updated control strategy.

8. A method for controlling complex terrain in a humanoid robot with visual degradation according to claim 3, characterized in that, The central landing area is further shielded, including: The rectangle size is determined based on a first preset ratio range for the depth map height and a second preset ratio range for the depth map width. The center of the rectangle is randomly determined within the third preset ratio range of the depth map height and the fourth preset ratio range of the depth map width; Crop the rectangular edges that extend beyond the image boundaries; Set the pixels within the rectangle to zero, and repeat the process of setting the pixels within the rectangle to zero on each newly acquired preprocessed depth map within a preset number of depth map update cycles.

9. A method for controlling complex terrain in a humanoid robot with visual degradation according to claim 4, characterized in that, The updated visual noise intensity is synthesized by weighting the previous visual noise intensity and the original visual noise intensity using an exponential moving average method, further including: Calculate the mean return, standard deviation of return, and coefficient of variation of return within a reward window of a preset number of completed training rounds; When the absolute value of the average return is lower than a preset positive number, the absolute value of the average return is replaced with the preset positive number; The visual noise intensity of the previous round and the original visual noise intensity are weighted and summed using a preset exponential moving average coefficient. The weighted summation result is cropped to a preset visual noise intensity range and mapped to the Gaussian depth noise standard deviation and the structured depth failure trigger probability.

10. A control system for humanoid robots in complex terrain designed to address visual degradation, characterized in that, include: The depth map preprocessing module is configured to acquire a depth map and perform invalid value replacement, distance clipping, and normalization on the depth map to obtain a preprocessed depth map. The visual degradation training module is configured to calculate the reward fluctuation from the round reward window during the training phase, update the visual noise intensity according to the reward fluctuation, map the visual noise intensity to the standard deviation of Gaussian depth noise and the structured depth failure trigger probability, apply Gaussian depth noise and continuous multi-frame structured depth failure triggered according to the structured depth failure trigger probability to the preprocessed depth map, and obtain the degradation depth map. The depth history cache module is configured to write degraded depth maps into the depth history cache according to the depth map update time. The ontology history encoding and modulation parameter generation module is configured to encode ontology history observations to obtain ontology history latent features, and input the current ontology observations and ontology history latent features into the modulation parameter generation network to obtain channel scaling parameters and channel bias parameters. The visual feature extraction module is configured to extract convolutional features from the depth map of each frame in the depth history cache, modulate the convolutional features based on the channel scaling parameter and the channel bias parameter, and perform gated cyclic temporal fusion on the modulated convolutional features of each frame to obtain the visual latent features. The motion control module is configured to input the current ontology observation, ontology historical latent features, and visual latent features into the control strategy network, output the joint target position, and control the humanoid robot to move in complex terrain based on the joint target position.

Images