Robot autonomous navigation method, device, robot and computer-readable storage medium

Abstract

This application relates to the field of robot control technology, and in particular to a robot autonomous navigation method, device, robot, and computer-readable storage medium. The method includes: acquiring the robot's real-time body state information, current position, target position, and depth perception information of the external environment; encoding the depth perception information, and concatenating the encoded depth perception information with the real-time body state information to obtain a unified feature representing the external environment and the robot's own state; inputting the unified feature and the target position into a preset navigation decision model to generate the robot's initial motion speed; calculating the straight-line distance between the current position and the target position in real time, and adjusting the initial motion speed based on the straight-line distance to obtain a stable desired speed parameter; generating and controlling the execution of an execution command to drive the robot's motion based on the stable desired speed parameter and the real-time body state information, navigating to the target position, and realizing autonomous navigation of the robot in a map-free scenario.

Description

Technical Field

[0001] This application relates to the field of robot control technology, and in particular to a robot autonomous navigation method, device, robot, and computer-readable storage medium. Background Technology

[0002] In real-world scenarios such as search and rescue, industrial inspection, and outdoor exploration, robots often face working environments with no pre-set maps, complex terrain, and dynamic changes. For example, collapsed terrain during rubble search and rescue, or ravines and rubble areas during outdoor inspection. These scenarios place core demands on the robot's autonomous navigation and motion control capabilities—the ability to accurately perceive the terrain and move stably in the absence of prior environmental information.

[0003] Currently, the mainstream solutions in the industry mainly use SLAM (Simultaneous Localization and Mapping) technology to build environmental maps in real time, and then use the maps as a basis for positioning and navigation.

[0004] However, this approach has significant limitations in complex 3D terrain without maps: SLAM technology requires a large amount of computing resources to process environmental data, and when the terrain changes dynamically (such as secondary collapse of ruins or temporary obstacles appearing in the inspection area), the map is prone to failure, leading to navigation interruption.

[0005] The aforementioned technical bottlenecks severely restrict the practical application of robots—in rubble search and rescue operations, navigation failures prevent robots from quickly approaching trapped individuals. Therefore, it is urgent to solve the problem of efficient autonomous navigation for robots in the absence of maps and privileged sensors.

[0006] The above content is only used to help understand the technical solution of this application and does not represent an admission that the above content is prior art. Summary of the Invention

[0007] The main purpose of this application is to provide a robot autonomous navigation method, device, robot, and computer-readable storage medium, which aims to achieve efficient autonomous navigation of robots in map-free scenarios.

[0008] To achieve the above objectives, this application provides a robot autonomous navigation method, which includes: Acquire the robot's real-time body state information, current position, target position, and depth perception information of the external environment; The depth perception information is encoded, and the encoded depth perception information is concatenated with the real-time body state information to obtain a unified feature representing the external environment and the robot's own state. Input the unified features and target position into the preset navigation decision model to generate the robot's initial movement speed; The straight-line distance between the current position and the target position is calculated in real time, and the initial movement speed is adjusted based on the straight-line distance to obtain a stable desired speed parameter; Based on the stable expected velocity parameters and real-time body state information, the system generates execution commands to drive the robot's movement and controls the robot to execute the commands and navigate to the target position.

[0009] In addition, to achieve the above objectives, this application also provides a robot autonomous navigation device, which includes: The high-level navigation strategy module is used to acquire the robot's real-time body state information, current position, target position, and depth perception information of the external environment; The high-level navigation strategy module is also used to encode depth perception information, and then splice the encoded depth perception information with real-time body state information to obtain a unified feature representing the external environment and the robot's own state. The high-level navigation strategy module is also used to input unified features and target positions into a preset navigation decision model to generate the robot's initial motion speed; The high-level navigation strategy module is also used to calculate the straight-line distance between the current position and the target position in real time, and adjust the initial movement speed based on the straight-line distance to obtain a stable desired speed parameter; The low-level motion strategy module is used to generate execution commands to drive the robot's motion based on the stable desired velocity parameters and real-time body state information, and to control the robot to execute the execution commands and navigate to the target position.

[0010] In addition, to achieve the above objectives, this application also provides a robot, which includes a memory, a processor, and a robot autonomous navigation program stored in the memory and executable on the processor. When the robot autonomous navigation program is executed by the processor, it implements the steps of the above-described robot autonomous navigation method.

[0011] In addition, to achieve the above objectives, this application also provides a computer-readable storage medium storing a robot autonomous navigation program, which, when executed by a processor, implements the steps of the robot autonomous navigation method described above.

[0012] This application provides a method for autonomous robot navigation. By acquiring the robot's real-time body state information, current position, target position, and depth perception information of the external environment, the robot can continuously monitor its own state, location, navigation target, and surrounding terrain data, providing a data foundation for subsequent navigation. The depth perception information is encoded and then concatenated with the real-time body state information to obtain unified features representing the external environment and the robot's own state. This integrates fragmented environmental and self-state information into unified and effective features, eliminating information fragmentation and providing structured input data for the subsequent navigation decision model. The unified features and target position are then input into a pre-defined navigation decision model to generate robot navigation data. The system first calculates the initial motion speed based on environmental and self-state characteristics and navigation target, providing a basic motion speed reference for the robot. Then, it calculates the straight-line distance between the current position and the target position in real time, and adjusts the initial motion speed based on this distance to obtain stable expected speed parameters. This optimizes the initial motion speed according to the real-time distance between the robot and the target position, resulting in stable speed parameters suitable for the given distance scenario. Finally, based on the stable expected speed parameters and real-time body state information, it generates execution commands to drive the robot's movement and controls the robot to execute these commands, navigating to the target position. This process transforms stable speed parameters into robot joint motion commands and controls their execution, thereby enabling autonomous navigation of the robot in map-less scenarios. Attached Figure Description

[0013] Figure 1 This is a flowchart illustrating the first embodiment of the robot autonomous navigation method of this application; Figure 2 This is a schematic diagram of the navigation algorithm framework involved in this application; Figure 3 This is a schematic diagram of the motion control framework based on reinforcement learning involved in this application; Figure 4 This is a schematic diagram of the end-to-end navigation strategy framework based on depth maps involved in this application; Figure 5 This is a schematic diagram of the structure of the robot autonomous navigation device involved in the embodiments of this application; Figure 6 This is a schematic diagram of the robot involved in the embodiments of this application.

[0014] The realization of the purpose, functional features and advantages of this application will be further explained in conjunction with the embodiments and with reference to the accompanying drawings. Detailed Implementation

[0015] This application provides a robot autonomous navigation method, referring to... Figure 1 As shown, Figure 1This is a flowchart illustrating the first embodiment of the robot autonomous navigation method of this application.

[0016] Exemplary embodiments will now be described in detail, examples of which are illustrated in the accompanying drawings. When the following description relates to the drawings, unless otherwise indicated, the same numbers in different drawings represent the same or similar elements. The embodiments described in the following exemplary embodiments do not represent all embodiments consistent with this application.

[0017] In scenarios such as rubble search and rescue and outdoor exploration, robots need to navigate autonomously in environments with complex terrain and no pre-set maps. Currently, the mainstream solution in the industry relies on SLAM technology to build maps in real time for positioning and navigation. However, this solution consumes a lot of computing resources, and when the terrain changes dynamically (such as rubble collapse or the appearance of temporary obstacles), the map is prone to failure, leading to navigation interruption, which is difficult to meet the needs of actual operations.

[0018] This application provides a robot autonomous navigation method, apparatus, robot, and computer-readable storage medium to achieve efficient autonomous navigation of robots in map-free scenarios.

[0019] It should be noted that the executing entity in this embodiment can be a computing service device with data processing, network communication, and program execution functions, such as a device capable of performing the above functions, like a robot. The following description uses a robot as an example to illustrate this embodiment and the subsequent embodiments.

[0020] The robot autonomous navigation method of this application includes the following implementation steps S10 to S50.

[0021] Step S10: Obtain the robot's real-time body state information, current position, target position, and depth perception information of the external environment; Optionally, as an example, the target application scenarios can be unstructured complex scenarios without preset maps or dynamic terrain scenarios lacking external positioning support, such as rubble search and rescue, outdoor exploration, and post-disaster reconstruction sites, which require robots to autonomously cope with sudden terrain changes and lack prior environmental information, covering the core requirements for the robot's "mapless perception-planning-control" full-process collaboration.

[0022] Optionally, the hierarchical navigation framework refers to a modular architecture that includes high-level navigation strategies and low-level motion strategies; the high-level navigation strategy refers to the module that undertakes global planning functions; and the low-level motion strategy refers to the module that undertakes local control functions.

[0023] Optionally, such as Figure 2 As shown, Figure 2This is a schematic diagram of the navigation algorithm framework involved in this application. This framework is used to achieve autonomous navigation of a robot in complex scenes. The framework includes an encoder, a navigation strategy, a motion strategy, and the robot body as the execution end; wherein, o t nav (Navigation Observation) represents navigation observation information, encompassing the robot's state data related to the target position; the depth map represents the external environmental terrain information collected by the robot; the encoder is used to process the depth map and output environmental feature information; a t nav (Navigation Action) represents the speed command and is the output of the navigation strategy; t loco (Locomotion Observation) represents motion observation information, including the robot's own body state data; a t loco (LocomotionAction, motion action command) represents motor commands and is the output of the motion strategy. Navigation observation information is directly input into the navigation strategy. Simultaneously, the depth map is input into the encoder for processing, and the encoder's output is synchronously input into the navigation strategy. The navigation strategy generates speed commands based on the above inputs and sends these speed commands to both the navigation strategy itself and the motion strategy. Motion observation information is input into the motion strategy, which combines the speed commands to generate motor commands. Finally, the motor commands are sent to the robot body to control the robot to complete the action. This framework achieves decoupling between global planning and local execution through a layered design of "high-level navigation strategy (responsible for global planning) + low-level motion strategy (responsible for local control)," ensuring the robot's autonomous navigation capability in complex environments.

[0024] Optionally, real-time body state information refers to a set of data reflecting the robot's own operating state; current position refers to the robot's spatial coordinates in the current environment; target position refers to the coordinates of the endpoint that the robot needs to reach; depth perception information refers to depth data representing the terrain of the external environment.

[0025] Optionally, combined Figure 2The navigation algorithm framework shown involves the robot acquiring the current joint position and angular velocity through joint encoders, and acquiring the body angular velocity, gravity direction vector in the body coordinate system, and body attitude through an inertial measurement unit (IMU). These data are integrated to obtain real-time body state information. The current position is calculated based on the body attitude and motion trajectory in the real-time body state information, while the endpoint coordinates sent by the external navigation system are received as the target position. The terrain around the current position (e.g., 1 meter forward and 0.5 meters to the left and right) is scanned and rendered by an airborne forward depth camera, and depth perception information representing the terrain of this range is output.

[0026] Optionally, the preset range for depth perception information acquisition is based on the robot's center of mass as the origin of the coordinate system. The X-axis (directly in front) covers -0.4 meters to 1.0 meters, and the Y-axis (left and right directions) covers -0.5 meters to 0.5 meters. The sampling resolution is 0.1 meters. This range is determined through experimental testing and is the optimal observation interval for adapting the robot's body perception and terrain interaction. The Scandot (high-precision local terrain point array) in the simulation stage is only provided during training and cannot be obtained in the deployment stage. It is necessary to learn the equivalent terrain features through depth map distillation.

[0027] This embodiment provides comprehensive and high-quality basic data for robot autonomous navigation through a clear hardware acquisition method and scenario-based processing logic, ensuring the accuracy of subsequent navigation decisions and motion control, and improving the robot's adaptability in complex scenarios.

[0028] Step S20: Encode the depth perception information, and then concatenate the encoded depth perception information with the real-time body state information to obtain a unified feature representing the external environment and the robot's own state. Optionally, encoding refers to the process of converting raw perception data into low-dimensional feature vectors; unified features refer to a comprehensive feature vector that integrates the external environment and the robot's own state, which is used as input for subsequent navigation decision models.

[0029] Optionally, the depth perception information is encoded using methods including but not limited to lightweight visual coding networks to extract environmental geometric features (such as obstacle positions and terrain slopes) and convert them into feature vectors of a preset dimension; real-time body state information is obtained, including but not limited to joint angular velocity and body angular velocity; the feature vectors and real-time body state information are concatenated according to the feature dimension to obtain a unified feature representing the external environment and the robot's own state.

[0030] Optionally, the encoding process employs a two-stage approach: first, a terrain-aware encoder is trained using high-fidelity terrain scanning information (Scandot, a structured ontology height-aware representation composed of high-precision local terrain points provided during the training phase), and then the geometric features of this encoder are distilled into a lightweight encoding module. During stitching, the real-time ontology state information is first standardized (e.g., the data is normalized to a preset range), and then fused with the encoded depth-aware information to form a unified feature. The lightweight encoding module includes, but is not limited to, a MobileViT-GRU (Mobile Vision Transformer-Gated Recurrent Unit) structure, where MobileViT captures spatial geometric features and GRU models temporal dynamics (i.e., the changing patterns of historical perception data).

[0031] This embodiment achieves effective integration of the external environment and the robot's own state through explicit coding and splicing logic, providing comprehensive feature support for subsequent navigation decisions and improving the robot's perception and decision-making accuracy in complex scenarios.

[0032] Step S30: Input the unified features and target position into the preset navigation decision model to generate the robot's initial movement speed; Optionally, the preset navigation decision model refers to a model that can output motion speed after training, including but not limited to a network structure based on reinforcement learning (RL, Reinforcement Learning, a machine learning method that learns the optimal policy by interacting with the environment and based on reward signals); the initial motion speed refers to the robot motion speed parameters generated by the model without distance adjustment, including planar linear velocity and angular velocity.

[0033] Optionally, a unified feature is acquired, which includes, but is not limited to, the fusion result of encoded depth perception information and real-time ontology state information; the target position of the robot is determined, which is the coordinate of the endpoint that the robot needs to reach; the unified feature and the target position are input into a preset navigation decision model, and the model learns the mapping relationship from the input to the initial motion velocity through training methods including but not limited to proximal policy optimization (PPO, a stable and efficient policy gradient algorithm), and outputs the planar linear velocity and angular velocity of the robot in the body coordinate system as the initial motion velocity.

[0034] Optionally, during the training of the preset navigation decision model, the core optimization direction is first clarified, including but not limited to shortening navigation time and reducing motion attitude fluctuations; then, auxiliary training objectives are introduced to improve the model's generalization, including but not limited to loss terms based on temporal state prediction (used to measure the accuracy of the predicted next moment's ontology state compared with the actual collected state) and task-related reward and penalty terms (used to guide the model to prioritize learning safe and efficient motion patterns, such as speed tracking rewards and collision avoidance penalties); the model parameters are updated through multi-objective collaborative optimization, enabling the model to adaptively output an initial motion speed that matches the current terrain complexity based on the unified features of the input and the target position, for example, outputting a higher linear speed on flat terrain and a lower linear speed while retaining turning margin on terrain with dense obstacles.

[0035] Optionally, in the planar linear velocity output by the navigation decision model, the X-axis (forward / backward) value ranges from -0.5 m / s to 0.8 m / s, the Y-axis (left / right translation) value ranges from -0.5 m / s to 0.5 m / s, and the Z-axis angular velocity ranges from -1 rad / s to 1 rad / s. The final velocity command will be scaled to this range to ensure motion safety.

[0036] This embodiment achieves an efficient conversion from environmental perception to velocity generation through a clear model input and output logic, providing reasonable initial parameters for subsequent velocity adjustment and motion control, and improving the robot's navigation decision-making efficiency in complex scenarios.

[0037] Step S40: Calculate the straight-line distance between the current position and the target position in real time, and adjust the initial motion speed based on the straight-line distance to obtain a stable desired speed parameter; Optionally, the straight-line distance refers to the spatial straight-line distance between the current position and the target position, used to determine the distance between the robot and the target; the preliminary motion speed refers to the unadjusted speed parameters output by the navigation decision model; the stable expected speed parameters refer to the speed parameters adapted to the current navigation scenario after distance adjustment, taking into account both navigation efficiency and motion stability.

[0038] Optionally, the robot's current position and target position are acquired in real time, and the straight-line distance between them is calculated by means including but not limited to estimation based on the robot's state (such as motion trajectory and angular velocity); the initial motion speed output by the navigation decision model is acquired, which includes planar linear velocity and angular velocity; the initial motion speed is adjusted based on the straight-line distance, for example, increasing the speed to optimize navigation efficiency when the straight-line distance is far (such as greater than 5 meters), and decreasing the speed to ensure motion stability when the straight-line distance is close (such as less than 1 meter), so as to obtain stable desired speed parameters.

[0039] Optionally, when adjusting the initial movement speed, first set distance judgment rules adapted to the navigation scenario, including but not limited to judgment criteria that distinguish between different distance ranges of far, medium and near (e.g., far distance: greater than 5 meters, medium distance: 1-5 meters, near distance: less than 1 meter); select the corresponding adjustment logic according to the range to which the straight-line distance belongs, for example, increase the initial movement speed by a preset ratio (e.g., 1.2 times) in the far distance range, and reduce the initial movement speed through a smooth attenuation algorithm (e.g., linear attenuation based on distance) in the near distance range; dynamically correct the speed during the adjustment process by combining terrain perception information, for example, appropriately reduce the speed adjustment range in complex terrain (e.g., change the original adjustment ratio of 1.2 times to 1.1 times) to ensure that the stable expected speed parameters are adapted to the terrain conditions and distance requirements.

[0040] This embodiment enables the robot to obtain reasonable speed parameters in different navigation stages by dynamically adjusting the speed based on distance, balancing navigation efficiency and motion stability, providing reliable input for subsequent motion command generation, and improving the robot's navigation adaptability in complex scenarios.

[0041] Step S50: Based on the stable desired velocity parameters and real-time body state information, generate execution commands to drive the robot's movement, and control the robot to execute the execution commands to navigate to the target position; Optionally, the execution command refers to the target motion trajectory that drives the robot's joint movements, including the angle change patterns and time points of each joint.

[0042] Optionally, the system acquires stable desired velocity parameters and real-time body state information; based on the joint positions and body postures in the stable desired velocity parameters and real-time body state information, it calculates the deviation between the joint positions and the ideal joint positions (joint postures adapted to stable desired velocity parameters), and outputs the deviation adjustment amount through a preset deviation adjustment algorithm including but not limited to the proportional-derivative adjustment algorithm (PD algorithm, a control algorithm that adjusts the deviation through proportional and derivative terms); based on the motion requirements of the stable desired velocity parameters, it calls preset basic gait joint trajectories (such as standard joint motion trajectories corresponding to walking and climbing), and superimposes the deviation adjustment amount and the basic gait joint trajectory according to the joint dimension to generate the target motion trajectory of each joint as an execution command; the execution command is sent to the joint drive module (the hardware module that controls the joint motion) to control each joint to move according to the target motion trajectory until the target position is reached.

[0043] Optionally, actual joint torque Calculated using a proportional-derivative (PD) controller:

[0044] in, and These are the proportional and differential gain parameters, respectively.

[0045] Optionally, when generating execution commands, the basic gait joint trajectory is first dynamically adjusted based on terrain perception information. For example, the amplitude of the hip joint trajectory is increased on sloping terrain to adapt to the slope. When calculating deviations, the joint angular velocity is combined for dynamic correction to avoid joint movement fluctuations. After the execution command is issued, the joint movement feedback (such as the actual joint position collected by the joint encoder) is monitored in real time. If the movement deviation exceeds the preset range (such as greater than 5°), the deviation adjustment amount is recalculated and the execution command is updated. Among them, the basic gait joint trajectory can be preset with multiple adaptation schemes according to different terrains (such as stairs, gravel ground), and automatically matched according to depth perception information when called.

[0046] This embodiment generates precise execution commands by combining speed requirements with the robot's state, enabling stable motion control of the robot in complex terrain, ensuring the smoothness and adaptability of the movements during navigation, and providing reliable execution support for the robot to reach the target location.

[0047] This embodiment acquires the robot's real-time body state information, current position, target position, and depth perception information of the external environment; encodes the depth perception information, and concatenates the encoded depth perception information with the real-time body state information to obtain a unified feature representing the external environment and the robot's own state; inputs the unified feature and the target position into a preset navigation decision model to generate the robot's initial movement speed; calculates the straight-line distance between the current position and the target position in real time, and adjusts the initial movement speed based on this straight-line distance to obtain a stable expected speed parameter; based on the stable expected speed parameter and the real-time body state information, generates and controls the execution of the command to drive the robot's movement, navigating to the target position, thus achieving autonomous navigation of the robot in a map-free scenario.

[0048] Furthermore, based on the above, a second embodiment of the robot autonomous navigation method of this embodiment is proposed. In some feasible embodiments, the system includes a target detection model, and the above step S10 includes the following implementation steps B201-B203.

[0049] Step B201: Collect the robot's current joint position, joint angular velocity, body angular velocity, gravity direction vector in the body coordinate system, and body posture, and integrate them to obtain real-time body state information; Step B202: Determine the robot's current position based on real-time body state information, and obtain the coordinates of the endpoint that the robot needs to reach from the external navigation system, as the target position; Step B203: Scan and render the terrain within a preset range around the current location, and output depth perception information representing the terrain within the preset range.

[0050] Optionally, RL (Reinforcement Learning): a machine learning method that learns optimal policies by interacting with the environment and based on reward signals, used to train motion control and navigation policies; PPO (Proximal Policy Optimization): a stable and efficient policy gradient algorithm used to optimize policy updates in reinforcement learning; GRU (Gated Recurrent Unit): a recurrent neural network structure used to process temporal data and extract representations of historical perceived states; MobileViT (Mobile Vision Transformer): a lightweight vision transformer model used to extract semantic representations from depth maps; Sim-to-Real: a method for transferring models trained in a simulation environment to run in a real robot system; Scandot: a structured ontology height-aware representation composed of high-precision local terrain lattices provided during the training phase, which is not available during deployment and requires learning through depth map distillation; Warp Renderer: a fast rendering module based on NVIDIA Warp, used in Isaac... The Gym simulation platform generates realistic depth maps with high concurrency, accelerating the visual training process; Latent Representation: a low-dimensional feature vector extracted from perceptual information by the encoder, used to represent the environmental state for use by the policy network; Contrastive Learning: an unsupervised learning method used to enhance the consistency and discriminativeness of representations in the time dimension, improving sample efficiency in motion policy training; Inverse Dynamics: a model that predicts the actions required for a change in a state, used to improve the decision relevance of representations in policy learning; IMU (Inertial Measurement Unit): a sensor used to measure the motion state of an object, such as angular velocity and acceleration.

[0051] Optionally, this embodiment is applied to unstructured complex scenarios without preset maps or dynamic terrain scenarios lacking external positioning support, such as rubble search and rescue, outdoor exploration, etc., to achieve autonomous robot navigation by combining a navigation algorithm framework (including encoder, navigation strategy, motion strategy, and robot body). The robot collects the current joint position and joint angular velocity through the joint encoder, and collects the body angular velocity, gravity direction vector in the body coordinate system, and body posture through the IMU. The robot integrates these data to obtain real-time body state information, which includes, but is not limited to, the aforementioned joint and body-related data reflecting the robot's own operating state; the robot calculates the current position (including but not limited to the robot's spatial coordinates in the current environment) based on the body posture and motion trajectory in the real-time body state information, and simultaneously receives the destination coordinates sent by the external navigation system as the target position (including but not limited to the destination coordinates that the robot needs to reach); the robot uses an onboard forward depth camera to scan a preset range around the current position (the preset range is based on the robot's position). The robot's center of mass is the origin of the coordinate system. The X-axis covers a range of -0.4 meters to 1.0 meters in front and the Y-axis covers a range of -0.5 meters to 0.5 meters in the left and right directions, with a sampling resolution of 0.1 meters. Terrain scanning and rendering are used to obtain a depth map, which serves as the main terrain perception input during deployment. In the simulation phase, domain randomization processing, including but not limited to random camera pose perturbation and pixel-by-pixel Gaussian noise, can be applied to it. In real-world scenarios, data processing methods, including but not limited to global mean interpolation of zero-value pixels and exponential moving average (with a smoothing coefficient set to 0.01) temporal smoothing, can be used. The final output is depth perception information representing the terrain within a preset range (including but not limited to depth data representing the terrain of the external environment).

[0052] Optionally, such as Figure 3 As shown, Figure 3 This is a schematic diagram of the motion control framework based on reinforcement learning involved in this application, which integrates reinforcement learning, self-supervised representation learning, and deep perception feature distillation, wherein a t loco Motion control actions refer to the commands that drive the robot's joints to move; MSEloss (mean squared error loss) is a loss function used to measure the difference between predicted and true values; inverse dynamics: inverse dynamics modeling, a model that predicts actions using adjacent states, used to enhance the correlation between features and control; contrastive learning: an unsupervised learning method that improves the consistency of feature representations by aligning positive samples and distinguishing negative samples; contrastive loss, a loss function used in contrastive learning to optimize feature representations; z t+1 propThe forward-predicted potential representation for the next time step is generated by the forward transition model; the forward transition model is a model that predicts the next time step state based on the current state and action; the inverse reconstruction model is a model that predicts actions based on neighboring states (i.e., the inverse dynamics model); the encoder is a module that converts raw perceptual information (such as ontology and depth data) into a low-dimensional potential representation; the Actor policy network is a module that outputs the robot's motion actions; the policy gradient is the gradient calculation method used in reinforcement learning to optimize the policy network; the Critic is a value network that evaluates the value of actions; stage 1 policy learning is the stage of training the motion policy using reinforcement learning: o t-9t prop Predicted ontology-aware information; Stage 2: Deep distillation: Deep distillation is the stage that transfers high-fidelity terrain features to a lightweight deep encoder; Reinforcement learning: A machine learning method that learns the optimal policy by interacting with the environment and based on reward signals; Encoder; Supervision: Reference signals used to train the model (such as the output of terrain reconstruction); GRU: Gated Recurrent Unit, a neural network structure for processing temporal data; MobileViT: Mobile Vision Transformer, a lightweight visual feature extraction module; t prop Predicted ontological perception information; Decoder: a module that restores the latent representation to the original perception information; Terrain reconstruction; The process of restoring terrain information through encoder-decoder; O t depth : Depth-sensing information; O t scan Terrain scan information; Z t depth : The latent representation of depth information; Z t prop : The latent representation of the prediction; Z t+1 prop : Predict the potential representation for the next time step; o t-10t-1 propThe framework includes two types of inputs: ontology-aware information and external depth-aware information. Phase 1 (Policy Learning): The ontology-aware information is processed by the encoder into an ontology latent representation, which is input to the Actor (policy network) to output motion control actions. Simultaneously, the Critic is input to evaluate the value of the actions. Both are trained through reinforcement learning via policy gradients. In the self-supervised representation learning module, the forward transition model generates a predicted value for the next latent representation based on the current latent representation and actions, and aligns it with the true latent representation using Contrastive Loss. The inverse reconstruction model (inverse dynamics model) predicts actions based on the latent representations of adjacent time steps, and is optimized using MSELoss (mean squared error loss). Both work together to improve the representation quality of the latent space. Phase 2 (Depth Distillation): High-fidelity terrain scan information (oscam) is used by the terrain reconstruction module (encoder + decoder) to generate a supervision signal. Its geometric features are then distilled into a lightweight depth encoder (MobileViT+GRU). This encoder processes external depth sensing information (odeptht) and outputs a depth latent representation (zdeptht), which is ultimately fused with the ontological latent representation. This allows the system to achieve terrain-aware motion control using only the original depth map. The entire framework, through a two-stage design and multi-module collaboration, achieves efficient integration of perception, feature representation, and motion control.

[0053] In this embodiment, the acquisition of depth perception information is achieved through the IsaacGym simulation platform, with LeggedGym serving as the code library for extending simulation functionality. The number of parallel environments in the GPU parallel depth rendering module does not exceed 1024. This parameter is set based on the GPU memory capacity; the larger the memory capacity, the more parallel environments can be supported. The intrinsic parameters of the airborne forward depth camera are based on the Intel RealSense D435 configuration, with an acquisition frequency of 200Hz. The reinforcement learning neural network strategy acquires camera data at a frequency of only 50Hz. The depth map missing pixel interpolation is processed by averaging the data from the first 10 frames, while temporal smoothing is achieved through sliding window averaging.

[0054] This embodiment provides comprehensive and high-quality basic data for robot autonomous navigation through a clear hardware acquisition method and scenario-based processing logic, ensuring the accuracy of subsequent navigation decisions and motion control, and improving the robot's adaptability in complex scenarios.

[0055] Furthermore, based on the above, in some feasible embodiments, step S20 includes the following implementation steps C301-C302.

[0056] Step C301: Encode the depth perception information, extract obstacle information and terrain slope information from the depth perception information, and map the extracted obstacle information and terrain slope information into feature vectors of a preset dimension; Step C302: Concatenate the feature vectors of the preset dimensions with the joint angular velocities and the body angular velocities according to the feature dimensions to form a unified feature representing the external environment and the robot's own state.

[0057] Optionally, depth perception information: depth data representing the external environment terrain, including obstacle information, terrain slope information, etc.; feature vector: converting the original perception information into a low-dimensional vector for use as input to the subsequent decision model.

[0058] Optionally, this embodiment is applied to unstructured complex scenarios without a preset map. It combines a navigation algorithm framework to achieve autonomous robot navigation. The depth perception information is encoded using a lightweight encoding module including but not limited to MobileViT-GRU. MobileViT captures the spatial geometry of obstacles and terrain slope in the depth map, while GRU models the temporal dynamics of this information. The extracted obstacle information and terrain slope information are mapped into feature vectors of a preset dimension (e.g., 32-dimensional). Data such as joint angular velocity and body angular velocity are obtained. The feature vector is then concatenated with the joint angular velocity and body angular velocity according to the feature dimensions to form a unified feature representing the external environment and the robot's own state.

[0059] Optionally, the encoding process employs a two-stage depth distillation process: first, the terrain-aware encoder is trained using high-fidelity terrain scan information (Scandot), and the terrain scan information is processed by an encoder-decoder autoencoder network to reconstruct the terrain. Then, the geometric features of the encoder are distilled into the MobileViT-GRU lightweight encoding module, so that the depth encoder uses the Scanot-based representation as the teacher signal for distillation. During stitching, the joint angular velocity and body angular velocity are first standardized, and then fused with the depth-aware information encoded by MobileViT-GRU to form a unified feature.

[0060] Optionally, the terrain-aware encoder adopts an MLP structure, with the intermediate layer dimension changing from 128 to 64 and the decoder dimension changing from 64 to 128. In the MobileViT-GRU module used for deep distillation, the GRU hidden layer dimension is 512. MobileViT adopts the existing lightweight structure and does not make additional customizations to the hierarchy and attention modules. The combination of MobileViT and GRU is used as a unique design to model visual historical memory. The distillation loss adopts L2 feature regression and no other loss types are introduced.

[0061] Optionally, during terrain reconstruction, the reconstruction loss formula for the encoder-decoder autoencoder network is:

[0062] Among them, high-fidelity terrain scanning Geometric latent features are encoded using an encoder-decoder autoencoder network. Decoder reconstruction And a reconstruction loss supervised encoder.

[0063] Optionally, in the depth encoder distillation stage, to extract the terrain-aware latent representation from the depth map, a lightweight MobileViT-GRU module is trained: MobileViT extracts local and global spatial geometric features, GRU models temporal dynamics, and the depth encoder distills the scan representation as the teacher signal, as shown in the formula: , Distillation loss is

[0064] The Warp camera ray casting module is integrated into LeggedGym, enabling GPU parallel rendering via NVIDIA Warp: it calculates the inverse of the camera intrinsic matrix for each pixel and combines it with the rotation matrix to define the unit ray direction. The origin of the ray is the camera position o, and the intersection point with the environment mesh is calculated. The "positive depth" is obtained by projecting t along the camera's principal axis.

[0065] The system outputs a physically consistent depth map. The simulation-to-real-world transfer adopts a strategy of "simulation domain randomization + real data processing": in the simulation, camera perturbation and Gaussian noise are added to the depth map to simulate sensor artifacts; in the real-world transfer, the zero-value pixels are interpolated with the global mean and smoothed with an exponential moving average to reduce inter-domain differences and improve transfer performance.

[0066] Optionally, the training process in this application follows a two-stage pipeline. Stage one is policy learning, where the policy, encoder, forward model, backward model, and scan encoder are jointly trained, with a total loss of...

[0067] The loss weight is set to integrate reinforcement learning and self-supervised objectives. The second stage is deep distillation, which trains a deep encoder to mimic a scanning encoder. The distillation loss is L2=Ldistill. By decoupling motion policy learning and perceptual distillation, high-quality scanning data can be used during training, and only depth sensor inference is required during deployment. This reduces task complexity, stabilizes visual learning, and accelerates convergence, making the policy robust and agile in complex scenarios.

[0068] Furthermore, based on the above, in some feasible embodiments, step S40 above may also include the following implementation steps D401 to D404.

[0069] Step D401: Calculate the straight-line distance between the current position and the target position in real time, and compare the straight-line distance with the preset far target determination threshold and the preset near target determination threshold; Step D402: If the straight-line distance is greater than the preset distant target determination threshold, the initial movement speed is adjusted according to the preset distant target adjustment rules to obtain a stable expected speed parameter that adapts to the navigation efficiency of distant targets; Step D403: If the straight-line distance is between the preset near target determination threshold and the preset far target determination threshold, then the initial movement speed is used as the stable expected speed parameter for adapting to the medium-distance scene. Step D404: If the straight-line distance is less than the preset near target determination threshold, the initial motion speed is adjusted according to the preset near target adjustment rules to obtain the stable expected speed parameters that are suitable for the motion stability of the near target.

[0070] Optionally, a preset far target determination threshold (L1) is used to distinguish the distance threshold for long-distance navigation scenarios, which is set to 5 meters in this embodiment; a preset near target determination threshold (L2) is used to distinguish the distance threshold for short-distance navigation scenarios, which is set to 1 meter in this embodiment; far target adjustment rules: strategies for adjusting the initial movement speed in long-distance scenarios, including but not limited to increasing by a factor of 1.2; near target adjustment rules: strategies for adjusting the initial movement speed in short-distance scenarios, including but not limited to linear attenuation combined with hysteresis judgment and dead zone suppression; hysteresis judgment: a state determination method to avoid boundary jitter by using the entry radius (L2=1 meter) and exit radius (L3=1.2 meters); dead zone suppression: a processing method to filter residual speed commands less than 0.05 m / s in near target scenarios.

[0071] Optionally, the straight-line distance between the current position and the target position is calculated in real time and compared with a preset far target determination threshold (5 meters) and a preset near target determination threshold (1 meter). If the straight-line distance is greater than 5 meters, the initial movement speed is adjusted according to the far target adjustment rule (1.2 times ratio) to obtain a stable expected speed parameter adapted to the navigation efficiency of far targets. If the straight-line distance is between 1 meter and 5 meters, the initial movement speed is used as the stable expected speed parameter adapted to the mid-range scenario. If the straight-line distance is less than 1 meter, the near target adjustment rule is activated. First, hysteresis judgment is used (entering the near target state when the straight-line distance is less than 1 meter, and exiting when it is greater than 1.2 meters) to avoid frequent switching. Then, dead zone suppression is performed on speed commands less than 0.05 m / s, and the initial movement speed is linearly decayed to obtain a stable expected speed parameter adapted to the stability of near target movement. During the adjustment process, terrain perception information is used for dynamic correction. For example, in sloping terrain, the far target adjustment ratio is changed to 1.1 times.

[0072] Optionally, by combining hysteresis judgment (e.g., entering with a radius of 1 meter and exiting with a radius of 1.2 meters) with dead zone suppression (e.g., suppression when the linear velocity is <0.05m / s), the strategy oscillation and circling behavior of the robot at the navigation endpoint can be completely avoided. Compared with traditional methods, the number of steps to converge at the endpoint is reduced to 0, which improves the actual deployment availability of navigation.

[0073] This embodiment combines segmented distance adaptation with hysteresis and dead zone mechanisms to adjust speed, balancing efficiency and stability at different navigation stages, avoiding speed oscillations in near-target scenarios, and providing precise input for motion control.

[0074] Furthermore, based on the content of any of the above embodiments, in some feasible embodiments, the above step S50 further includes the following implementation steps E501-E503.

[0075] Step E501: Based on the stable desired velocity parameters, the current joint position and body posture in the real-time body state information, calculate the deviation between the joint position and the ideal joint position that adapts to the stable desired velocity parameters, and output the adjustment amount of the deviation as the residual joint angle through the preset deviation adjustment algorithm. Step E502: Based on the motion requirements of the stable desired velocity parameters, call the preset basic gait joint trajectory, superimpose the residual joint angle with the preset basic gait joint trajectory according to the joint dimension, generate the target motion trajectory of each joint, and use it as the execution command to drive the robot's motion. Step E503: Send the execution command to the robot's joint drive module to control each joint to move in real time according to the target motion trajectory until the robot reaches the target position.

[0076] Optionally, residual joint angle: the deviation of the joint angle relative to the robot's default standing posture, used to fine-tune the posture to adapt to different motion requirements; proportional-derivative (PD) controller: a feedback control algorithm that calculates the control quantity through a proportional term (responding to the current deviation) and a derivative term (suppressing the rate of change of deviation); joint torque: the torque of the force that drives the robot's joint rotation, which is the core input of the joint drive module; speed command: the command generated by the navigation strategy that specifies the robot's expected translational and rotational motion parameters; motor action: the specific operating state of the motor required to achieve joint movement, including rotation angle, speed, etc.

[0077] Optionally, after receiving the speed command from the navigation strategy, the motion strategy first calculates the deviation between the current joint position and the ideal joint position based on the real-time body state (including joint position and body posture) and the stable desired speed parameters, and outputs the residual joint angle corresponding to the deviation. Then, it calls the preset proportional-derivative (PD) controller, takes the residual joint angle as input, and calculates the joint torque corresponding to each joint by combining the preset proportional gain (e.g., 30) and derivative gain (e.g., 11). Subsequently, the motion strategy converts the joint torque into motor actions that can be executed by the motor, and generates execution commands to drive the robot's movement. Finally, the execution commands are sent to the joint drive module to control the motors to move according to the commands, so as to realize the stable movement of the robot according to the speed command.

[0078] Optionally, proprioceptive input From the joint encoder and inertial measurement unit (IMU), including the following: Let velocity be the angular velocity of the robot body. Gravity direction vector in body coordinate system Joint position Joint angular velocity The action at the previous moment During the training phase, to improve the accuracy of policy and value estimation, the network has access to privileged information—scanots in ontology coordinates, denoted as... Its coverage area is forward meters, around Meters, with a sampling resolution of 0.1 meters, ultimately generating a The height matrix effectively encodes the local terrain around the robot.

[0079] Optionally, a forward depth image with a field of view of 87° is also provided during the training phase. For perception distillation, unlike the scandots used for training, depth images will serve as the primary terrain-aware input during deployment.

[0080] Optionally, output action This indicates a posture relative to the default standing posture. Residual joint angles:

[0081] This residual mechanism allows the strategy to fine-tune the stable attitude to adapt to different terrains.

[0082] Actual joint torque Calculated using a proportional-derivative (PD) controller:

[0083] in, and These are the proportional and differential gain parameters, respectively.

[0084] Optionally, the navigation decision model to be trained is trained using the PPO algorithm. The hidden layers of both the policy network and the value network are set to three layers, with the number of neurons being 512, 256, and 128 respectively. The bilinear projection matrix of the contrastive forward prediction has a dimension of 20, and the dimension of the latent state Z has a dimension of 32. The loss function for back-dynamic modeling is the mean squared error, and the weights of both the contrastive loss and the back-dynamic loss are set to 0.01. This weight is determined to be the optimal value through empirical parameter tuning.

[0085] Optionally, in the reinforcement learning motion control framework of this embodiment, latent space modeling is a key step in achieving efficient policy learning. The robot first processes continuous ontology state windows. (Encoding multi-dimensional proprioceptive information, including command velocity, body angular velocity, joint position, etc., into latent states) This encoding process can extract the core features of the ontology state and achieve a low-dimensional and compact representation of information.

[0086] Simultaneously, the terrain encoder processes the privileged information from the training phase—the height scan map (Scandot)—and outputs the terrain latent state. This state encapsulates the geometric features of the local terrain surrounding the robot. Subsequently, the latent state, along with the terrain latent state and the current ontology state, is input into the policy network. Its mathematical expression is: Based on these inputs, the policy network outputs residual joint angles relative to the default standing posture, which guide the robot's movements. Additionally, these inputs are also fed into the value network, mathematically expressed as:

[0087] The value network is used to evaluate the value of actions in the current state, providing a reference for strategy optimization.

[0088] Optionally, the speed command update frequency of the navigation strategy is 10Hz (updated once every 0.1 seconds), and the joint drive frequency of the low-level motion control is 50Hz (executed once every 0.02 seconds). After receiving the depth map, the navigation strategy outputs a 32-dimensional latent embedding through the MobileViT encoder, which is directly concatenated with the ontology features and then input into the decision model without additional feature fusion processing.

[0089] Optionally, in the self-supervised representation learning stage of this embodiment, inverse dynamics modeling is a key means to improve the relevance of potential representation decisions. Specifically, an inverse model is trained. The model uses the potential states at adjacent time points. Given the input, predict the action at the corresponding time.

[0090] Its loss function is the mean squared error loss:

[0091] Among them, a t loco To represent real actions, minimizing this loss encourages the latent representation to encode more information directly related to the decision action, filtering out redundant content irrelevant to the task. Combining the forward dynamics model (used to predict the latent state at the next moment) with this backward dynamics model can significantly improve the reinforcement learning algorithm's understanding of action-state relationships, making it more targeted when exploring the environment, more efficient in the optimization process, and ultimately improving the dynamic response and robustness of the policy in complex terrain.

[0092] This embodiment uses a proportional-derivative controller to accurately calculate joint torque and optimize motor action by combining residual joint angles, ensuring that speed commands are efficiently converted into actual motion and improving the accuracy and stability of motion control.

[0093] Furthermore, based on the content of any of the above embodiments, in some feasible embodiments, the robot autonomous navigation method further includes the following implementation steps F501 to F502 before step S30.

[0094] Step F501: Obtain the optimization objective of the navigation decision model to be trained, the state prediction loss generated by the preset comparative forward prediction module, and the key components of the preset reward function. The key components include the velocity tracking reward, attitude stabilization penalty, and collision avoidance penalty. The state prediction loss is the difference between the predicted next-time body state information and the actual collected next-time body state information. Step F502: Input the optimization objective, key components, and state prediction loss into the navigation decision model to be trained to obtain a preset navigation decision model that meets the optimization objective.

[0095] Optionally, contrastive forward prediction: a self-supervised representation learning method that trains a forward model to predict the next potential state based on the current potential state and action. It uses InfoNCE loss to align the predicted state with the true state and distinguish it from negative samples within the batch, improving the temporal consistency of the state representation. State prediction loss: a metric measuring the difference between the predicted next-time ontology state information and the actually collected next-time ontology state information, including but not limited to InfoNCE loss calculated based on the bilinear projection matrix, as shown in the formula...

[0096] Where Zt is the current potential state, and W is the bilinear projection matrix. The learnable bilinear projection matrix is ​​the denominator representing all latent states in the current mini-batch. Summation above (excluding those from other transitions) ), predicted Compared with positive samples Align with other negative samples within the batch. Distinguishing features: Reward function: A combination of signals that guides the model to learn the optimal strategy. Key components include velocity tracking reward (e.g., linear velocity tracking weight 1.5, angular velocity tracking weight 0.5, encouraging tracking of the desired velocity), attitude stability penalty (e.g., pitch angle alignment reward weight 0.1, vertical velocity penalty weight -1.0, suppressing unbalanced attitude), collision avoidance penalty (e.g., weight -10, reducing collision risk), and also includes power consumption penalty, joint acceleration penalty, etc.; Navigation decision model: A model used to generate robot navigation speed commands, including but not limited to a network structure based on the PPO algorithm. The input is the body state (e.g., joint position, IMU data) and depth map encoding features (extracted by the MobileViT-GRU module), and the output is planar linear velocity (x-axis -0.5~0.8m / s, y-axis -0.5~0.5m / s) and angular velocity (-1~1rad / s).

[0097] Optionally, the formula and weight of the reward items are detailed in Table 1 below:

[0098] Table 1 Among them, the multi-dimensional reward items for robot motion control include reward items such as linear velocity / angular velocity tracking and pitch angle alignment, as well as penalty items such as heading correction and vertical velocity penalty. Each reward item is quantified by mathematical expressions such as exponent and norm, and the weights are determined through experimental optimization to achieve a balance between speed tracking, attitude stability and motion efficiency through guidance strategies.

[0099] In this embodiment, the optimization objective of the navigation decision model to be trained is obtained (including but not limited to minimizing the state prediction loss and maximizing the reward function value to achieve navigation in complex terrain without a map). A pre-defined comparative forward prediction module generates a state prediction loss based on the current potential state and action using InfoNCE loss. This loss reflects the difference between the predicted next-moment ontology state information and the actual acquired next-moment ontology state information. The key components of the pre-defined reward function include a velocity tracking reward (linear velocity and angular velocity tracking), an attitude stability penalty (pitch angle and vertical velocity penalty), and a collision avoidance penalty. The weights of each component are determined experimentally (e.g., linear velocity tracking weight 1.5). The optimization objective, key components, and state prediction loss are input into the navigation decision model to be trained (using the PPO algorithm, with 512, 256, and 128 neurons in the network hidden layers). During training, the features of the depth map encoded by MobileViT-GRU and the ontology state features are combined to finally obtain a pre-defined navigation decision model that meets the optimization objective.

[0100] Optionally, to further stabilize the course, we introduce a yaw angle penalty in the world coordinate system:

[0101] And introduce automatic angular velocity correction action. .

[0102] Optionally, in unstructured terrain such as slopes, the sampling process allows for broader directional control, improving terrain adaptability and throughput. To encourage the robot to jump over obstacles rather than avoid them, we designed an additional reward mechanism based on pitch angle alignment. This is achieved by projecting the foreground scandots point cloud onto the xz plane and fitting the slope k using the least squares method. Sampling interval: , Slope fitting formula:

[0103] The target pitch angle is:

[0104] Pitch alignment reward is:

[0105] in Using the scaling factor, this method avoids the problem of full 3D fitting, is simple to calculate, and is suitable for real-time training.

[0106] Alternatively, alternatively, such as Figure 4 As shown, Figure 4This is a schematic diagram of the depth map-based end-to-end navigation strategy framework involved in this application. It directly maps the airborne forward depth map and ontology perception information to high-level velocity commands, simplifying the learning process and strengthening the coupling between perception and control. Specifically, the airborne forward depth map (otdepth) is first input into the MobileViT module to extract terrain geometric features and output a latent depth representation (ztdepth). Simultaneously, ontology perception information (including the aircraft's linear velocity (vtbase), angular velocity (ωt), attitude (θtbase), target displacement (Δg), and the previous velocity command (at)) is input. 1nav and ztdepth are concatenated to form a complete observation input. This fused observation is simultaneously fed into the Actor (policy network) and Critic (value network): the Actor outputs the high-level velocity command atnav at the current moment as a reference for the downstream motion controller's action; the Critic evaluates the state value vt corresponding to this action and feeds back optimization signals to the Actor through policy gradients, driving the policy to learn velocity commands that are more adapted to terrain and navigation needs, thus realizing an end-to-end closed loop of "perception-decision".

[0107] This embodiment achieves effective integration of the external environment and the robot's own state through explicit coding and splicing logic, providing comprehensive feature support for subsequent navigation decisions and improving the robot's perception and decision-making accuracy in complex scenarios.

[0108] Optionally, in the observation space, the navigation observation at each time step t of otnav integrates the body state and visual perception: the body part includes the linear velocity, angular velocity, roll-pitch-yaw attitude, and displacement between the robot's current position and the target position.

[0109] The robot body is located at the following position: The target location is Simultaneously, the action from the previous moment is incorporated to enhance temporal smoothness; visual perception is represented by a 106×60 depth map, which is processed by the MobileViT encoder into a 32-dimensional latent embedding ztdepth, and finally concatenated with the ontology features to form a complete observation. .

[0110] Optionally, in the action space, the higher-level strategy outputs velocity commands for planar linear velocity (x / y axis) and angular velocity.

[0111] Among them, compared to only containing The traditional design is better suited to the flexible movements of quadruped robots, such as lateral movement, and the reward function is based on the target distance. With the core as the basis, a weighted sum is used. The “approximation” term is expressed by a piecewise scale σ(dt).

[0112] Among them, for distant targets σ=1 and for near targets σ=5, the following is used:

[0113] Balancing long-range propulsion and short-range convergence; the "aircraft system heading alignment" term is penalized exponentially.

[0114] Suppressing yaw error, the yaw error of the aircraft system is denoted as ,in, The target displacement is represented in the machine system; the "velocity toward the target" term only awards the velocity component in the target direction.

[0115] It also includes penalties for approaching the target, such as deceleration, stopping, and collision, with weights set experimentally (e.g., weight of 2 for approaching the target, penalty of -10 for collision).

[0116] To ensure stable target convergence, motion scaling with hysteresis and dead zone is designed: through...

[0117] (Using rs=0.6, k=2) Smoothly scale commands with distance; introduce a binary state zt to mark the near-target region (entry radius 0.2m, exit radius 0.4m) to avoid boundary jitter; when near the target, perform dead zone suppression on commands less than the threshold δ (x / y axis 0.1m / s, angular velocity 0.2rad / s), and finally...

[0118] in, m、 , m、 m, and (m / s, m / s, rad / s), combining scaling and suppression results to achieve smooth behavior of "alignment-deceleration-stop".

[0119] Optionally, the end-to-end depth-only navigation strategy process is as follows: the depth map is processed by a lightweight MobileViT encoder to obtain a compact latent representation, which is then concatenated with ontology features to form a complete observation vector; this fused vector is input into a shared multi-layer network containing a strategy head and a value head, and the strategy head outputs a continuous action vector containing x / y axis velocities and yaw rate. ,in The system represents a parameterized stochastic strategy. The entire system is trained from scratch in the simulation using the PPO algorithm, relying only on depth and body state. It does not require privileged information or maps, making it easy to deploy directly to real robots. This strategy can implicitly learn three-dimensional geometric interpretation capabilities and can choose to bypass or cross obstacles based on the height of the obstacles. With the full-body mobility of the quadruped robot, it achieves highly adaptable and efficient navigation that surpasses classic two-dimensional path planning.

[0120] Optionally, several penalties can be introduced to ensure navigation safety and stability, as shown in Table 2. The navigation strategy reward items and weights are as follows:

[0121] Table 2 Optionally, in the end-to-end depth-only navigation strategy, the depth map is encoded using MobileViT and concatenated with ontology features. This concatenation is then input into a shared network containing both a policy head and a value head, and the output is a velocity command. The system is trained in simulation using the PPO algorithm, requiring no privileged information or map. It can be directly deployed to real robots and implicitly learns 3D geometric interpretation capabilities (such as choosing to bypass or cross obstacles based on obstacle height), achieving more efficient navigation tailored to quadruped robots.

[0122] This embodiment improves the model's adaptability to terrain and navigation stability by fusing comparative forward prediction state prediction loss with a multi-dimensional reward function, thus meeting the real-time control requirements in mapless scenarios.

[0123] Furthermore, based on the content of any of the above embodiments, in some feasible embodiments, the above-described robot autonomous navigation method further includes steps G601-G602 before step S20.

[0124] Step G601: If the robot is in a simulation environment, perform domain adaptation processing on the depth perception information, including random camera pose perturbation, pixel-by-pixel Gaussian noise or occlusion-type pixel loss, to obtain the domain-adapted depth perception information, and use the domain-adapted depth perception information as the depth perception information to be encoded. Step G602: If the robot is in a real environment, perform global mean interpolation and temporal smoothing on the zero-value missing pixels in the depth perception information to obtain the depth perception information after global mean interpolation and temporal smoothing, and use the depth perception information after global mean interpolation and temporal smoothing as the depth perception information for encoding.

[0125] Optionally, simulation-side randomization: a method of applying systematic perturbations to depth sensing information in a simulation environment to simulate real sensor artifacts, including random camera pose perturbations (including pitch angle and translation offset), pixel-by-pixel Gaussian noise, and occlusion-type dropout guided by Sobel edges; real-side zero-value interpolation: in a real environment, zero-value missing pixels in the depth sensing information are filled by calculating the global mean of all currently valid pixels; exponential moving average temporal smoothing: in a real environment, the interpolated depth sensing information is processed by using an exponential moving average algorithm (smoothing coefficient set to 0.01) to reduce noise and improve the stability of the data in the temporal dimension.

[0126] In this embodiment, if the robot is in a simulation environment, the depth perception information is subjected to simulation-side randomization processing. The random camera pose perturbation is set to pitch ±2° and small translation. The variance of the pixel-by-pixel Gaussian noise is configured according to the characteristics of the real sensor. Occlusion-type dropout uses Sobel edge detection results to determine occlusion areas and simulate pixel loss. The processed depth perception information is used as the subsequent encoding input. If the robot is in a real environment, for zero-value missing pixels in the depth perception information, the global mean of all current valid pixels is calculated for interpolation. Then, an exponential moving average algorithm with a smoothing coefficient of 0.01 is used for temporal smoothing. The processed depth perception information is used as the subsequent encoding input. This processing flow effectively reduces inter-domain differences by simulating real noise on the simulation side and optimizing data quality on the real side, providing reliable input for the depth encoder and supporting stable perception and navigation of the robot in complex terrain.

[0127] Optionally, the real robot deployment uses CloudDeep Technology's quadruped robot platform, with an NVIDIA Jetson NX processor; the exponential moving average smoothing coefficient of the depth map is set to 0.01, and the global mean of zero-value pixel interpolation is the real-time calculated mean of the current effective pixels; the model has been trained and tested in both simulation and real environments on complex terrains such as slopes, stairs, and gravel. The GPU parallel depth rendering module has been tested to achieve a rendering speed of 20,000 frames per second under hundreds of parallel environments, which greatly improves training efficiency.

[0128] Optionally, at the perception input and representation level, in addition to the forward depth camera paired with MobileViT-GRU lightweight encoding, a multimodal fusion scheme can be used: replace monocular depth with binocular / structured light depth and event cameras, or combine 16-line / 32-line low-beam LiDAR to achieve sparse-dense fusion. The encoder can be ConvNeXt, EfficientNet, etc. to replace MobileViT, and the temporal modeling can be TCN or lightweight Transformer to replace GRU, which can enrich the perception dimension and improve the accuracy of feature extraction in complex environments.

[0129] Alternatively, along the distillation path, in addition to Scandots as teacher representation and L2 feature regression as loss, contrastive distillation and mutual information maximization distillation can achieve more accurate teacher-student alignment and enhance the terrain adaptability of deep representation.

[0130] Optionally, at the interface between low-level control and high-level navigation, the motion space can be decoupled from "velocity command - residual joint pose" and reduced to vx,ω to improve motion stability. Upgrading to local sub-targets can enhance navigation scalability. In terms of training paradigm, the SAC / TQC plasma policy algorithm is more efficient than PPO samples. The sparse arrival reward combined with the walkability dense reward can simplify reward design and ensure navigation safety.

[0131] Optionally, in the Sim-to-Real phase, the nearest local geometric prior provided by lightweight SLAM / odometer can address the challenges of deep reliability without disrupting the mapless paradigm.

[0132] In addition, this application also provides a robot autonomous navigation device, please refer to... Figure 5 , Figure 5 This is a schematic diagram of the robot autonomous navigation device according to the embodiments of this application. The robot autonomous navigation device provided in this application includes: The high-level navigation strategy module H01 is used to acquire the robot's real-time body status information, current position, target position, and depth perception information of the external environment; The high-level navigation strategy module H02 is also used to encode depth perception information, and then splice the encoded depth perception information with the real-time body state information to obtain a unified feature representing the external environment and the robot's own state. The high-level navigation strategy module H03 is also used to input unified features and target positions into a preset navigation decision model to generate the robot's initial motion speed; The high-level navigation strategy module H04 is also used to calculate the straight-line distance between the current position and the target position in real time, and adjust the initial movement speed based on the straight-line distance to obtain a stable desired speed parameter; The low-level motion strategy module H05 is used to generate execution commands to drive the robot's motion based on the stable desired velocity parameters and real-time body state information, and to control the robot to execute the execution commands and navigate to the target position.

[0133] The robot autonomous navigation device provided in this application, employing the robot autonomous navigation method in the above embodiments, can solve the problem of efficient autonomous navigation of robots without maps and privileged sensors. Compared with the prior art, the beneficial effects of the robot autonomous navigation system provided in this application are the same as those of the robot autonomous navigation method provided in the above embodiments, and other technical features of this robot autonomous navigation system are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0134] In addition, this application also provides a robot. Please refer to... Figure 6 , Figure 6 This is a schematic diagram of the robot involved in the embodiments of this application.

[0135] This application provides a robot, which includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to perform the robot autonomous navigation method in Embodiment 1 above.

[0136] The following is for reference. Figure 6 , Figure 6 This is a schematic diagram of the structure of a robot involved in the embodiments of this application, showing a structural schematic diagram suitable for implementing the robot of the embodiments of this application. Figure 6 The robot shown is merely an example and should not be construed as limiting the functionality and scope of use of the embodiments of this application.

[0137] like Figure 6 As shown, the robot may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory (ROM) 1002 or a program loaded from a storage device 1003 into a random access memory (RAM) 1004. The RAM 1004 also stores various programs and data required for robot operation. The processing unit 1001, ROM 1002, and RAM 1004 are interconnected via a bus 1005. An input / output (I / O) interface 1006 is also connected to the bus. Typically, the following devices can be connected to the I / O interface 1006: input devices 1007 including, for example, a touchscreen, touchpad, keyboard, mouse, image sensor, microphone, accelerometer, gyroscope, etc.; output devices 1008 including, for example, a liquid crystal display (LCD), speaker, vibrator, etc.; storage devices 1003 including, for example, magnetic tape, hard disk, etc.; and communication devices 1009. Communication device 1009 allows the robot to communicate wirelessly or wiredly with other devices to exchange data. While the figure shows a robot with various devices, it should be understood that implementing or possessing all of the shown devices is not required. More or fewer devices may be implemented alternatively.

[0138] Specifically, according to the embodiments disclosed in this application, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments disclosed in this application include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from ROM 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this application.

[0139] The robot provided in this application, employing the autonomous navigation method described in the above embodiments, can solve the problem of efficient autonomous navigation for robots without maps and privileged sensors. Compared with the prior art, the beneficial effects of the robot provided in this application are the same as those of the autonomous navigation method provided in the above embodiments, and other technical features of the robot are the same as those disclosed in the previous embodiment method, and will not be repeated here.

[0140] Furthermore, this application provides a computer-readable storage medium. This computer-readable storage medium stores a robot autonomous navigation program, which, when executed by a processor, implements the steps of the aforementioned robot autonomous navigation method.

[0141] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or system that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or system. Unless otherwise specified, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or system that includes that element.

[0142] The sequence numbers of the embodiments in this application are for descriptive purposes only and do not represent the superiority or inferiority of the embodiments.

[0143] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a software product. This computer software product is stored in a computer-readable storage medium (such as ROM / RAM, magnetic disk, optical disk) as described above, and includes several instructions to cause a device (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0144] The above are merely preferred embodiments of this application and do not limit the patent scope of this application. Any equivalent structural or procedural transformations made using the content of this application's specification and drawings, or direct or indirect applications in other related technical fields, are similarly included within the patent protection scope of this application.

Claims

1. A method for autonomous navigation of a robot, characterized in that, The robot autonomous navigation method includes: Acquire the robot's real-time body state information, current position, target position, and depth perception information of the external environment; The depth perception information is encoded, and obstacle information and terrain slope information are extracted from the depth perception information. The extracted obstacle information and terrain slope information are then mapped into feature vectors of a preset dimension. The feature vectors of the preset dimension are concatenated with the joint angular velocity and the body angular velocity according to the feature dimension to form a unified feature that characterizes the external environment and the robot's own state. The unified features and the target position are input into a preset navigation decision model to generate the robot's initial movement speed; The straight-line distance between the current position and the target position is calculated in real time, and the straight-line distance is compared with a preset far target determination threshold and a preset near target determination threshold. If the straight-line distance is greater than the preset far target determination threshold, the initial movement speed is adjusted according to the preset far target adjustment rules to obtain a stable expected speed parameter that adapts to the navigation efficiency of the far target. If the straight-line distance is between the preset near target determination threshold and the preset far target determination threshold, then the initial movement speed is used as the stable expected speed parameter for adapting to medium-distance scenarios. If the straight-line distance is less than the preset near target determination threshold, the initial movement speed is adjusted according to the preset near target adjustment rules to obtain a stable expected speed parameter that adapts to the stability of near target movement; Based on the stable desired velocity parameters, the current joint position and body posture in the real-time body state information, the deviation between the joint position and the ideal joint position that adapts to the stable desired velocity parameters is calculated, and the adjustment amount of the deviation is output as the residual joint angle through a preset deviation adjustment algorithm. Based on the motion requirements of the stable desired speed parameters, the preset basic gait joint trajectory is invoked, and the residual joint angle is superimposed on the preset basic gait joint trajectory according to the joint dimension to generate the target motion trajectory of each joint, which serves as the execution command to drive the robot's motion. The execution command is sent to the joint drive module of the robot to control each joint to move in real time according to the target motion trajectory until the robot reaches the target position.

2. The robot autonomous navigation method as described in claim 1, characterized in that, The steps for acquiring the robot's real-time body state information, current position, target position, and depth perception information of the external environment include: The robot's current joint position, joint angular velocity, body angular velocity, gravity direction vector in the body coordinate system, and body posture are collected and integrated to obtain the real-time body state information; The robot's current position is determined based on the real-time body state information, and the coordinates of the destination that the robot needs to reach, sent by the external navigation system, are obtained as the target position. The terrain within a preset range surrounding the current location is scanned and rendered, and depth perception information representing the terrain within the preset range is output.

3. The robot autonomous navigation method as described in claim 2, characterized in that, Before the step of inputting the unified features and the target position into a preset navigation decision model to generate the robot's initial motion speed, the method further includes: The optimization objective of the navigation decision model to be trained is obtained, along with the state prediction loss generated by the pre-defined comparative forward prediction module and the key components of the pre-defined reward function. The key components include a velocity tracking reward, an attitude stabilization penalty, and a collision avoidance penalty. The state prediction loss is the difference between the predicted next-moment body state information and the actual collected next-moment body state information. The optimization objective, the key components, and the state prediction loss are input into the navigation decision model to be trained to obtain the preset navigation decision model that meets the optimization objective.

4. The robot autonomous navigation method as described in claim 3, characterized in that, Before the step of encoding the depth-sensing information, the method further includes: If the robot is in a simulation environment, then the depth perception information is subjected to domain adaptation processing, including random camera pose perturbation, pixel-by-pixel Gaussian noise, or occlusion-type pixel loss, to obtain the domain-adapted depth perception information, and the domain-adapted depth perception information is used as the encoded depth perception information. If the robot is in a real environment, then the zero-value missing pixels in the depth perception information are subjected to global mean interpolation and temporal smoothing of effective pixels to obtain the depth perception information after global mean interpolation and temporal smoothing, and the depth perception information after global mean interpolation and temporal smoothing is used as the depth perception information to be encoded.

5. A robot autonomous navigation device, characterized in that, The robot autonomous navigation device includes: The high-level navigation strategy module is used to acquire the robot's real-time body state information, current position, target position, and depth perception information of the external environment; The high-level navigation strategy module is further configured to encode the depth perception information, and concatenate the encoded depth perception information with the real-time body state information to obtain a unified feature representing the external environment and the robot's own state. Specifically, the high-level navigation strategy module is further configured to encode the depth perception information, extract obstacle information and terrain slope information from the depth perception information, map the extracted obstacle information and terrain slope information into a feature vector of a preset dimension, and concatenate the feature vector of the preset dimension with the joint angular velocity and the body angular velocity according to the feature dimension to form a unified feature representing the external environment and the robot's own state. The high-level navigation strategy module is also used to input the unified features and the target position into a preset navigation decision model to generate the robot's initial movement speed; The high-level navigation strategy module is further configured to calculate the straight-line distance between the current position and the target position in real time, and adjust the initial movement speed based on the straight-line distance to obtain a stable expected speed parameter. Specifically, the high-level navigation strategy module is further configured to calculate the straight-line distance between the current position and the target position in real time, and compare the straight-line distance with a preset far target determination threshold and a preset near target determination threshold. If the straight-line distance is greater than the preset far target determination threshold, the initial movement speed is adjusted according to a preset far target adjustment rule to obtain a stable expected speed parameter adapted to the navigation efficiency of far targets. If the straight-line distance is between the preset near target determination threshold and the preset far target determination threshold, the initial movement speed is used as a stable expected speed parameter adapted to medium-range scenarios. If the straight-line distance is less than the preset near target determination threshold, the initial movement speed is adjusted according to a preset near target adjustment rule to obtain a stable expected speed parameter adapted to the stability of near target movement. A low-level motion strategy module is used to generate execution instructions to drive the robot's motion based on the stable desired velocity parameters and the real-time body state information, and control the robot to execute the execution instructions to navigate to the target position. Specifically, the low-level motion strategy module is used to calculate the deviation between the joint position and the ideal joint position adapted to the stable desired velocity parameters based on the stable desired velocity parameters, the current joint position in the real-time body state information, and the body posture; output the adjustment amount of the deviation as the residual joint angle through a preset deviation adjustment algorithm; based on the motion requirements of the stable desired velocity parameters, call the preset basic gait joint trajectory, and superimpose the residual joint angle and the preset basic gait joint trajectory according to the joint dimension to generate the target motion trajectory of each joint as the execution instructions to drive the robot's motion; send the execution instructions to the robot's joint drive module to control each joint to move in real time according to the target motion trajectory until the robot reaches the target position.

6. A robot, characterized in that, The robot includes a processor, a memory, and a robot autonomous navigation program stored in the memory that can be executed by the processor, wherein when the robot autonomous navigation program is executed by the processor, it implements the steps of the robot autonomous navigation method as described in any one of claims 1 to 4.

7. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a robot autonomous navigation program, wherein when the robot autonomous navigation program is executed by a processor, it implements the steps of the robot autonomous navigation method as described in any one of claims 1 to 4.

Images