Robotic swarm vision navigation method based on differentiable physics and surrogate model

Abstract

A visual navigation method for robot swarms based on differentiable physics and agent models is proposed. A high-level navigation policy network processes raw image information acquired from a depth camera, using a deep learning model to understand the environmental geometry, identify obstacles and dynamic agents, and output velocity commands for obstacle avoidance, collision avoidance, and target navigation. A low-level motion policy network translates these velocity commands into joint position and torque commands for the quadruped robot's legs, generating stable and flexible body movements to execute velocity tracking strategies. After training using a forward-backward asymmetric optimization framework, swarm control is achieved in the online phase. This invention improves training sample efficiency and policy transferability in scenarios relying solely on local visual perception, while overcoming the curse of dimensionality in multi-agent learning. It can be flexibly extended to large-scale robot swarms and achieves zero-sample transfer from simulation to real-world scenarios.

Description

Technical Field

[0001] This invention relates to a technology in the field of robot navigation, specifically a visual navigation method for robot swarms based on differentiable physics and surrogate models. Background Technology

[0002] Traditional quadruped robot swarm control relies heavily on global localization, ideal communication, or simplified models, making stable deployment difficult in dynamic, unstructured environments. Existing robot navigation and localization technologies suffer from the curse of dimensionality, low training efficiency, and difficulty in scaling to large-scale robot swarms. Summary of the Invention

[0003] To address the aforementioned shortcomings of existing technologies, this invention proposes a visual navigation method for robot swarms based on differentiable physics and surrogate models. This method embeds prior physical knowledge into the policy learning process in the form of differentiable optimization. Through a hierarchical control architecture and a forward-backward asymmetric optimization mechanism, it can improve training sample efficiency and policy transferability in scenarios that rely solely on local visual perception, while overcoming the curse of dimensionality in multi-agent learning. It can be flexibly extended to large-scale robot swarms and achieve zero-sample transfer from simulation to real-world scenarios.

[0004] This invention is achieved through the following technical solution:

[0005] This invention relates to a visual navigation method for robot swarms based on differentiable physics and surrogate models, comprising:

[0006] Step 1: Construct a high-fidelity simulation environment with differentiable physics simulation, including a quadruped robot;

[0007] Step 2: Construct a high-level navigation strategy network. Based on the raw image information acquired by the depth camera, use a deep learning model to understand the environmental geometry, identify obstacles and dynamic intelligent agents, and output speed commands to achieve obstacle avoidance, collision avoidance and target navigation. Construct a low-level motion strategy network to convert the speed commands into joint position and torque commands for the quadruped robot's legs, and generate stable and flexible body movements to execute the speed tracking strategy.

[0008] Step 3: Construct a differentiable physical proxy model to approximate the backpropagation process of the high-level navigation policy network and the low-level motion policy network in Step 2 during the training phase.

[0009] Step 4: Training is performed using a forward-backward asymmetric optimization framework. In the forward execution phase, a high-level navigation strategy network and a low-level motion strategy network are used for inference in the high-fidelity simulation environment built in Step 1 to accurately capture dense contact details in robot motion. In the backpropagation phase, gradient updates are performed on the high-level navigation strategy network and the low-level motion strategy network in the differentiable physical proxy model built in Step 3 to obtain smooth gradient signals.

[0010] Step 5: Deploy the high-level navigation strategy network and low-level motion strategy network trained in Step 4 to a real-world experimental scenario to achieve cluster control.

[0011] This invention relates to a robot swarm visual navigation system for implementing the above-described method, comprising: a control layer consisting of a navigation strategy module and a motion strategy module; an optimization layer consisting of a mass proxy model and a rigid body proxy model; and a simulation environment module. Specifically: the navigation strategy module extracts visual features and performs temporal decision processing based on depth image information acquired by an onboard depth camera to obtain a desired velocity command; the motion strategy module performs state-action mapping processing based on the desired velocity command output by the navigation strategy module and the robot's body perception information to obtain joint control commands; the simulation environment module performs physical simulation calculations based on the joint control commands, updates the robot's state, and feeds back the simulated depth image and body state to the control layer; the mass proxy model performs differentiable dynamics calculations based on the input and output of the navigation strategy module to obtain a gradient optimization signal for the navigation strategy; the rigid body proxy model performs differentiable rigid body dynamics calculations based on the input and output of the motion strategy module to obtain a gradient optimization signal for the motion strategy; and the optimization layer backpropagates the gradient optimization signal to the control layer to update the network parameters of the navigation strategy module and the motion strategy module.

[0012] Technical effect

[0013] Compared with existing technologies, this invention, through a layered decoupled navigation-motion control architecture and a corresponding forward-backward asymmetric proxy model optimization mechanism, achieves similar performance with only 2% of the simulation sample size of the PPO method, reducing training time by 50% (actual data: this invention achieved a reward value of -1.89±0.23 in 1.5 hours of training, while the PPO algorithm required 2.9 hours to achieve a reward value of -2.11±0.20). In a circular fence scenario, this invention achieved a task completion rate of 98.96% with 96 robots. In a complex maze scenario, even when expanded to 512 robots, the task completion rate remained above 97%, and the time required to reach a 75% completion rate increased approximately linearly with the scale. This invention can smoothly increase speed according to speed commands (gradually increasing the target speed from 0.75 m / s to 1.5 m / s, the actual average speed of this invention increases almost linearly with the command increase, and maintains a high task completion rate at all speeds). In contrast, the actual speed of the PPO algorithm is not sensitive to changes in commands, and at high-speed commands (>1.2 m / s), the speed remains low. (The task completion rate decreased significantly at m / s); the present invention achieved a pass rate of over 80% in an extremely narrow passage with a width of 0.6m (the success rate of the PPO strategy dropped precipitously at a width of 0.9m); in five real-world scenarios including forests, narrow bridges, Chinese pavilions, and cluttered rooms, six Yushu Go2 robots were deployed directly with zero samples, relying entirely on local depth vision (without communication or global localization). Attached Figure Description

[0014] Figure 1 This is a schematic diagram of the system structure of the present invention;

[0015] Figure 2 This is a schematic diagram of a differentiable physical proxy model;

[0016] Figure 3 This is a schematic diagram illustrating the experimental results of zero-shot transfer learning in real-world multi-scenario verification of the present invention;

[0017] In the image: (A) a narrow bridge requiring group coordination, (B) a forest environment with natural obstacles, (C) a Chinese pavilion with narrow spaces, (D) a narrow gap requiring careful manipulation, and (E) a cluttered room with randomly placed obstacles, (A1-A4) noisy depth images captured by the yellow robot at different time steps.

[0018] Figure 4 This is a schematic diagram comparing the training sample efficiency of the present invention and the PPO algorithm;

[0019] Figure 5 This is a schematic diagram illustrating the scalability analysis of the present invention in a circular fence scenario;

[0020] Figure 6 This is a schematic diagram illustrating the scalability analysis of the present invention in complex maze scenarios;

[0021] Figure 7 This is a schematic diagram illustrating the biological-like group cooperative behavior emerging from the quadruped robot swarm trained according to the present invention.

[0022] Figure 8 This is a schematic diagram comparing the controllability of the present invention and the PPO algorithm in improving traffic speed;

[0023] Figure 9 This is a schematic diagram comparing the present invention with the PPO algorithm in fine control of narrow regions. Detailed Implementation

[0024] like Figure 1 and Figure 2 As shown in this embodiment, a robot swarm visual navigation system based on differentiable physics and a surrogate model is proposed. The system includes a control layer consisting of a navigation strategy module and a motion strategy module, an optimization layer consisting of a mass surrogate model and a rigid body surrogate model, and a simulation environment module. Based on this system, this embodiment constructs a quadruped robot swarm control model based on differentiable physics and trains a decentralized navigation and obstacle avoidance strategy based on local vision within it. The multi-robot swarm navigation task is defined as an optimization problem, where each robot is modeled as a discrete-time dynamic system with a continuous state and control input space. The state evolution of the system is described by a differentiable function. At each time step, the robot outputs control commands through the policy network based on the current observations (including depth image and body perception), updates its state after execution, and receives a loss signal.

[0025] This embodiment relates to a visual navigation method for robot swarms based on differentiable physics and surrogate models, including:

[0026] Step 1: Build a simulation environment based on the NVIDIA IsaacGym platform, specifically including:

[0027] 1.1 Geometric Modeling: In Blender 3D modeling software, create a scene model with high detail based on the scale of the real world.

[0028] 1.2 Model Export: Export the completed scene model in a universal format (OBJ, USD).

[0029] 1.3 Scene Loading and Instantiation: In Isaac Gym, the exported model files are loaded through the corresponding API interface, and the flexible invocation, randomized placement, and batch instantiation of scenes are implemented in the training code.

[0030] The simulation environment includes diverse and complex scenes such as open fields, mazes, sheep pens, and forests.

[0031] Step 2: Constructing as follows Figure 2 The differentiable physics proxy model shown includes a differentiable particle model and a differentiable rigid body model. It is used to calculate the corresponding gradients during backpropagation when training high-level navigation strategies and low-level motion strategies.

[0032] The differentiable physical proxy model is as follows: ,in: and The model outputs for time step k+1 and time step k are respectively. For time step k, and These are surrogate models and real system dynamics, respectively. is the time step k slack variable, representing the residual between the surrogate model and the real system dynamics.

[0033] The discrete-time update equation for the differentiable particle model in the aforementioned differentiable physical proxy model is: ,in: For the robot in time step Location, For time steps speed, The time step for integration.

[0034] The dynamic equations of the differentiable single rigid body model in the aforementioned differentiable physical proxy model are as follows: , , , ,in: For the robot's position, For speed, Let the azimuth be expressed as a unit quaternion. Angular velocity, The net external torque acting on the robot's rigid body. Here is the robot's rotational inertia matrix. The control input is the ground reaction force from its legs. .

[0035] The third step is to construct a high-level navigation strategy network based on a convolutional recurrent neural network (CRNN), and obtain a two-dimensional velocity command consisting of linear velocity and yaw rate based on the depth image and the two-dimensional reference velocity generated in real time from the robot's current position and the target point.

[0036] The convolutional recurrent neural network includes: three layers of LeakyReLU activated convolutional layers with kernel sizes of 2, 3, and 3, and output channels of 32, 64, and 128 respectively; a linear layer; a GRU unit; and a fully connected layer. The depth image is processed by the three convolutional layers to extract features, which are then flattened and mapped to 192-dimensional visual features by the linear layer. The reference velocity is mapped to 192-dimensional observation features by a linear layer with a weight scaling factor of 0.5. The visual features and observation features are added element-wise and input to a 192-dimensional GRU unit in the hidden layer for temporal modeling. Finally, a two-dimensional velocity command is output through the fully connected layer.

[0037] The hidden state of the GRU unit is passed between time steps, enabling the policy to have short-term memory and motion intention inference capabilities.

[0038] The fourth step is to construct a low-level motion policy network based on a multilayer perceptron (MLP) to generate joint angle commands based on two-dimensional velocity commands.

[0039] The low-level motion strategy network includes two hidden layers, each with 256 neurons, using the ReLU activation function. Its inputs are the target velocity and yaw angle (4D), phase variable (2D), base angular velocity (3D), base attitude angle (2D), 12 joint angles, and 12 joint velocities output by the high-level navigation strategy network. The output is a 12-dimensional joint angle command.

[0040] The joint angle command is converted into joint torque by a simplified PD controller, and then mapped into foot force through the foot Jacobian matrix to drive the rigid body model update.

[0041] Step 5: Construct a multi-objective loss function and train the policy network constructed in steps 3 and 4 using a gradient accumulation mechanism. This includes:

[0042] 5.1 Initialization: Initialize the high-level navigation policy network and the low-level motion policy network separately, and set their respective gradient accumulation steps. At the same time, set the initial robot state and global target position to lay the foundation for subsequent collaborative training.

[0043] 5.2 Forward Interaction and Data Acquisition Loop: In the simulation environment, the high-level navigation strategy outputs the desired motion commands based on the depth image and target velocity, while the low-level motion strategy generates specific joint position commands based on the robot's body state. The simulation environment executes the commands and updates the state, while simultaneously calculating the loss functions of the navigation and motion strategies.

[0044] 5.3 Gradient Accumulation and Policy Update: A gradient accumulation mechanism is adopted. After the navigation strategy accumulates a set number of steps of loss, the gradient is calculated and the network parameters are updated through a differentiable mass surrogate model. Similarly, after the motion strategy accumulates a corresponding number of steps of loss, the parameters are updated with the help of a differentiable rigid body surrogate model to improve training stability.

[0045] The gradient accumulation mechanism involves setting each simulation environment to contain 96 robot instances running in parallel. For the high-level navigation strategy, the gradient is calculated based on a differentiable point mass surrogate model every 80 simulation steps of accumulated loss, and the network parameters are updated using the AdamW optimizer with an initial learning rate of 0.001, followed by cosine annealing to 0.00001. For the low-level motion strategy, the gradient is calculated based on a differentiable rigid body surrogate model every 20 simulation steps of accumulated loss, also using the AdamW optimizer with an initial learning rate of 0.001, followed by cosine annealing to 0.0001. The total number of training steps is 5000 optimization steps for the motion strategy and 1250 optimization steps for the navigation strategy. Domain randomization is introduced during training: gait frequencies are randomly sampled between 1.7 and 2.5 Hz, and the maximum permissible speed is randomly sampled between 0.5 and 1.5 m / s to enhance the robustness and generalization ability of the strategy.

[0046] 5.4 Iterative Convergence: Repeat the forward interaction and policy update loop until the loss functions of the navigation policy and motion policy converge, completing the entire training process.

[0047] The training is performed using a gradient descent method based on differentiable physics. The loss is recorded at each simulation step, and after accumulating a certain number of steps, the gradient is calculated through backpropagation using a surrogate model to update the network parameters.

[0048] The multi-objective loss function includes: the loss function of the high-level navigation strategy and the loss function of the low-level motion strategy.

[0049] The loss function of the high-level navigation strategy Among them: speed loss Weight Obstacle avoidance loss Weight Collision loss Weight Overspeed loss Weight Acceleration loss Weight ; Directional loss Weight ; Target direction loss Weight Time loss Weight ;

[0050] The loss function of the low-level motion strategy Among them: velocity tracking loss Weight Attitude tracking loss Weight Angular velocity tracking loss Weight loss of fuselage height Weight Gravity projection loss Weight Action regularization loss Weight Torque regularization loss Weight Swing leg trajectory loss Weight .

[0051] Step 6: Deploy the policy network trained in steps 3 and 4 onto a quadruped robot in a real-world experimental scenario, using zero-sample transfer, i.e., whether the policy trained in simulation can be directly deployed onto a real quadruped robot without any additional fine-tuning.

[0052] Through practical application experiments, the robot swarm visual navigation method based on differentiable physics and a proxy model of this invention was tested on the Unitree Robotics Go2 quadruped robot platform and the Intel RealSense D435i depth camera in a completely decentralized manner (no communication between robots) and relying solely on local depth vision (input resolution 16×12 pixels) for real-time decision-making. Figure 3 As shown, in five challenging real-world scenarios—a forest, a narrow bridge, a Chinese-style pavilion, and a cluttered room—six robots successfully transitioned from simulation to reality. Figure 4 As shown, this invention demonstrates a significant advantage in sample efficiency, requiring only about 2% of the simulation sample size of the PPO method to achieve a similar performance level, and obtaining a higher final reward value after convergence. It also achieves an acceleration of about 50% in actual training time.

[0053] like Figure 5 and Figure 6 As shown, the quantitative evaluation results are presented in circular fence and complex maze scenarios: In the fence scenario with up to 96 robots, the task completion rate reached 98.96%; in the maze scenario with 512 robots, the task completion rate remained above 97%, and the time required to reach 75% completion rate showed a stable, approximately linear increase with the number of robots.

[0054] Refined Motion Control and Velocity Tracking Capabilities: This invention achieves precise tracking and refined motion control of velocity commands by combining low-level motion strategies with a differentiable rigid body proxy model. For example... Figure 8 As shown, when the target speed gradually increases from 0.75 m / s to 1.5 m / s, the strategy of this invention can respond accurately, with the actual average speed increasing linearly with the increase of the instruction, and maintaining a high task completion rate at all speeds. In contrast, the PPO strategy exhibits poor instruction following ability.

[0055] like Figure 9 As shown, the present invention demonstrates fine-grained control capabilities in narrow passage scenarios: when the passage width decreases from 1.0m to 0.5m, the strategy of this invention maintains a pass rate of over 80% even when the width drops to 0.6m, while the success rate of the PPO strategy drops precipitously when the width is less than 0.9m. This indicates that the present invention, through the deterministic, low-noise gradient provided by differentiable physics, supports centimeter-level precision fine-tuning of the motion trajectory, exhibiting significant advantages in extremely narrow spaces.

[0056] Experiments verified that in unstructured terrain, the robot can overcome visual interference and achieve robust obstacle avoidance and formation maintenance among continuous obstacles. In extreme spaces, the robot exhibits centimeter-level perception accuracy and fine-grained coordination capabilities, demonstrating coordination modes such as proactive yielding and alternating passage. In high-density dynamic interaction scenarios, the robot effectively solves complex conflict problems such as two-way vehicle passing and congestion avoidance through real-time path adjustment and intermittent pauses. These results demonstrate that, while maintaining fully distributed decision-making and local perception constraints, the present invention achieves efficient and robust zero-shot simulation to real-world transfer.

[0057] Compared to existing technologies that are difficult to implement due to the "curse of dimensionality," this invention achieves similar performance with only about 2% of the simulation sample size of the PPO method, reducing training time by approximately 50%. It can achieve group control of up to 512 robots, with group task completion time increasing approximately linearly with scale, while also smoothly increasing speed with speed commands, demonstrating significantly enhanced speed compliance compared to the PPO algorithm. Compared to the PPO algorithm, this invention can navigate through narrower gaps more accurately. It can be directly deployed from simulation to real-world environments without additional training. The navigation algorithm of this invention remains effective even in conditions where global positioning information is unavailable and communication is lacking, without relying on state sharing between robots or external positioning facilities. Thanks to the decoupling design of high-level navigation strategies and low-level motion strategies in the hierarchical control architecture, the system can achieve fully autonomous operation relying only on an onboard depth camera and limited local computing resources. This feature not only simplifies system design but also provides significant advantages in large-scale, low-cost robot swarm deployments.

[0058] The above-described specific implementations can be partially adjusted by those skilled in the art in different ways without departing from the principles and purpose of the present invention. The scope of protection of the present invention is defined by the claims and is not limited to the above-described specific implementations. All implementation schemes within the scope of the claims are bound by the present invention.

Claims

1. A visual navigation method for robot swarms based on differentiable physics and surrogate models, characterized in that, include: Step 1: Construct a high-fidelity simulation environment with differentiable physics simulation, including a quadruped robot; Step 2: Construct a high-level navigation strategy network. Based on the raw image information acquired by the depth camera, use a deep learning model to understand the environmental geometry, identify obstacles and dynamic intelligent agents, and output speed commands to achieve obstacle avoidance, collision avoidance and target navigation. A low-level motion strategy network is constructed to translate velocity commands into joint position and torque commands for the quadruped robot's legs, generating stable and flexible body movements to execute a velocity tracking strategy. Step 3: Construct a differentiable physical proxy model to approximate the backpropagation process of the high-level navigation policy network and the low-level motion policy network in Step 2 that are not differentiable during the training phase; Step 4: Training is performed using a forward-backward asymmetric optimization framework. In the forward execution phase, a high-level navigation strategy network and a low-level motion strategy network are used for inference in the high-fidelity simulation environment constructed in Step 1 to accurately capture the dense contact details in the robot's motion. In the backpropagation phase, gradient updates are performed on the high-level navigation policy network and the low-level motion policy network in the differentiable physical proxy model constructed in step 3 to obtain smooth gradient signals. Step 5: Deploy the high-level navigation strategy network and low-level motion strategy network trained in Step 4 to a real-world experimental scenario to achieve cluster control.

2. The robot swarm visual navigation method based on differentiable physics and surrogate model according to claim 1, characterized in that, Step 1 specifically includes: 1.1 Geometric Modeling: In Blender 3D modeling software, a highly detailed scene model is created based on the scale of the real world; 1.2 Model Export: Export the completed scene model in a universal format (OBJ, USD); 1.3 Scene Loading and Instantiation: In IsaacGym, the exported model files are loaded through the corresponding API interface, and the flexible invocation, randomized placement, and batch instantiation of scenes are implemented in the training code; The simulation environment includes diverse and complex scenes such as open fields, mazes, sheep pens, and forests.

3. The robot swarm visual navigation method based on differentiable physics and surrogate models according to claim 1, characterized in that, The high-level navigation strategy network is based on a convolutional recurrent neural network, which obtains a two-dimensional velocity command consisting of linear velocity and yaw rate based on the depth image and the two-dimensional reference velocity generated in real time from the robot's current position and the target point. The low-level motion strategy network is based on a multilayer perceptron (MLP) and generates joint angle commands based on two-dimensional velocity commands.

4. The robot swarm visual navigation method based on differentiable physics and surrogate model according to claim 3, characterized in that, The convolutional recurrent neural network includes: three layers of LeakyReLU activated convolutional layers with kernel sizes of 2, 3, and 3, and output channels of 32, 64, and 128 respectively; a linear layer; a GRU unit; and a fully connected layer. The depth image is processed by three convolutional layers to extract features, which are then flattened and mapped to 192-dimensional visual features by a linear layer. The reference velocity is mapped to 192-dimensional observation features by a linear layer with a weight scaling factor of 0.

5. The visual features and observation features are added element-wise and input to a 192-dimensional GRU unit in the hidden layer for temporal modeling. The fully connected layer outputs a two-dimensional velocity command. The hidden state of the GRU unit is passed between time steps, enabling the policy to have short-term memory and motion intention inference capabilities.

5. The robot swarm visual navigation method based on differentiable physics and surrogate model according to claim 3, characterized in that, The low-level motion strategy network includes two hidden layers, each with 256 neurons, using the ReLU activation function. Its inputs are the target velocity and yaw angle (4D), phase variable (2D), base angular velocity (3D), base attitude angle (2D), 12 joint angles, and 12 joint velocities output by the high-level navigation strategy network. The output is a 12-dimensional joint angle command. The joint angle command is converted into joint torque by a simplified PD controller, and then mapped into foot force through the foot Jacobian matrix to drive the rigid body model update.

6. The robot swarm visual navigation method based on differentiable physics and surrogate model according to claim 1, characterized in that, The differentiable physical proxy model includes a differentiable particle model and a differentiable rigid body model, which are used to proxy the calculation of the corresponding gradients during backpropagation when training high-level navigation strategies and low-level motion strategies. The differentiable physical proxy model is as follows: ,in: and The outputs of the model at time step k+1 and time step k are respectively. For time step k, and These are surrogate models and real system dynamics, respectively. is the time step k slack variable, representing the residual between the surrogate model and the real system dynamics.

7. The robot swarm visual navigation method based on differentiable physics and surrogate models according to claim 6, characterized in that, The discrete-time update equation for the differentiable particle model is: ,in: For the robot in time step Location, For time step speed, For the time step of integration; The dynamic equations of the differentiable single rigid body model are as follows: , , , ,in: For the robot's position, For speed, Let the azimuth be expressed as a unit quaternion. Angular velocity, The net external torque acting on the rigid body of the robot. Here is the robot's rotational inertia matrix, and the control input is the ground reaction force from its legs. .

8. The robot swarm visual navigation method based on differentiable physics and surrogate model according to claim 1, characterized in that, Step 4 specifically includes: 5.1 Initialization: Initialize the high-level navigation strategy network and the low-level motion strategy network respectively, and set their respective gradient accumulation steps. At the same time, set the initial robot state and global target position to lay the foundation for subsequent collaborative training. 5.2 Forward Interaction and Data Acquisition Loop: In the simulation environment, the high-level navigation strategy outputs the desired motion command based on the depth image and target velocity, while the low-level motion strategy generates specific joint position commands based on the robot's body state. The simulation environment executes the commands and updates the state, while simultaneously calculating the loss functions of the navigation strategy and the motion strategy. 5.3 Gradient Accumulation and Policy Update: A gradient accumulation mechanism is adopted. After the navigation strategy accumulates a set number of steps of loss, the gradient is calculated and the network parameters are updated through a differentiable mass surrogate model. Similarly, after the motion strategy accumulates a corresponding number of steps of loss, the parameters are updated with the help of a differentiable rigid body surrogate model to improve training stability. 5.4 Iterative Convergence: Repeat the forward interaction and policy update loop until the loss functions of the navigation policy and motion policy converge, completing the entire training process.

9. The robot swarm visual navigation method based on differentiable physics and surrogate model according to claim 8, characterized in that, The gradient accumulation mechanism refers to the following: each simulation environment contains 96 robot instances running in parallel. For the high-level navigation strategy, the gradient is calculated based on a differentiable mass surrogate model every 80 simulation steps of accumulated loss, and the network parameters are updated using the AdamW optimizer with an initial learning rate of 0.001, followed by cosine annealing to 0.00001. For the low-level motion strategy, the gradient is calculated based on a differentiable rigid body surrogate model every 20 simulation steps of accumulated loss, also using the AdamW optimizer with an initial learning rate of 0.001, followed by cosine annealing to 0.0001. The total number of training steps is 5000 optimization steps for the motion strategy and 1250 optimization steps for the navigation strategy. Domain randomization is introduced during training: gait frequency is randomly sampled between 1.7 and 2.5 Hz, and the maximum allowable speed is randomly sampled between 0.5 and 1.5 m / s to enhance the robustness and generalization ability of the strategy. The training is performed using a gradient descent method based on differentiable physics. The loss is recorded at each simulation step, and after accumulating a certain number of steps, the gradient is calculated through backpropagation using a surrogate model to update the network parameters. The multi-objective loss function includes: a loss function for a high-level navigation strategy and a loss function for a low-level motion strategy, wherein: the loss function for the high-level navigation strategy... Among them: speed loss Obstacle avoidance loss Collision damage Overspeed loss Acceleration loss Directional loss Loss in the target direction Time loss ; Loss function of low-level motion strategy Among them: velocity tracking loss attitude tracking loss Angular velocity tracking loss loss of fuselage height Gravity projection loss Action regularization loss Torque regularization loss Swing leg trajectory loss , , , , , , , , , , , , , , and As weight.

10. A robot swarm visual navigation system implementing the method of any one of claims 1-9, characterized in that, include: The system comprises a control layer consisting of a navigation strategy module and a motion strategy module, an optimization layer consisting of a mass proxy model and a rigid body proxy model, and a simulation environment module. Specifically: the navigation strategy module extracts visual features and performs temporal decision processing based on depth image information acquired by an onboard depth camera to obtain the desired velocity command; the motion strategy module performs state-action mapping processing based on the desired velocity command output by the navigation strategy module and the robot's body perception information to obtain joint control commands; the simulation environment module performs physical simulation calculations based on the joint control commands, updates the robot's state, and feeds back the simulated depth image and body state to the control layer; the mass proxy model performs differentiable dynamics calculations based on the input and output of the navigation strategy module to obtain the gradient optimization signal for the navigation strategy; the rigid body proxy model performs differentiable rigid body dynamics calculations based on the input and output of the motion strategy module to obtain the gradient optimization signal for the motion strategy; and the optimization layer backpropagates the gradient optimization signal to the control layer to update the network parameters of the navigation strategy module and the motion strategy module.

Images