Navigation robot self-adaptive control system and method for low-friction working condition

Abstract

The invention discloses a navigation robot self-adaptive control system and method for a low-friction working condition, and belongs to the technical field of robot control. The system comprises a sensing module, an embedded processing unit and a motor driving circuit. The embedded processing unit is internally provided with an LSTM-based meta reinforcement learning network and is used for extracting environmental implicit physical characteristics from multi-source heterogeneous observation data and outputting two paths of instructions in parallel: one path is a kinematics instruction corrected by a dynamic inverse solution module by using a dynamic feedforward compensation coefficient and is used for controlling a speed closed loop of an FOC driver; and the other path is a self-adaptive torque clamping coefficient which is deduced and generated based on the environmental friction limit and is used for dynamically setting a current limit threshold value of the FOC driver. According to the invention, a dual-frequency heterogeneous architecture of NPU reasoning and CPU control is adopted, and soft and hard closed loops from data driving perception to physical bottom layer execution are realized. The system can effectively restrain wheel slipping on a low-friction road surface, and the navigation efficiency and robustness of the robot in an extreme environment are remarkably improved.

Description

Technical Field

[0001] This invention belongs to the field of mobile robot control technology, specifically referring to an adaptive control system and method for navigation robots in low-friction conditions that combines meta-reinforcement learning environmental feature inference with embedded low-level actuator feedback coupling. Background Technology

[0002] In the fields of industrial automation and special operations, mobile robots often need to operate in extreme, non-cooperative environments such as copper electrolysis workshops, ice surfaces, and oily floors. These environments typically have extremely low coefficients of surface friction. And sensor noise caused by strong magnetic field interference (such as lidar noise) ).

[0003] Traditional path planning algorithms (such as) Algorithms are typically planned based on geometric constraints and lack the ability to perceive environmental dynamic parameters. Under low-friction conditions, due to insufficient adhesion between the drive wheels and the ground, the rated torque output by the motor often causes wheel slippage, resulting in severe distortion between the odometer data fed back by the encoder and the actual displacement, leading to an extremely low success rate for navigation tasks.

[0004] On the other hand, while conventional deep reinforcement learning algorithms (such as the PPO algorithm) possess a certain degree of environmental adaptability, their fixed policy network parameters prevent them from sensing sudden changes in the physical properties of the environment online. In environments with drastic fluctuations in friction, conventional algorithms often output highly oscillating velocity commands to correct positional errors caused by slippage, resulting in a jagged robot trajectory and significantly increasing task execution time and system energy consumption. Therefore, achieving robust yet efficient adaptive navigation under complex physical constraints and overcoming the technological gap where the perception layer cannot guide the physical flow limitations of the execution layer is a pressing technical problem that needs to be solved. Summary of the Invention

[0005] The purpose of this invention is to provide an adaptive control system and method for navigation robots in low-friction conditions. By deploying a meta-reinforcement learning network in the embedded main control unit, the system infers the physical characteristics of the environment in real time and converts these characteristics into kinematic commands and dynamic current limiting commands for the underlying motor driver. This physically suppresses wheel idling and significantly improves navigation efficiency.

[0006] To achieve the above objectives, the technical solution adopted by this invention is as follows: an adaptive control system for a navigation robot under low-friction conditions, comprising: a perception module for real-time acquisition of the robot's environmental geometry, body posture, and drive wheel motion state information, constructing a multi-source heterogeneous observation vector; an embedded processing unit, which integrates a deep learning inference engine for running a meta-reinforcement learning network based on spatiotemporal context awareness; the embedded processing unit is configured to input the multi-source heterogeneous observation vector into the meta-reinforcement learning network, extract the implicit physical features of the environment through a Long Short-Term Memory (LSTM) unit, and output kinematic commands and physical constraint commands in parallel; a motor drive circuit, including a field-oriented control (FOC) driver, communicatively connected to the embedded processing unit; the kinematic commands include target values ​​of linear velocity and angular velocity after feedforward compensation via the network's internal inverse dynamics module, used to control the speed closed loop of the motor drive circuit; the physical constraint commands include an adaptive torque clamping coefficient generated based on the decoding of implicit physical features of the environment. It is used to dynamically set the saturation threshold of the internal current loop of the motor drive circuit; while performing speed closed-loop tracking, the motor drive circuit limits the torque component of the stator current within the dynamic current limit threshold, thereby achieving anti-slip adaptive control.

[0007] Specifically, the input end of the meta-reinforcement learning network is equipped with a feature decoupling module, which is used to decouple the multi-source heterogeneous observation vector into three independent feature streams: spatial perception stream: containing normalized distance point cloud data collected by lidar, used to characterize the geometric boundaries of the environment; task state stream: containing the target point orientation and thermal imaging sensor readings, used to characterize the navigation task progress; physical fingerprint stream: containing the high-frequency vibration amplitude of the inertial measurement unit (IMU) and the differential data of the wheel speed encoder, used to provide the long short-term memory (LSTM) unit with implicit inference of the friction coefficient and magnetic field interference intensity of the current road surface.

[0008] Specifically, the meta-reinforcement learning network has structurally independent navigation action branches and physical constraint branches at the output of the Long Short-Term Memory (LSTM) units: the navigation action branches adopt a multilayer perceptron structure to map the hidden states output by the LSTM units into action probability distributions and generate kinematic instructions; the kinematic instructions contain dynamic feedforward compensation coefficients identified by the network based on historical observation sequences. The compensation component; the physical constraint branch adopts a parameterized boundary projection structure to map the same implicit state into a scalar form of adaptive torque clamping coefficient. The range of values ​​for this coefficient is: .

[0009] Specifically, the computational logic of the physical constraint branch is as follows: the high-dimensional LSTM hidden state vector is projected into a scalar through a fully connected layer, and the sigmoid activation function is used to map the scalar to a normalized confidence interval; when the hidden state represents an environment under low friction or strong magnetic interference conditions, the adaptive torque clamping coefficient... It automatically reduces current, thereby tightening the current limit boundary of the motor drive circuit.

[0010] Specifically, the dynamic current limiting threshold in the motor drive circuit The calculation formula is ,in The maximum physical current allowed by the motor drive circuit; when the torque command output by the speed loop attempts to exceed the dynamic current limit threshold. At this time, the FOC driver will force the quadrature-axis current to clamp at the threshold and suppress integral saturation of the speed loop.

[0011] The embedded processing unit adopts a dual-frequency heterogeneous control architecture: the meta-reinforcement learning network runs in the neural network accelerator (NPU) and updates the adaptive torque clamping coefficients at the first frequency. And kinematic commands; the communication control of the motor drive circuit runs in the real-time kernel of the central processing unit (CPU), and reads the latest adaptive torque clamping coefficient at a second frequency. The kinematic instructions are written to the driver register; wherein the second frequency is at least 10 times higher than the first frequency.

[0012] An adaptive control method for a navigation robot under low-friction conditions, applied to the adaptive control system of the navigation robot under low-friction conditions, includes the following steps: S1: Acquire multi-source sensor data through a sensing module and perform layer normalization processing to generate multi-source heterogeneous observation vectors; S2: Input the multi-source heterogeneous observation vectors into a meta-reinforcement learning network, and update the latent state features representing the physical properties of the environment through long short-term memory (LSTM) units; S3: Utilize the latent state features to simultaneously generate kinematic commands including dynamic feedforward compensation and an adaptive torque clamping coefficient representing the road surface adhesion limit through parallel decision branches. S4: Send kinematic commands to the speed loop input of the FOC driver, while utilizing the adaptive torque clamping coefficient. Calculate the dynamic current limit threshold and write it to the torque limit register of the FOC driver; S5: The FOC driver drives the motor to run according to the kinematic instructions and monitors the real-time phase current. When the current reaches the dynamic current limit threshold, it triggers hardware current limiting protection.

[0013] Specifically, the training process of the meta-reinforcement learning network includes: constructing a training scenario in a simulation environment that includes random friction coefficients and sensor noise; performing end-to-end training of the network using the proximal policy optimization (PPO) algorithm; introducing a low-friction zone overspeed risk penalty term into the reward function, which is calculated only using the environmental truth value during the offline training phase; and training long short-term memory (LSTM) units to inversely identify slip hysteresis features from the IMU vibration sequence and establish a mapping relationship between them and the feedforward compensation amount through the backpropagation time-backward propagation (BPTT) mechanism.

[0014] A computer-readable storage medium having a computer program stored thereon, the program being executed by a processor, describes an adaptive control method for a navigation robot in low-friction conditions.

[0015] The beneficial effects of this invention are: Significantly improves navigation efficiency: By transforming the "soft" features extracted by the algorithm into "hard" constraints at the motor's bottom layer, this invention reduces the average number of task execution steps from 268.2 steps in the benchmark algorithm to 85.7 steps under extremely slippery conditions with a friction coefficient of 0.38, improving efficiency by 213% and effectively eliminating trajectory oscillation.

[0016] Extremely high robustness: This invention maintains an 88.0% task success rate in extreme environments with both strong noise and low friction, completely solving the problem of traditional geometric programming algorithms failing due to slippage.

[0017] Real-time performance through hardware and software integration: By leveraging the embedded NPU to accelerate inference and directly coupling it with the underlying FOC current loop, millisecond-level dynamic response is achieved, ensuring the safe operation of the robot under physical boundary conditions. Attached Figure Description

[0018] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation on the scope. For those skilled in the art, other related drawings can be obtained based on these drawings without creative effort.

[0019] Figure 1 This is a diagram of the overall hardware architecture of the navigation robot system provided in this embodiment of the invention.

[0020] Figure 2 This is a schematic diagram of a dual-channel closed-loop control logic based on spatiotemporal context awareness provided in an embodiment of the present invention.

[0021] Figure 3 This is a schematic diagram of the underlying physical current limiting closed-loop control principle provided in an embodiment of the present invention. Figure 4 This is a response curve of the nonlinear slip dynamics model under low friction conditions provided in the embodiments of the present invention.

[0022] Figure 5 This is a detailed topology diagram of the STC-MetaNet network provided in an embodiment of the present invention.

[0023] Figure 6 This is a comparison chart of navigation performance under low-friction conditions and benchmark algorithms provided by an embodiment of the present invention. Detailed Implementation

[0024] The present invention will now be described in further detail with reference to the accompanying drawings and specific embodiments.

[0025] Example 1: As Figure 1-6 As shown, an adaptive control system for a navigation robot designed for low-friction conditions includes a sensing module, an embedded processing unit, and a motor drive circuit.

[0026] The perception module is used to collect real-time environmental geometric information (environmental point cloud data), body attitude information (body attitude angular velocity data), and drive wheel motion state information (real-time rotational speed pulse data of the drive wheels), and to construct multi-source heterogeneous observation vectors. .

[0027] The embedded processing unit integrates a neural network accelerator (NPU) for running a meta-reinforcement learning inference model based on a long short-term memory (LSTM) network; the embedded processing unit is configured to process multi-source heterogeneous observation vectors at the current time. Based on the previous hidden state, update the hidden state vector representing the current road surface friction constraint characteristics. Simultaneously, through network training, kinematic commands and physical constraints can be output.

[0028] The motor drive circuit includes a field-oriented control (FOC) driver, whose signal input terminal is connected to the embedded processing unit via a fieldbus; it receives kinematic commands from the embedded processing unit, and the embedded processing unit responds according to the implicit state vector. Calculate the adaptive torque clamping coefficient And based on the adaptive torque clamping coefficient A dynamic current limiting command is generated and sent to the motor drive circuit.

[0029] The motor drive circuit adjusts the torque current component in the stator current of the power motor in real time according to the instruction. The limiting threshold for shaft current.

[0030] Furthermore, the embedded processing unit adopts a dual-rate sampling control strategy, including a slow inference loop running on the NPU and a fast control loop running on the CPU real-time core; The slow inference loop has a frequency of 20Hz and is used to update the hidden state vector. And calculate the adaptive torque clamping coefficient. and kinematic commands; the fast control loop operates at a frequency of no less than 500 Hz to read the latest adaptive torque clamping coefficients via direct memory access (DMA). The kinematic commands are then written into the control register of the motor drive circuit via the CAN-FD bus.

[0031] Furthermore, the specific calculation logic for the dynamic current limiting command is as follows: .

[0032] in, for Shaft dynamic current limit threshold, This is the maximum physical current allowed by the motor drive circuit. When the hidden state vector When the environmental friction coefficient decreases, The value decreases, thus limiting the motor's output electromagnetic torque to not exceed the maximum static friction force of the road surface.

[0033] The hardware structure and software operating environment of this invention will be described in detail below with reference to actual conditions.

[0034] Example 2: Embedded Heterogeneous Hardware Architecture and Signal Link like Figure 1 As shown, the present invention provides an adaptive control system for a navigation robot for low-friction working conditions, which adopts an embedded heterogeneous architecture that integrates "perception-decision-execution".

[0035] The multi-source sensing module includes: LiDAR: Communicates with the main control unit via Ethernet / UDP protocol to transmit 360-dimensional environmental distance point cloud data in real time.

[0036] Inertial Measurement Unit (IMU): Reports triaxial acceleration and angular velocity at a frequency of 200Hz via the SPI bus to sense minute slippage of the fuselage on low-friction surfaces.

[0037] Wheel speed encoder: A hardware peripheral connected to the processor's quadrature pulse input (QEI) to count the drive wheel pulses in real time and compare them with the IMU's integral speed to detect the slip rate.

[0038] Core processing unit: The system uses a high-performance SoC chip (such as Rockchip RK3588), which integrates: CPU (Cortex-A76): Runs a real-time operating system and a fast control loop (500Hz), responsible for low-level communication and bus control.

[0039] NPU (Neural Processing Unit): With a computing power of no less than 6 TOPS, it is specifically designed to deploy the STC-MetaNet algorithm model and perform time-consuming inference tasks (20Hz).

[0040] The underlying execution circuitry includes a FOC (Field-Oriented Control) motor driver. The main control unit and the driver are connected via a high-bandwidth CAN-FD bus. This bus is used not only to send torque commands but also to transmit feedback from within the motor. Shaft (torque component) and The real-time sampling current of the shaft (excitation component) forms a physical closed loop.

[0041] Example 3: Inspection Tasks and Environment Construction in the Electrolysis Workshop This embodiment constructs a high-fidelity testing environment that simulates a hydrometallurgical electrolysis workshop.

[0042] Physical conditions: Set the coefficient of dynamic friction of the ground. (in This indicates a uniform distribution (used to simulate the randomness of oil and water films), and it has a standard deviation of... Random physical disturbances; superposition standard deviation of lidar Gaussian white noise (simulating strong magnetic field interference).

[0043] Nonlinear slip dynamics modeling: To address the slippage between the drive wheels and the ground under low-friction conditions, this embodiment establishes a first-order inertial hysteresis dynamic model in a simulation environment to simulate the real physical motion response. Its mathematical expression is as follows:

[0044] In the formula, for The robot's actual linear velocity at any given moment. This represents the theoretical linear velocity corresponding to the motor command. The coefficient of friction with the current road surface Positively correlated dynamic feedforward compensation coefficient ( ).

[0045] When on a high-friction road surface The robot responds quickly; When on an oily and slippery road surface ( ), The speed is significantly reduced, resulting in the actual speed lagging far behind the commanded speed.

[0046] One of the core innovations of this invention lies in using the LSTM unit in STC-MetaNet to backward identify the dynamic feedforward compensation coefficients from historical observation sequences. And use this coefficient to perform feedforward compensation for kinematic commands.

[0047] Task State Machine (FSM): The system pre-sets several environmental monitoring target point sequences. and charging pile coordinates .

[0048] Inspection mode: The robot needs to arrive sequentially according to the optimal path. When Euclidean distance And speed At that time, thermal imaging acquisition is triggered and the system switches to the next target.

[0049] Recharge Mode: When all environmental monitoring target points are completed or the battery SOC is <20%, the navigation target will be forcibly switched to [target mode]. Implement precise parking.

[0050] Example 4: STC-MetaNet Spatiotemporal Context Aware Network.

[0051] This invention designs a "Spatiotemporal Context Meta-Neural Network" (STC-MetaNet) as a reinforcement learning strategy carrier, deployed in an NPU. Its mathematical model is as follows: Definition and feature decoupling of multidimensional heterogeneous observation space: To address the issue of the discrepancy in scale between lidar data and underlying physical state data, and to enable the network to perceive non-visual features in the environment (such as magnetic field interference and slippage), this invention constructs a multi-source heterogeneous observation vector. And decouple it into three key feature subspaces at the input layer: Spatial perception flow ( ): Contains multi-dimensional (e.g., 360-dimensional) LiDAR point cloud data for sensing geometric boundaries; Task state stream ( ): Includes target distance, azimuth angle, and real-time voltage readings from several (e.g., 8) thermal imaging sensors (used to determine whether the inspection point is effectively covered); Physical fingerprint stream ( The data includes: IMU high-frequency vibration amplitude, wheel speed encoder differential values, and electrolyte area markings. The IMU vibration characteristics are used by STC-MetaNet to implicitly infer the current level of magnetic field interference and road surface slipperiness.

[0052] In this embodiment, a 360-line radar is selected and eight thermal imaging points are set up, thus providing a multi-source heterogeneous observation vector. The total dimension is 375.

[0053] Subsequently, the system introduces LayerNorm to standardize the heterogeneous data and eliminate gradient imbalance:

[0054] In the formula, The normalized feature vectors, The original feature vector is the input to the network; The mean of the input features; The variance of the input features; To prevent extremely small fractions with a denominator of zero, numerical stability is guaranteed; sign This represents the Hadamard Product, an element-wise multiplication operator used to multiply normalized features by a learnable scaling factor. Perform element-wise weighted summation; This is a learnable translation factor. The formula, through layer normalization, eliminates the dimensional differences in the input data, accelerating the network's convergence process.

[0055] Gated feature fusion: By dynamically adjusting the weights of spatial perception and dynamic information using a gating mechanism, fused features are generated. :

[0056] In the formula This refers to the Hyperbolic Tangent Activation Function, which is used to introduce non-linear properties and map the output value to... Interval; symbol The feature concatenation operator is used to concatenate and fuse the processed spatial feature vector with the temporal feature vector along the channel dimension. , , These are the learnable weight matrices for the fusion layer, spatial feature extraction layer, and temporal feature extraction layer, respectively; This is the bias vector of the fusion layer; and These are the normalized temporal and spatial features output from the previous layer, respectively.

[0057] LSTM long-term memory optimization: To capture continuous slippage characteristics, this invention innovatively initializes the LSTM Forget Gate Bias to 1.0. This forces the network to maintain a high “memory rate” in the early stages of training, enabling it to infer the low coefficient of friction of the environment from sensor biases over the past 50-100 steps.

[0058] Furthermore, the STC-MetaNet network utilizes the temporal memory capability of LSTM to learn the implicit cross-modal correlation between "IMU vibration characteristics" and "LiDAR noise". When high-frequency vibrations caused by non-physical collisions are detected in the IMU (i.e., interference characteristics of strong magnetic fields on MEMS gyroscopes), the network automatically reduces the confidence weight of LiDAR data and relies more on odometry and wheel speed feedback, thereby maintaining the stability of positioning and control in areas with strong magnetic interference.

[0059] In this architecture, the hidden state output by the LSTM It not only serves as a carrier of temporal memory but also acts as a 'meta-variable'. Without updating network weights, it characterizes the physical properties of the current environment (such as friction coefficient and disturbance intensity) through the dynamic evolution of activation values, thereby achieving the meta-learning capability of 'in-context adaptation'.

[0060] Parallel decision output mechanism: To achieve decoupled control between motion planning and physical constraints, the STC-MetaNet network designs two parallel decision branches (Dual-Branch Output) after the LSTM temporal feature extraction layer. Navigation Action Branch: This branch uses a multilayer perceptron (MLP) structure as the policy head.

[0061] It contains several layers of neurons and nonlinear activation functions, used to process the high-dimensional hidden states output by the LSTM. Projected onto the robot's continuous motion space.

[0062] This branch outputs statistical parameters characterizing the probability distribution of actions, and generates normalized linear velocity commands via random sampling or a deterministic policy gradient algorithm. ) and angular velocity command ( ).

[0063] This instruction represents the "motor intention" generated by the agent in its current perceptual state in order to maximize long-term cumulative rewards.

[0064] Physical Constraint Branch: This branch adopts the "Parametric Safety Boundary Projection" structure.

[0065] It consists of fully connected layers combined with a sigmoid saturation activation function, used to store hidden states. This high-dimensional eigenvector is reduced in dimensionality to an adaptive torque clamping coefficient in scalar form. (range of values) ).

[0066] Unlike the "driving" output of the action branch, this branch outputs an active "inhibitory signal".

[0067] It quantifies the agent's confidence in the physical adhesion limits of the current environment, thereby constructing a dynamically adjustable physical safety boundary for the motion commands generated by the navigation action branches, ensuring that the final power output always converges within the physical limits of the road friction circle.

[0068] Example 5: Reinforcement learning training mechanism based on Meta-PPO.

[0069] The above network is trained using the Proximal Policy Optimization (PPO) algorithm. The core mechanism includes: Composite reward function.

[0070] To guide robots in balancing "multi-point inspection efficiency" and "low-friction safe driving" in highly dynamic environments, this invention designs a composite reward function that includes sparse rewards and dense guidance. .

[0071] The total reward formula is defined as follows:

[0072] The specific definitions are as follows: Potential-oriented rewards ( ): To guide the robot to approach the current target point (inspection point or charging station) along the shortest path, a dense reward based on distance difference is used:

[0073] in Let be the Euclidean distance between the robot and the target at the current moment, and let be the coefficient. This encourages the robot to move closer to its goal with each step.

[0074] Inspection achievement reward ( ): When the robot enters the effective radius of the inspection point ( When ), sparse rewards are triggered:

[0075] This enhances the agent's understanding of the task of "reaching and covering" specific coordinates.

[0076] Motion smoothing and energy penalty ( ): To suppress control command jitter and reduce system power consumption, the following penalty term is introduced:

[0077] Motion smoothness term: Motion smoothness penalty weight The sudden change in the punishment action command protects the motor gearbox; Energy consumption constraint: Energy consumption penalty weight Penalize excessive linear velocity and angular velocity ; Time penalty term: weight of constant regularization term This encourages robots to complete tasks in the fewest possible time steps.

[0078] Penalty for speeding in low friction zone ): Nonlinear safety constraints were designed to address the slippery and hazardous areas unique to electrolysis workshops. When the robot is in the slippery area and its linear velocity exceeds a safety threshold, a penalty is applied:

[0079] The slippage penalty weighting coefficient in the formula , The current linear velocity, This is the current speed threshold. This forced policy network learns a conditioned reflex of "environmental awareness - deceleration": once it passes... Sensing a low-friction environment, the speed must be actively reduced to avoid penalties, which is consistent with the present invention. The axis current limiting mechanism logically forms a "soft and hard closed loop".

[0080] It should be noted that the above-mentioned risk penalty mechanism based on absolute position ( This is only applied to the 'offline training phase' of the policy network.

[0081] During the offline training phase, the risk penalty term in the reward function ( The negative reward acts as a "supervisory signal," suppressing the output gain of the fully connected layer through a backpropagation algorithm when the robot outputs excessive current on a low-friction surface, causing slippage.

[0082] During this stage, the simulation training platform uses the global environmental truth information it has acquired (including the preset coordinates of the slippery area and the distribution of friction coefficients) to calculate and feed back precise reward and punishment signals in order to supervise and guide the updating of the policy network parameters.

[0083] During the robot's 'online deployment phase' (i.e., the physical operation phase), the robot no longer relies on and does not possess the aforementioned global truth information. At this point, the fully trained STC-MetaNet network has internalized the mapping relationship between environmental features and dynamic responses. It can autonomously assess the current slip risk using only real-time observation data from LiDAR, IMU, and wheel speedometers, through the temporal inference capability of LSTM, and generate corresponding deceleration and current-limiting control commands, thereby achieving independent adaptive control after leaving the training environment.

[0084] Terminal status reward ( ): Success: Complete all checkpoints or reach the finish line. Reward ; Failure: Collision or driving off the map boundary, penalty And terminate the round.

[0085] Full sequence backtracking training: The advantage function is calculated using generalized advantage estimation (GAE), and the network parameters are updated via backpropagation over time (BPTT). This allows the network to understand that the current smooth driving is due to an active deceleration decision made seconds ago.

[0086] Progressive training techniques based on dynamic courses: To address the difficulty of direct training of mobile robots in extremely non-cooperative environments with low friction and strong interference, this invention proposes a three-stage dynamic curriculum training strategy that guides the policy network to convergence by gradually increasing the environmental entropy. Phase 1: Construction of Basic Kinematics Environmental settings: Set the coefficient of dynamic friction of the ground. (High adhesion), no obstacles, and LiDAR noise injection is turned off.

[0087] Training objective: To enable the robot to master basic obstacle avoidance logic and path planning capabilities for 8-point inspection. The stage advancement threshold is set as the average success rate over the most recent 100 episodes. .

[0088] Technical effect: Quickly initializes policy network weights, avoiding ineffective exploration caused by frequent slippage in the early stages of training.

[0089] Phase Two: Strengthening Robustness Against Disturbances Environment setting: Introduce random dynamic obstacles to reduce the ground dynamic friction coefficient to [value missing]. (Simulating conventional ground) and injecting sensor data. Gaussian white noise.

[0090] Training objective: To force the Actor network to adapt to slight sensor drift and dynamic disturbances, and to learn conservative driving strategies under uncertainty.

[0091] Phase 3: Generalization of Extreme Operating Conditions Environment settings: Load a complete high-fidelity model of the electrolysis workshop and set the ground dynamic friction coefficient. (Simulated oil / ice surface), standard deviation of lidar noise .

[0092] Core mechanism: Activating the BPTT training path of the LSTM. At this point, the network is forced to use the memory units of the LSTM to identify the slip trend from the vibration sequence of the IMU.

[0093] Training objective: To learn to collaboratively output compensated kinematic and physical constraint commands, achieving accurate trajectory tracking while actively suppressing slippage at the physical layer until the average reward curve converges. Early Stopping Mechanism: Introducing a monitor to evaluate the training status in real time, when the "inspection completion rate" is met... "and the average reward variance" "After 50 epochs, training is automatically terminated and the best weights (Best Checkpoint) are saved to prevent the network from overfitting."

[0094] Example 6: Low-level physical current limiting and closed-loop control This is a key step in transforming algorithmic "soft" decisions into hardware "hard" constraints (such as...). Figure 2 (as shown) Adaptive factor generation: Every 50ms, the NPU exits the LSTM hidden state. Decoding the confidence level of friction force .

[0095] Specifically, a fully connected layer maps the high-dimensional hidden state vector to a scalar, and then a sigmoid activation function is used to constrain its numerical range. Within the interval, the normalized confidence level characterizes the current environmental friction coefficient.

[0096] Dynamic current loop limitation: The CPU transmits via the CAN-FD bus Write to the FOC driver register and execute the current limiting calculation formula:

[0097] in, As mentioned above Shaft dynamic current limit threshold; This is the adaptive torque clamping coefficient output by the NPU; This is the inherent rated current constant of the motor.

[0098] when (Normal road surface): Unleash full power; when (Ice surface / oil stains detected): Clamp to .

[0099] In addition to LSTM, other neural network models with long-term temporal correlation mining capabilities, such as GRU (Gated Recurrent Unit), TCN (Temporal Convolutional Network), or Transformer, can also be used for temporal inference networks. Any technical solution that infers physical environment characteristics based on historical observation sequences and adjusts the parameters of the underlying actuator accordingly falls within the protection scope of this invention.

[0100] Explicit decoding mechanism for implicit features: In this embodiment, from the hidden state To adaptive torque coefficient The mapping is not a simple linear transformation, but a composite structure of "fully connected layer projection + Sigmoid nonlinear activation".

[0101] Specifically, the system utilizes a fully connected layer to encode the high-dimensional hidden state vector. The projection is a scalar, and its value is then strictly constrained by the Sigmoid activation function. Within the interval, the physical confidence level is defined (i.e., 1 represents complete reliability and 0 represents complete slippage).

[0102] The weight parameters of this decoding structure are not preset by manual rules, but are automatically learned through a trial-and-error feedback mechanism of meta-reinforcement learning (Meta-RL). It should be noted that the calculation logic of the dynamic current limiting command includes, but is not limited to, the linear mapping described above. In other embodiments, the adaptive torque clamping coefficient... With dynamic current limit threshold Mapping can also be achieved using lookup tables or nonlinear fitting curves (such as sigmoid S-curves) to accommodate the nonlinear characteristics of different motor models in the magnetic saturation region. For real-time considerations, this embodiment preferably employs a computationally efficient linear calculation mode.

[0103] Physical effect: This mechanism forces the electromagnetic torque output by the motor. Always satisfied This eliminates the possibility of the drive wheel spinning freely due to exceeding the static friction limit from a physical source.

[0104] Ultimately, during the online deployment phase, the aforementioned decoding structure serves as a fixed "cognitive-control bridge." The LSTM's implicit cognition of the environment is transformed in real-time into explicit underlying adaptive torque clamping coefficients. This mechanism directly applies to the current loop of the FOC driver. This mechanism achieves a closed loop from "data-driven perception" to "physical underlying constraints", thus realizing adaptive anti-slip control that is completely independent of truth value dependence.

[0105] Example 7: Verification of Beneficial Effects 100 field tests were conducted under a stringent simulation environment with a friction coefficient of 0.38 and a noise variance of 0.12: Compared to A*: Due to the lack of dynamic constraints, a sideslip collision occurs at the turn, with a success rate of 0%.

[0106] Compared to Base-PPO: It has no LSTM memory and exhibits frequent "slippage-correction" oscillations, with an average of 268.2 steps.

[0107] This invention: leveraging STC-MetaNet's environmental awareness and Axis current limiting results in smooth trajectory, a success rate of 88.0%, an average number of steps reduced to 85.7 (efficiency improved by 213%), and an average reduction in motor energy consumption of 40%.

[0108] It should be noted that although this embodiment uses the RK3588 chip and LSTM network as examples, the scope of protection of this invention is not limited thereto: The core processing unit can also use computing platforms with AI acceleration capabilities, such as GPU modules, FPGAs (Field Programmable Gate Arrays) or DSPs (Digital Signal Processors); It will be apparent to those skilled in the art that the present invention is not limited to the details of the exemplary embodiments described above, such as the model parameters of the divergent multi-line laser, and that the invention can be implemented in other specific forms without departing from the spirit or essential characteristics of the invention. Therefore, the embodiments should be considered in all respects as exemplary and non-limiting, and the scope of the invention is defined by the appended claims rather than the foregoing description. Thus, all variations falling within the meaning and scope of equivalents of the claims are intended to be included within the present invention. No reference numerals in the claims should be construed as limiting the scope of the claims.

[0109] Furthermore, it should be understood that although this specification describes embodiments, not every embodiment contains only one independent technical solution. This narrative style is merely for clarity. Those skilled in the art should consider the specification as a whole, and the technical solutions in each embodiment can also be appropriately combined to form other embodiments that can be understood by those skilled in the art.

Claims

1. An adaptive control system for a navigation robot operating in low-friction conditions, characterized in that, include: The perception module is used to collect the robot's environmental geometry information, body posture information and drive wheel motion state information in real time, and to construct multi-source heterogeneous observation vectors. An embedded processing unit, integrating a deep learning inference engine, is used to run a spatiotemporally context-aware meta-reinforcement learning network. The embedded processing unit is configured to input the multi-source heterogeneous observation vectors into the meta-reinforcement learning network, extract implicit physical features of the environment using its internal Long Short-Term Memory (LSTM) unit, and output kinematic and physical constraint commands in parallel. A motor drive circuit, including a field-oriented control (FOC) driver, is communicatively connected to the embedded processing unit. The kinematic commands include target values ​​for linear and angular velocities after feedforward compensation via the network's internal inverse dynamics module, used to control the speed closed loop of the motor drive circuit. The physical constraint commands include adaptive torque clamping coefficients generated based on the decoding of implicit physical features of the environment. It is used to dynamically set the dynamic current limit threshold of the internal current loop of the motor drive circuit; while performing speed closed-loop tracking, the motor drive circuit limits the torque component of the stator current within the dynamic current limit threshold.

2. The adaptive control system for a navigation robot in low-friction conditions according to claim 1, characterized in that, The input of the meta-reinforcement learning network is equipped with a feature decoupling module, which decouples the multi-source heterogeneous observation vector into three independent feature streams: spatial perception stream: containing normalized distance point cloud data collected by lidar, used to characterize the geometric boundaries of the environment; task state stream: containing the target point orientation and thermal imaging sensor readings, used to characterize the navigation task progress; physical fingerprint stream: containing the high-frequency vibration amplitude of the inertial measurement unit (IMU) and the differential data of the wheel speed encoder, used for the long short-term memory (LSTM) unit to implicitly infer the friction coefficient and magnetic field interference intensity of the current road surface.

3. The adaptive control system for a navigation robot under low-friction conditions according to claim 1, characterized in that, The meta-reinforcement learning network has structurally independent navigation action branches and physical constraint branches at the output of the Long Short-Term Memory (LSTM) units. The navigation action branches adopt a multilayer perceptron structure to map the hidden states output by the LSTM units into action probability distributions and generate kinematic instructions. The kinematic instructions contain dynamic feedforward compensation coefficients identified by the network based on historical observation sequences. The compensation component; the physical constraint branch adopts a parameterized boundary projection structure to map the same implicit state into a scalar form of adaptive torque clamping coefficient. The range of values ​​for this coefficient is: .

4. The adaptive control system for a navigation robot under low-friction conditions according to claim 3, characterized in that, The specific computational logic of the physical constraint branch is as follows: the high-dimensional LSTM hidden state vector is projected into a scalar through a fully connected layer, and the sigmoid activation function is used to map the scalar to a normalized confidence interval; when the hidden state represents an environment under low friction or strong magnetic interference conditions, the adaptive torque clamping coefficient... It automatically reduces current, thereby tightening the current limit boundary of the motor drive circuit.

5. The adaptive control system for a navigation robot under low-friction conditions according to claim 1, characterized in that, Dynamic current limiting threshold in motor drive circuit The calculation formula is ,in This is the maximum physical current allowed by the motor drive circuit. When the load torque of the drive wheel attempts to exceed the limit due to slippage. At that time, the FOC driver forces the quadrature axis current. Clamping in At this point, the integral effect of the velocity loop is cut off.

6. The adaptive control system for a navigation robot under low-friction conditions according to claim 1, characterized in that, The embedded processing unit adopts a dual-frequency heterogeneous control architecture: the meta-reinforcement learning network runs in the neural network accelerator (NPU) and updates the adaptive torque clamping coefficients at the first frequency. The communication control of the motor drive circuit runs in the real-time core of the central processing unit (CPU), and reads the latest adaptive torque clamping coefficient at a second frequency. And write it into the driver register; wherein the second frequency is at least 10 times higher than the first frequency.

7. An adaptive control method for a navigation robot under low-friction conditions, applied to the adaptive control system for a navigation robot under low-friction conditions as described in any one of claims 1-6, characterized in that, Includes the following steps: S1: Acquire multi-source sensor data through the sensing module and perform layer normalization to generate multi-source heterogeneous observation vectors; S2: Input the multi-source heterogeneous observation vectors into a meta-reinforcement learning network, and update the latent state features representing the physical properties of the environment through Long Short-Term Memory (LSTM) units; S3: Utilize the latent state features to simultaneously generate kinematic commands including dynamic feedforward compensation and adaptive torque clamping coefficients representing the road adhesion limit through parallel decision branches. S4: Send kinematic commands to the speed loop input of the FOC driver, while utilizing the adaptive torque clamping coefficient. Calculate the dynamic current limit threshold and write it to the torque limit register of the FOC driver; S5: The FOC driver drives the motor to run according to the kinematic instructions and monitors the real-time phase current. When the current reaches the dynamic current limit threshold, it triggers hardware current limiting protection.

8. The adaptive control method for a navigation robot under low-friction conditions according to claim 7, characterized in that, The training process of the meta-reinforcement learning network includes: constructing a training scenario in a simulation environment that includes random friction coefficients and sensor noise; performing end-to-end training of the network using the near-end policy optimization PPO algorithm; introducing a low-friction zone overspeed risk penalty term into the reward function, which is calculated only using the environmental truth value during the offline training phase; and training long short-term memory (LSTM) units to inversely identify slip hysteresis features from the IMU vibration sequence and establish a mapping relationship between them and the feedforward compensation amount through the backpropagation time-backward propagation (BPTT) mechanism.

9. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by the processor, the program implements the adaptive control method for navigation robots in low-friction conditions as described in claim 7.

Images