Self-balancing two-wheeled vehicle control system based on autonomous following reinforcement learning

Abstract

The invention relates to the technical field of self-balancing two-wheeled vehicle control, and discloses a self-balancing two-wheeled vehicle control system based on autonomous following reinforcement learning, and the system comprises a sensing and positioning module which is used for collecting and processing multi-source sensor data in real time, and generating vehicle state information; the behavior decision module is used for constructing a driving mode decision for guiding a macroscopic driving behavior; the motion planning and control module is used for generating an actual motion instruction for driving the vehicle; the execution module is used for analyzing the generated actual action instruction and converting the actual action instruction into a driving signal for the self-balancing two-wheeled vehicle; and the vehicle state and dynamics model module is used for managing vehicle state information. A dynamic stability domain predictor is introduced, and bidirectional information coupling between a high-level strategy network and the predictor is established, so that an expected action instruction generated by a high-level decision is constrained in a physically feasible safe action space in real time before execution, and the physical stability of a vehicle in a complex dynamic state is ensured through a closed-loop feedback mechanism.

Description

Technical Field

[0001] This invention relates to the field of self-balancing two-wheeled vehicle control technology, specifically a control system for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning. Background Technology

[0002] With the rapid development of robotics and artificial intelligence, self-balancing two-wheeled vehicles, especially electric vehicles with front and rear dual-wheel layouts, which have advantages of high mobility and small size, have become an important research hotspot in urban transportation and personal travel. To achieve autonomous operation of these vehicles, it is crucial to build a powerful control system. This system needs to perform various driving tasks safely and efficiently in complex dynamic environments, thereby meeting the future intelligent transportation requirements for flexibility and autonomy.

[0003] Existing autonomous control systems for self-balancing two-wheeled vehicles typically employ a traditional layered architecture of perception-decision-planning-control. In this architecture, the sensor module acquires vehicle status and environmental data and transmits it to upper-level modules. The decision-making module allocates macro-level tasks based on preset rules or simple algorithms, such as deciding whether the vehicle should enter automatic following mode or obstacle avoidance mode. Subsequently, the planning and control module generates specific motion trajectories and instructions, which are then executed by the lower-level actuators. For example, in automatic following tasks, the higher-level decision-making module typically calculates the distance to the vehicle ahead and directly outputs a target speed, which is then converted into a drive signal by the lower-level controller. In simple line-following tasks, a preset geometric path is used, and a simple PID controller is employed for tracking.

[0004] However, existing autonomous control systems for self-balancing two-wheeled vehicles suffer from a fundamental disconnect between high-level decision-making and low-level physical constraints when task complexity increases, particularly when involving medium-to-high speed driving, large dynamic maneuvers, or complex environmental interactions. Idealized commands generated at the upper level often fail to adequately consider the vehicle's physical limits at the current speed and posture, making safe execution of commands difficult at the lower level. In high-dynamic conditions such as turning around on narrow roads at medium-to-high speeds, traditional path models cannot guarantee the physical feasibility of the trajectory and the stability of the vehicle, forcing the system to slow down or stop to complete the task, severely impacting efficiency and smoothness. Furthermore, existing systems typically rely on preset rules and fixed parameters, lacking self-learning and adaptive capabilities, making it difficult to make intelligent and smooth decisions when facing unpredictable dynamic environments (such as complex pedestrian traffic or sudden road conditions). Therefore, this invention provides a control system for self-balancing two-wheeled vehicles based on autonomous following reinforcement learning to address the shortcomings of existing technologies. Summary of the Invention

[0005] To address the shortcomings of existing technologies, this invention provides a control system for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning. This system solves the problem that existing self-balancing two-wheeled vehicles struggle to balance following performance, path smoothness, and vehicle physical stability in dynamic conditions such as turning around on narrow roads.

[0006] To achieve the above objectives, the present invention provides the following technical solution:

[0007] The first aspect of this invention provides a control system for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning, comprising:

[0008] The perception and localization module is used to collect, process, and fuse multi-source sensor data in real time to generate vehicle state information for the self-balancing two-wheeled vehicle and the motion state of the target it is following. Vehicle state information includes, but is not limited to, vehicle attitude, linear velocity, and road boundary information.

[0009] The behavior decision module is used to construct driving mode decisions to guide macro driving behavior based on the generated vehicle status information and the motion state of the target being followed. The driving mode decisions include autonomous following mode and intelligent decision mode to cope with different driving scenarios.

[0010] The motion planning and control module is used to generate actual action commands for driving the vehicle based on the driving mode decision constructed by the behavior decision module. The actual action commands are the target linear acceleration and target angular velocity of the self-balancing two-wheeled vehicle.

[0011] The execution module is used to receive the actual action commands generated by the motion planning and control module, and parse and convert them into drive signals for the physical execution mechanism of the self-balancing two-wheeled vehicle.

[0012] The vehicle state and dynamics model module is used to centrally manage the vehicle state information and provide a nonlinear dynamics model of the self-balancing two-wheeled vehicle. This dynamics model describes the coupling relationship between the vehicle's tilting, steering, and driving behaviors, providing physical constraints for the planning process of the motion planning and control module.

[0013] Preferably, when the driving mode decision is autonomous following mode, the system adopts a hierarchical collaborative control scheme. This scheme aims to achieve excellent following performance while ensuring the vehicle's dynamic stability by establishing bidirectional information coupling between high-level driving intentions and low-level physical safety guarantees. This scheme is implemented collaboratively by a high-level policy network unit, a dynamic stability domain predictor unit, and an action projection and feedback unit.

[0014] Among them, the high-level policy network unit uses deep reinforcement learning technology to generate an ideal, physically unconstrained desired action instruction based on environmental information related to different following tasks (such as vehicle-to-vehicle or vehicle-to-person following).

[0015] The dynamic stability domain predictor unit is a key technological innovation. Its input not only includes the vehicle's current physical state but also cleverly couples desired action commands from the higher-level policy network unit, enabling it to proactively predict an action space that ensures the vehicle's dynamic safety. This safe action space defines the upper and lower limits of the linear acceleration and angular velocity that the vehicle can execute under the current physical state.

[0016] The action projection and feedback unit acts as a bridge connecting high-level intentions and low-level reality. It projects the desired action instructions generated by the high-level unit into the safe action space to generate the final actual action instructions, thus ensuring the physical safety of any executed instructions. Furthermore, this unit generates a stability margin penalty signal by calculating the deviation between the desired action and the actual action, and feeds it back to the high-level policy network unit. This two-way feedback mechanism guides the network to learn safer policies that are closer to physical limits during subsequent training.

[0017] Preferably, the formula for calculating the stability margin penalty is:

[0018] ;

[0019] in, and They represent the times at time 1 and 2 respectively. The lower and upper limits of linear acceleration in the safe action space output by the dynamic stability domain predictor unit; and They represent the times at time 1 and 2 respectively. The lower and upper limits of angular velocity in the safe action space output by the dynamic stability domain predictor unit; and They represent the times at time 1 and 2 respectively. The final actual linear acceleration and actual angular velocity are measured. and They represent the times at time 1 and 2 respectively. The desired linear acceleration and desired angular velocity output by the high-level policy network unit; To stabilize the punishment signal; It is a very small positive number that prevents the denominator from being zero.

[0020] Preferably, when the driving mode decision is set to intelligent decision-making mode, the intelligent decision-making mode includes obstacle avoidance and U-turn on narrow roads. By utilizing the path splicing and smoothing unit and the path tracking control unit, a U-turn function that conforms to dynamic constraints at medium and high speeds is achieved. The path splicing and smoothing unit, based on the dynamic model provided by the vehicle state and dynamics model module, solves an optimal control problem that includes dynamic constraints, road boundary constraints, actuator saturation constraints, and stability constraints, planning and generating a globally continuous, smooth, and optimal spatiotemporal trajectory that conforms to the vehicle's motion laws. This trajectory specifies the vehicle's position, speed, and heading at each moment. The path tracking control unit tracks this trajectory in real time, calculates and generates actual action commands to precisely guide the vehicle to complete the U-turn. This path planning and tracking scheme ensures the controllability and safety of the vehicle under complex maneuvering conditions.

[0021] A second aspect of the present invention provides a control method for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning, the method being executed in the control system and comprising the following steps:

[0022] S100: Perform state perception and information fusion to obtain vehicle state information of the self-balancing two-wheeled vehicle and the motion state of the target being followed.

[0023] S200 makes autonomous decisions on driving modes. Based on the vehicle status information of the self-balancing two-wheeled vehicle and the motion status of the target being followed, it decides on the current driving mode in autonomous following mode and intelligent decision mode. Autonomous following mode includes vehicle-to-vehicle following and vehicle-to-person following. Intelligent decision mode includes obstacle avoidance and U-turn on narrow roads.

[0024] S300: When the driving mode is autonomous following mode, the desired action command is generated by the high-level policy network unit, the safe action space is predicted by the dynamic stability domain predictor unit, and the desired action command is projected into the safe action space to generate the actual action command.

[0025] S400: When the driving mode is intelligent decision-making mode, an optimal spatiotemporal trajectory is planned based on the dynamic model of the self-balancing two-wheeled vehicle, and actual action commands are generated according to the optimal spatiotemporal trajectory.

[0026] The S500 performs instruction parsing and physical execution, receives the generated actual action instructions, and converts them into drive signals for the self-balancing two-wheeled vehicle.

[0027] This invention provides a control system for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning. It has the following beneficial effects:

[0028] 1. This invention introduces a dynamic stability domain predictor and establishes bidirectional information coupling between the high-level policy network and the predictor, so that the expected action instructions generated by the high-level decision are constrained in real time within the physically feasible safe action space before execution. This closed-loop feedback mechanism ensures the physical stability of the vehicle under complex dynamics. Furthermore, by feeding the stability penalty signal back to the reinforcement learning network, it guides the learning of a safer and more conservative driving strategy, thereby reducing the risk of instability caused by model errors or external disturbances.

[0029] 2. This invention employs an optimal control method based on a vehicle dynamics model to plan a globally smooth spatiotemporal trajectory that satisfies dynamic constraints at medium to high speeds. This planning process uses the dynamics model, road boundaries, and stability constraints as hard conditions, ensuring that the generated trajectory is geometrically and physically executable. Combined with path tracking using model predictive control, this invention can guide vehicles to complete dynamic U-turns in confined spaces, thus enhancing the system's functional boundaries and application scope.

[0030] 3. Employing reinforcement learning technology enables the high-level policy network to autonomously optimize its driving strategy through continuous interaction and feedback with the environment. This learning capability allows the system to intelligently adjust its following behavior and make smooth and safe driving decisions based on dynamic changes in different scenarios, such as complex vehicle-to-vehicle following or precise vehicle-to-pedestrian following tasks. This autonomous decision-making and adaptive capability reduces the extensive manual parameter tuning work required by traditional control methods, thereby providing users with a smooth and reliable driving experience. Attached Figure Description

[0031] Figure 1 This is a system architecture diagram of the present invention;

[0032] Figure 2 This is a schematic diagram of the finite state machine for the behavior decision module of the present invention;

[0033] Figure 3 This is a flowchart of the method steps of the present invention. Detailed Implementation

[0034] The technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0035] See attached document Figure 1 - Appendix Figure 2 , Figure 1 This is a schematic diagram of a control system architecture according to an embodiment of the present invention. Figure 2This is a schematic diagram of the finite state machine for the behavior decision module. This invention provides a control system for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning. This system is deployed in the onboard computing unit of the self-balancing two-wheeled vehicle and is used to realize advanced driving functions such as autonomous following in complex environments and U-turns on narrow roads at medium to high speeds. The control system for the self-balancing two-wheeled vehicle based on autonomous following reinforcement learning in this embodiment includes: a perception and localization module 100, a behavior decision module 200, a motion planning and control module 300, an execution module 400, and a vehicle state and dynamics model module 500.

[0036] The perception and positioning module 100, serving as the system's data input, integrates various onboard sensors, such as an inertial measurement unit (IMU), wheel speed encoder, lidar, millimeter-wave radar, vision sensors, and an ultra-wideband (UWB) receiver for specific scenarios. This module is responsible for real-time acquisition and processing of raw sensor data to generate structured information about the vehicle itself and its external environment. Its outputs include: high-precision vehicle positioning information (position, heading), motion state (vehicle speed, acceleration), attitude information (vehicle tilt angle, angular velocity), and target information from the external environment. Depending on the application scenario, this module can provide specific target information, such as:

[0037] For vehicle following scenarios: Provides the position, speed, and category of the target vehicle ahead.

[0038] For pedestrian following scenarios: the UWB receiver provides the precise location and movement status of authorized pedestrians.

[0039] The vehicle state and dynamics model module 500 is used to centrally manage and maintain the real-time state of the vehicle and provide computational support for the vehicle dynamics model. This module receives vehicle state information from the perception and localization module 100, forming a unified and consistent vehicle state vector. Simultaneously, this module internally embeds a nonlinear dynamics model of a self-balancing two-wheeled vehicle. This model describes the coupling relationship between the vehicle's tilting, steering, and driving behaviors, providing an accurate physical model basis for dynamic stability domain prediction and U-turn trajectory planning in the motion planning and control module 300.

[0040] The behavior decision module 200 is the high-level decision-making center of the system. This module receives external environmental information from the perception and positioning module 100 and the current vehicle state from the vehicle state and dynamics model module 500. Based on preset driving tasks (such as navigation instructions) and an understanding of the current traffic scene, it determines the macro-driving behavior that the vehicle should perform. For example, when the system identifies a target vehicle that needs to be followed ahead, the behavior decision module 200 will output an autonomous following mode instruction; when the system detects a narrow dead end ahead or receives a U-turn navigation instruction, the module will output a narrow road U-turn mode instruction. In addition, this invention also supports a specific autonomous following mode application scenario: when the system detects an authorized pedestrian wearing a wristband nearby through UWB positioning and identifies a following task, the module will also output an autonomous following mode instruction, but will simultaneously specify the following target as the pedestrian. When the target ahead is moving or stopped somewhere, the current vehicle issues a decision instruction (follow, emergency stop), and its output mode instruction is sent to the motion planning and control module 300.

[0041] The motion planning and control module 300 is the core computational module for implementing specific driving operations. Based on the mode commands issued by the behavior decision module 200, this module activates corresponding sub-functional modules. In autonomous following mode, this module employs a controller based on hierarchical constraint reinforcement learning, which includes a high-level policy network and a dynamic stability domain predictor, to generate real-time control commands that balance following efficiency and driving safety. In narrow-road U-turn mode, this module activates a path planner, generating a space-time optimal trajectory for narrow-road U-turns that meets dynamic constraints at medium to high speeds, based on the model provided by the vehicle state and dynamics model module 500. Regardless of the mode, the motion planning and control module 300 ultimately outputs a uniformly formatted motion control target, such as the desired linear acceleration and angular velocity.

[0042] The execution module 400, as the final output of the self-balancing two-wheeled vehicle control system, is responsible for converting the abstract control commands from the upper-level modules into direct control of the vehicle hardware. This module receives motion control targets from the motion planning and control module 300 and, through the built-in low-level controller (such as a PID controller or torque controller), parses them into precise control signals for the drive wheel motors and steering actuators, such as pulse width modulation (PWM) signals or bus messages, thereby directly driving the vehicle to complete actions such as acceleration, deceleration, and steering.

[0043] During the operation of the self-balancing two-wheeled vehicle control system, the aforementioned modules work collaboratively. The perception and positioning module 100 continuously updates environmental and self-state information. The behavior decision-making module 200 makes high-level decisions based on this information. The motion planning and control module 300 plans and calculates specific motion commands based on the decision results. Finally, the execution module 400 implements the commands. The vehicle state and dynamics model module 500 provides crucial state and model data support for decision-making and planning throughout this process, forming a complete closed loop of information perception, decision-making, planning, and control.

[0044] In this embodiment of the invention, the perception and positioning module 100 and the vehicle state and dynamics model module 500 work together to form the information foundation for all subsequent intelligent decisions and planning. Their goal is to provide the system with accurate, real-time, and multi-dimensional information about its own state and the external environment.

[0045] The sensing and positioning module 100 includes multiple sensor information processing units. In one embodiment, the module includes an inertial measurement unit (IMU) data processing unit, a wheel speed encoder data processing unit, a radar point cloud processing unit, a visual image processing unit, and a position and distance monitoring unit.

[0046] The inertial measurement and processing unit (IMU) processes raw data from IMU sensors, which typically include a three-axis accelerometer and a three-axis gyroscope. This unit outputs real-time attitude information of the self-balancing two-wheeled vehicle at a high frequency (e.g., 200 Hz) by integrating, filtering (e.g., using a Kalman filter), and calculating attitude (e.g., using a quaternion method) from the raw acceleration and angular velocity data. Specifically, the attitude information includes at least the vehicle roll angle (i.e., tilt angle). Pitch angle and angular velocity at these two angles This information serves as direct input for dynamic stability domain prediction and balance control.

[0047] The wheel speed encoder data processing unit is responsible for acquiring and parsing the encoder signals installed on the drive wheels. By counting and calculating the number of pulses output by the encoder per unit time, this unit can accurately obtain the real-time rotational speeds of the left and right drive wheels. Based on the rotational speeds of the two wheels and the known wheel radii, the vehicle's linear velocity and yaw angular velocity can be further calculated. The linear velocity output by this unit is one of the core state variables in the vehicle state and dynamics model module 500.

[0048] The radar point cloud processing unit processes 3D point cloud data acquired from lidar sensors (e.g., 16-line or 32-line lidar). This unit performs ground point filtering, clustering, and target identification on the raw point cloud. The point cloud is segmented into independent object clusters using a clustering algorithm (e.g., DBSCAN), and then features are extracted and classified for each object cluster to identify obstacles such as vehicles, pedestrians, and road curbs.

[0049] For vehicles identified as following targets ahead, the unit accurately calculates their longitudinal distance, lateral distance, size, and heading angle relative to the vehicle by fitting bounding boxes to their point cloud clusters. Simultaneously, by employing target tracking algorithms across consecutive frames (such as ICP algorithms based on point cloud matching or multi-target tracking filters), the instantaneous velocity vector of the target is calculated, and its linear velocity is decomposed from it. This information constitutes the main part of the state required by the high-level policy network in the autonomous follow mode.

[0050] The visual image processing unit processes image data acquired from the forward-facing camera. This unit utilizes deep learning object detection networks (such as the YOLO series models) to identify elements in the image, such as vehicles and lane lines. For autonomous following tasks, this unit can supplement or redundancy the LiDAR, providing target vehicle identification and distance estimation. Furthermore, this unit identifies the boundary of the current driving lane through lane line detection algorithms, providing the system with a basis for lateral control and information to determine road width, the latter being a crucial prerequisite for triggering the narrow-road U-turn mode.

[0051] The position and distance monitoring unit is specifically designed to process signals from UWB receivers. In one embodiment, multiple UWB receivers are installed on the vehicle, while the target (e.g., a pedestrian) wears a UWB transmitting wristband. This unit uses high-precision ranging and positioning algorithms to calculate the precise three-dimensional coordinates of the wristband-wearing target relative to the vehicle in real time, including longitudinal distance, lateral offset, and height information. Its operating principle is based on technologies such as Time Difference of Arrival (TDOA) or Angle of Arrival (AoA), which overcomes multipath effects and provides centimeter-level positioning accuracy in complex environments. This unit can also calculate the target's real-time motion state, including velocity and acceleration, using position data from consecutive frames. This precise distance, lateral offset, and relative velocity information measured by UWB is the core input to the high-level policy network in vehicle-to-pedestrian scenarios, ensuring that the vehicle can accurately follow the pedestrian when they move and stop when they stop.

[0052] The perception and localization module 100 also includes a multi-sensor fusion localization unit. This unit utilizes extended Kalman filter (EKF) or particle filter algorithms to fuse localization information from IMU, wheel speed encoder, and GPS / RTK (if equipped), and combines it with point cloud maps or visual feature maps generated by LiDAR SLAM or visual SLAM technologies to achieve real-time vehicle localization on a high-precision map. The output localization information includes the vehicle's position in the global coordinate system. and heading angle High-precision positioning is the foundation for performing global path planning functions such as U-turns on narrow roads.

[0053] The vehicle state and dynamics model module 500 is responsible for integrating and managing the discrete information output by the perception and localization module 100, forming a unified vehicle state vector that is synchronously updated in each control cycle. In one embodiment, the state vector At least include:

[0054] ;

[0055] In the formula, Indicates at time The vehicle state vector; This indicates the vehicle's lateral position in the global coordinate system; This indicates the vehicle's longitudinal position in the global coordinate system; This represents the vehicle's yaw angle, which is the angle between the vehicle's direction of travel and the X-axis of the global coordinate system. This indicates the vehicle's current linear velocity, which is the magnitude of the velocity of the vehicle's center of mass in the forward direction. The roll angle represents the vehicle's tilt angle, which is the angle at which the vehicle tilts around the axis of travel. This is a key variable for self-balancing two-wheeled vehicles to maintain balance. This indicates that the vector may also contain other physical quantities related to the vehicle's state, such as angular velocity and angular acceleration.

[0056] This module ensures that all other modules within the system access the same conflict-free vehicle status data at all times. Furthermore, this module internally stores the nonlinear dynamics equations of the self-balancing two-wheeled vehicle, which describe the driving torque applied to the wheels. With vehicle state change rate The relationship between them can be simplified as follows:

[0057] ;

[0058] in, It is a nonlinear function, the specific form of which is determined by physical parameters such as the vehicle's mass, center of mass position, and moment of inertia. This is the driving torque. (This is from the dynamic model.) It is invoked by the offline training of the dynamic stability domain predictor and the narrow road U-turn path planner to ensure that the generated control strategy and planned path conform to the physical motion law of the vehicle.

[0059] The behavior decision module 200 autonomously determines the most suitable driving mode based on the current driving environment, task objectives, and the vehicle's own status, and generates corresponding instructions to send to the motion planning and control module 300. In this embodiment, the behavior decision module 200 is mainly responsible for making decisions and switching between autonomous following mode and narrow road U-turn mode.

[0060] The behavior decision module 200 can be implemented based on a finite state machine (FSM). This state machine contains at least the following states: initialization state, autonomous following state, narrow-path U-turn preparation state, narrow-path U-turn execution state, and task completion state. The module internally maintains a current state variable and transitions between these states according to a set of predefined rules.

[0061] To enable state judgment and transition, the behavior decision module 200 includes a scene understanding unit and a task management unit.

[0062] The scene understanding unit is responsible for semantic analysis and interpretation of the fused information from the perception and localization module 100. This unit receives road geometry information, obstacle information, and location information, and determines whether the current scene meets the triggering conditions for a specific driving mode. For example, for the triggering of the narrow road U-turn mode, the scene understanding unit continuously monitors the following set of conditions:

[0063] Road width conditions: This unit receives lane line or road boundary information from the vision or lidar processing unit and calculates the width of the currently drivable area. When the width meets preset conditions, for example... (in For vehicle body length, If the coefficient is 2.5, then the narrow road condition is considered to be met.

[0064] Forward path termination condition: This unit analyzes navigation path information or perceived forward environment. When it detects that the road ahead is a dead end or the end of the global navigation path has been reached and a return is required, the path termination condition is considered met.

[0065] Dynamic environmental safety conditions: This unit assesses the presence of dynamic obstacles around the vehicle. Only when there are obstacles within the area covered by the vehicle's rear and U-turn path, within the estimated U-turn time window... The safety condition is considered met only when there are no other traffic participants in the area. When multiple conditions are met simultaneously, the scene understanding unit generates a U-turn request signal.

[0066] Switching from Autonomous Following to Narrow Road U-Turn: When the vehicle is in autonomous following mode, the task management unit continuously listens for "narrow road U-turn requests" from the scene understanding unit or external user commands. Once a valid request is received, and the vehicle's current speed is below a preset mode switching threshold, the task management unit will switch to autonomous following. The task management unit will then trigger a state transition, switching the system state to the narrow road U-turn preparation state and sending a command to the motion planning and control module 300 to enter the U-turn preparation state.

[0067] Switching back to autonomous mode from a narrow road U-turn: When the vehicle is performing a narrow road U-turn, the task management unit waits for a "U-turn complete" signal from the motion planning and control module 300. Once this signal is received, indicating that the physical U-turn has been completed, the task management unit switches the system state back to autonomous following mode or task complete mode, so that the vehicle can continue to perform subsequent following tasks or stop and wait.

[0068] The decision output of the behavior decision module 200 is a structured instruction. This instruction not only contains the command for mode switching but may also include some key parameters required for that mode. For example, when triggering the narrow road U-turn mode, its output structured instruction... It can be defined as:

[0069]

[0070] in, The field clearly defines the target driving mode; For at any time Structured instructions output by the behavior decision module 200; The field is a set of parameters that contains the initial conditions required by the motion planning module for path planning, such as the current road width. and recommended cornering speed This structured instruction structure ensures clear and accurate information transmission between the decision-making and planning control layers. Through this rule-based and state machine-based implementation, the behavioral decision module 200 is able to make reliable, predictable, and safe macro-behavioral decisions in complex driving environments.

[0071] When the behavior decision module 200 determines that the vehicle has entered autonomous following mode, the motion planning and control module 300 will activate this hierarchical collaborative control scheme. By decoupling the control task into high-level human-like driving intention generation and low-level dynamic physical safety assurance, and establishing bidirectional information coupling between the two, intelligent, smooth, and absolutely safe driving decisions can be achieved.

[0072] This solution is specifically implemented through the collaborative efforts of a high-level policy network unit, a dynamic stability domain predictor unit, and an action projection and feedback unit. It can support various following application scenarios, such as a vehicle following another vehicle or a vehicle following a pedestrian.

[0073] The high-level policy network unit functions to simulate an experienced rider, focusing on understanding the following scenario and making smooth, macro-level driving intentions that align with driving logic. This unit is specifically implemented as a deep reinforcement learning-based neural network, for example, employing an actor-critic architecture.

[0074] The input to this network unit is a processed high-level state vector from the sensing and localization module 100. This vector eliminates physical quantities directly related to the vehicle's underlying balance, containing only the environmental information needed to perform the following task. The composition of this vector can be flexibly adjusted according to different following scenarios.

[0075] When the target being followed is a vehicle ahead: Input vector Include (Longitudinal distance from the vehicle in front) (Relative velocity) (lateral trajectory offset) and (Relative heading angle). In this scenario, the perception and localization module 100 primarily acquires the target vehicle's state through LiDAR, visual sensors, etc. This vector eliminates physical quantities directly related to the vehicle's underlying balance and only contains environmental information required to perform the following task:

[0076]

[0077] in, The longitudinal distance to the vehicle in front. For relative velocity, This is a lateral trajectory offset. This is the relative heading angle.

[0078] When following a pedestrian wearing a UWB wristband: the sensing and positioning module 100 calculates the precise relative position and distance of the target wristband to the vehicle in real time using a UWB receiver installed on the vehicle. At this time, the input variables... Include (Real-time distance to pedestrians measured by UWB) (Lateral offset measured by UWB) and (Relative speed). This UWB-based following mode enables precise following, where the vehicle moves only when the pedestrian moves, making it particularly suitable for low-speed scenarios such as parking lots and parks. In this mode, the system continuously monitors UWB signals to determine whether a pedestrian is moving and adjusts the vehicle's start / stop status accordingly.

[0079] Regardless of the following scenario, the output of this network unit is an ideal, physically unconstrained desired action vector. :

[0080] ;

[0081] in, For the desired vehicle linear acceleration, This represents the desired vehicle steering angular velocity.

[0082] To guide the network unit in learning a strategy that balances efficiency and boundary awareness, the reward function used for its training... Designed as a composite function, with task rewards And the stability margin penalty of feedback constitute:

[0083] ;

[0084] in, This represents the weighting coefficient for the penalty item. Task Reward This function, responsible for guiding the network to learn accurate and smooth following behavior, can be composed of multiple weighted sub-items depending on the following scenario and task objective. For example, in a vehicle-following-pedestrian scenario, the reward function can additionally penalize distance and speed errors using a Gaussian potential function to ensure the vehicle can stably maintain a preset safe distance from the pedestrian and achieve smooth start-stop. (Stability margin penalty) This is a crucial feedback signal.

[0085] The dynamic stability domain predictor unit functions as a proactive safety guardian, providing the vehicle with a safe operating space that will not become unstable under the current physical conditions. This unit is specifically a pre-trained, offline deep neural network or Gaussian process regression model.

[0086] A key feature of this predictor unit lies in its input. It receives input not only from the vehicle state and dynamics model module 500... Vehicle physical state vector at any moment (in Current vehicle speed The angle of inclination. (For tilt angular velocity), it also receives the desired action vector output from the higher-level policy network unit. Its complete input vector is formed by concatenating the two:

[0087] ;

[0088] in, Indicates at time The complete input vector to the predictor; Indicates at time The vehicle's physical state vector; Indicates at time The expected action vector output by the high-level policy network unit.

[0089] By controlling the driving intentions of higher levels By incorporating input, the predictor can anticipate the vehicle's dynamic trends after executing the intention, thereby outputting a more forward-looking and dynamically changing safety action space. :

[0090] ;

[0091] in, Indicates at time Minimum and maximum values ​​of linear acceleration that can be safely executed; Indicates at time The minimum and maximum angular velocities that can be safely executed.

[0092] This space defines the upper and lower limits of linear acceleration and angular velocity that the vehicle can execute at the current moment to ensure dynamic stability.

[0093] The Motion Projection and Feedback Unit serves as a bridge connecting high-level intentions with low-level physical reality, responsible for enforcing safety constraints and generating critical feedback signals. This unit comprises two sub-functions.

[0094] The first sub-function is action projection. This function receives the desired action output by the higher-level policy network unit. and the safe action space output by the dynamic stability predictor unit It projects desired actions that might exceed the safety boundary into the safety space through a deterministic pruning operation, generating the actual actions that are ultimately executed by the execution module 400. .

[0095] ;

[0096] ;

[0097] in, and They represent the times at time 1 and 2 respectively. The final actual linear acceleration and actual angular velocity are measured. and They represent the times at time 1 and 2 respectively. The desired linear acceleration and desired angular velocity output by the high-level policy network unit; and They represent the times at time 1 and 2 respectively. The lower and upper limits of linear acceleration in the safe action space output by the dynamic stability domain predictor unit; and They represent the times at time 1 and 2 respectively. The lower and upper limits of angular velocity in the safe action space output by the dynamic stability domain predictor unit; This indicates that the smaller of the two input values ​​is selected; This indicates that the larger of the two input values ​​is selected.

[0098] This operation ensures that regardless of the high-level strategy decisions, the final physical execution instructions always remain within an absolutely safe feasible domain.

[0099] The second sub-function is the calculation of the stability margin penalty. This is the core of achieving bidirectional information coupling. After completing the action projection, this unit calculates the desired action. With actual actions The deviation between them is quantified into a stable penalty signal. The magnitude of this signal reflects the extent to which high-level strategic decisions touch upon or even attempt to exceed current physical limits. Its calculation formula is as follows:

[0100] ;

[0101] in, and They represent the times at time 1 and 2 respectively. The lower and upper limits of linear acceleration in the safe action space output by the dynamic stability domain predictor unit; and They represent the times at time 1 and 2 respectively. The lower and upper limits of angular velocity in the safe action space output by the dynamic stability domain predictor unit; and They represent the times at time 1 and 2 respectively. The final actual linear acceleration and actual angular velocity are measured. and They represent the times at time 1 and 2 respectively. The desired linear acceleration and desired angular velocity output by the high-level policy network unit; To stabilize the punishment signal; It is a very small positive number that prevents the denominator from being zero. The calculated... The signal is then fed back to the reward function of the higher-level policy network unit.

[0102] When the behavior decision module 200 issues a command to enter the narrow road U-turn mode, the motion planning and control module 300 will activate this function to generate and execute a U-turn path for the self-balancing two-wheeled vehicle that meets dynamic constraints at medium to high speeds (e.g., 20-40 km / h) and can be completed in a narrow space (e.g., 1.5 times the width of the vehicle body).

[0103] This function is specifically implemented by the path generation unit, the path splicing and smoothing unit, and the path tracking control unit working together.

[0104] The path generation unit is responsible for planning the core geometric path for the U-turn maneuver. Considering the large-angle cornering characteristics of self-balancing two-wheeled vehicles at medium to high speeds, traditional paths based on Ackerman steering geometry (such as spiral curves) are no longer applicable. In this embodiment, the unit generates the skeleton of the U-turn path by using a segmented combination of geometric curves. This path skeleton consists of three basic curves: an entry transition curve, a center arc curve, and an exit transition curve.

[0105] Entry / exit transition curves: Parametric Bézier curves or polynomial spirals are used to smoothly connect straight driving segments and central circular arc segments, allowing the curvature of the path to change continuously and avoiding abrupt changes. This is crucial for maintaining vehicle stability at high speeds.

[0106] Central circular arc curve: its radius It is based on the current road width It is determined by the dynamic constraints of the vehicle's minimum turning radius to ensure that the core turning portion can be completed within the road boundary.

[0107] The path splicing and smoothing unit receives the segmented geometric path skeleton output by the path generation unit and optimizes it to generate a globally continuous, smooth spatiotemporal trajectory that conforms to the vehicle dynamics model.

[0108] This unit first discretizes the path into a series of path points. ,in Represents the discretized first... 1 path point; Indicates the first The coordinates of each path point in two-dimensional space are then determined. An objective function is then constructed. This function aims to minimize the overall cost of the entire U-turn process, such as total time, energy consumption, or passenger discomfort. A specific objective function. It could be:

[0109] ;

[0110] in, It is the total turning time; It is the vehicle's acceleration; It is jerk; It is the path curvature; These are the weighting coefficients for each cost; This represents the time-dependent component. Minimizing jerk aims to improve ride smoothness, while minimizing the rate of curvature change ensures smooth steering.

[0111] During the optimization process, the unit imposes a series of constraints that ensure the generated trajectory is physically feasible:

[0112] Vehicle dynamics constraints: These constraints relate the vehicle state to the nonlinear dynamics model provided in the dynamics model module 500. (in For control inputs, such as driving torque and steering torque; Indicates at time The vehicle state vector is used as an equality constraint to ensure that every point on the trajectory satisfies the vehicle's motion law.

[0113] Road boundary constraints: require all points on the trajectory to be bounded by the road boundary constraints. All must be located inside the road boundary obtained by the perception module.

[0114] Actuator saturation constraint: for control input The range is limited, for example, the maximum / minimum torque of the drive motor and the maximum angular velocity of the steering actuator.

[0115] Stability constraint: This is a key constraint that requires the vehicle's tilt angle to be constant at any point on the trajectory. and vehicle speed All combinations must be within a stable region, i.e., satisfy... , where the function The maximum tilt angle required to maintain balance is defined for a given velocity and path curvature; It is the path curvature; The vehicle speed is represented at time t.

[0116] By solving the aforementioned constrained optimal control problem (e.g., using numerical optimization algorithms such as interior-point methods or sequential quadratic programming), the unit ultimately outputs an optimal four-dimensional spatiotemporal trajectory. The trajectory is for each moment. All specified the vehicle's position, speed, and heading, among which These represent the vehicle's time. The horizontal and vertical positions; Indicates the vehicle's time linear velocity; Indicates time The heading angle (yaw).

[0117] The path tracking control unit is responsible for accurately tracking the optimal spatiotemporal trajectory generated by the path splicing and smoothing unit. This unit employs a nonlinear model predictive control (NMPC) controller.

[0118] In each control cycle, the NMPC controller acquires the current state of the vehicle. and the reference trajectory within a future time horizon. The controller solves a small optimal control problem online, aiming to minimize the deviation between the vehicle's actual trajectory and the reference trajectory in the future time domain, while satisfying the vehicle's dynamics and stability constraints. The controller's output is the expected value at the current time step.

[0119] The actual control quantity applied, i.e., the desired linear acceleration. and angular velocity .

[0120] During path tracking, this unit reuses the dynamic stability domain predictor unit already built in autonomous following mode. A real-time safety constraint from this predictor is added to the NMPC constraints. This ensures that even under extreme conditions like high-speed U-turns, where tracking deviations occur due to model errors or external disturbances, the controller's output commands will never cause the vehicle to become unstable. This combination of planning and real-time safety monitoring provides dual safety assurance for medium-to-high-speed narrow-road U-turns, achieving a balance between efficiency and safety. Ultimately, the unit's output... and It is sent to execution module 400 for execution.

[0121] The execution module 400 is the final physical execution end of the control system of the present invention, which accurately converts the control commands described by the motion planning and control module 300 into direct drive signals for the underlying hardware execution mechanism.

[0122] The inverse kinematics unit is calculated based on the differential drive model of a self-balancing two-wheeled vehicle. First, based on the desired linear acceleration... Calculate the target linear velocity at the next moment. In one embodiment, this calculation can be performed by integrating the current velocity:

[0123] ;

[0124] in, It is the vehicle's actual linear velocity at the current moment; It is a control cycle; Indicates at time The final actual linear acceleration is then applied. Subsequently, the unit adjusts the target linear velocity accordingly. and actual angular velocity Calculate the target linear velocity of each of the left and right wheels. and :

[0125] ;

[0126] ;

[0127] in, and These represent the target linear velocities of the right and left wheels, respectively. Indicates the target linear velocity of the vehicle; The track width of a vehicle is the distance between the centers of its left and right drive wheels. Indicates time The actual angular velocity that is ultimately executed.

[0128] After obtaining the target speeds of the left and right wheels, the execution module 400 activates two parallel low-level speed closed-loop controllers to perform precise speed tracking control on the left and right drive wheels respectively. In one embodiment, the speed controller for each wheel employs a proportional-integral-derivative (PID) controller. Taking the right wheel as an example, its PID controller at time... control output The calculation method is as follows:

[0129] ;

[0130] in, At any moment The control output of the right wheel; The target speed of the right wheel The current actual speed measured by the wheel speed encoder The error between; It is the proportional gain, used to respond to the current error; It is the integral gain, used to eliminate the steady-state error of the system; It is the differential gain, used to predict the trend of error changes, suppress overshoot, and improve the system response speed; The integral term representing the error; The differential term represents the error.

[0131] The control quantity output by the PID controller It is a numerical value that needs to be converted into a physical signal that the motor driver can recognize.

[0132] The execution module 400 includes a signal conversion unit. This unit converts control signals... The linear mapping is represented by the duty cycle of a pulse width modulation (PWM) signal. This PWM signal is then sent to the motor driver (e.g., an H-bridge drive circuit) of the corresponding wheel. The motor driver adjusts the average voltage applied across the motor armature based on the duty cycle of the PWM signal, thereby controlling the motor's torque and speed.

[0133] In another embodiment, if the vehicle uses an intelligent motor controller that supports bus communication (e.g., CAN bus), the signal conversion unit encodes the output value of the PID controller (e.g., target speed or target current) into a CAN message of a specific format and sends it to the corresponding motor controller via the CAN bus. After receiving the message, the motor controller automatically completes the underlying current loop and speed loop control.

[0134] The execution module 400 ensures that whether it is the delicate acceleration, deceleration and steering in autonomous following mode or the large dynamic maneuvering in narrow road U-turn mode, the decisions of each module can be translated into the physical actions of the vehicle, thus closing the entire perception-decision-planning-control loop.

[0135] See attached document Figure 3 , Figure 3 This is a flowchart of a control method according to an embodiment of the present invention. The present invention provides a control method for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning. This method is executed in an onboard control system and includes the following steps:

[0136] The S100, through its onboard perception and positioning module, collects and processes real-time data from multiple sensors, including LiDAR, cameras, inertial measurement units, wheel speed encoders, and ultra-wideband (UWB) receivers. This data is fused to accurately obtain the vehicle's physical state information, including position, speed, vehicle tilt angle, and angular velocity, while simultaneously acquiring external environmental information. This external environmental information can be tailored to different application scenarios, such as the status of following vehicles, the precise position and movement of pedestrians wearing UWB wristbands, and information on road boundaries and obstacles.

[0137] In S200, the behavior decision-making module autonomously judges and selects from multiple predefined driving modes based on the acquired environmental information and preset driving tasks. If the current scenario is a regular road following vehicle, the decision is made to enter the autonomous following mode; if an authorized pedestrian wearing a UWB wristband is detected ahead and a following task is required, the decision is also made to enter the autonomous following mode, but the target will be specified as the pedestrian; if a narrow dead end is detected ahead or a U-turn instruction is received, the decision is made to enter the narrow road U-turn mode.

[0138] The S300 generates corresponding actual action commands based on the determined driving mode, specifically including:

[0139] S310, when the driving mode decided in step S200 is autonomous following mode, the following sub-steps are executed:

[0140] S311, Generate the desired action instruction. Based on environmental information related to the following task, the high-level policy network unit outputs a physically unconstrained desired action instruction aimed at optimizing following performance. This instruction includes the desired linear acceleration and angular velocity.

[0141] S312, predicts the dynamic safety action space. The dynamic stability domain predictor unit combines the vehicle's current physical state and the generated desired action command to predict and output a time-varying safety action space that can ensure the vehicle's dynamic stability. This space defines the range of currently feasible linear acceleration and angular velocity.

[0142] S313, Generate actual action instructions. The action projection and feedback unit projects the desired action instruction onto the range defined by the dynamic safe action space to obtain a final, safely executable actual action instruction. Simultaneously, the action projection and feedback unit calculates the deviation between the desired action and the actual action, generates a stable penalty signal, and feeds it back into the subsequent training and updates of the high-level policy network.

[0143] S320, when the driving mode decided in step S200 is intelligent decision-making mode, the following sub-steps are executed:

[0144] S321, Optimal Spatiotemporal Trajectory Planning. Based on the current road width and vehicle dynamics model, the path planning unit generates and optimizes a globally smooth, optimal spatiotemporal turning trajectory that can complete a U-turn while maintaining medium to high speeds and conforms to vehicle dynamics and stability constraints.

[0145] S322, Generate actual action commands. The path tracking control unit uses methods such as model predictive control to calculate in real time the actual action commands for accurately tracking the optimal timing air conditioning head trajectory and the current state of the vehicle.

[0146] The S400 performs command parsing and physical execution. The execution module receives the generated, uniformly formatted actual action commands. This module first decomposes the command into independent speed or torque targets for the left and right drive wheels, and then, through the underlying closed-loop controller, converts these targets into precise motor drive signals and sends them to the motor controller to drive the vehicle to complete the corresponding physical actions.

[0147] S500, cyclically executes steps S100 to S400, achieving continuous, intelligent and safe driving of the self-balancing two-wheeled vehicle in dynamic environments through high-frequency perception, decision-making, planning and control cycles.

[0148] Although embodiments of the invention have been shown and described, it will be understood by those skilled in the art that various changes, modifications, substitutions and alterations can be made to these embodiments without departing from the principles and spirit of the invention, the scope of which is defined by the appended claims and their equivalents.

Claims

1. A control system for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning, characterized in that, include: The perception and positioning module is used to collect, process and fuse multi-source sensor data in real time to generate vehicle status information of the self-balancing two-wheeled vehicle and the motion state of the target being followed. The behavior decision module is used to construct driving mode decisions to guide macro driving behavior based on the generated vehicle status information and the motion state of the target being followed. The driving mode decisions include autonomous following mode and intelligent decision mode. The motion planning and control module is used to generate actual motion commands for driving the self-balancing two-wheeled vehicle based on the constructed driving mode decision. The actual motion commands are the target linear acceleration and target angular velocity of the self-balancing two-wheeled vehicle. The execution module is used to parse the generated actual action commands and convert them into drive signals for the self-balancing two-wheeled vehicle; The vehicle state and dynamics model module is used to manage the vehicle state information of the self-balancing two-wheeled vehicle and provide the dynamics model of the self-balancing two-wheeled vehicle for the planning of the motion planning and control module.

2. The control system for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning according to claim 1, characterized in that, The sensing and positioning module includes: An inertial measurement and processing unit is used to acquire the attitude information of a self-balancing two-wheeled vehicle. Wheel speed encoder data processing unit, used to obtain the linear velocity of the self-balancing two-wheeled vehicle; The radar point cloud processing unit is used to identify targets following ahead and road boundaries; The location and distance monitoring unit is used to acquire the target location and monitor the target distance via ultra-wideband (UWB).

3. The control system for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning according to claim 1, characterized in that, The behavioral decision-making module includes: The scene understanding unit is used to construct a mode triggering condition determination result based on external environment information. The mode triggering conditions include road width conditions and path termination conditions. The task management unit is used to construct the state transition of the self-balancing two-wheeled vehicle between autonomous following state, narrow road turning preparation state and narrow road turning execution state based on the mode trigger condition determination result, so as to generate driving mode decision.

4. The control system for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning according to claim 1, characterized in that, When the driving mode decision is autonomous following mode, the motion planning and control module includes: The high-level policy network unit is used to generate the desired action instructions based on the follow-up task information in the external environment. The dynamic stability domain predictor unit is used to combine the vehicle state information of the self-balancing two-wheeled vehicle with the generated expected action command to predict the safe action space of the vehicle's dynamic stability. The motion projection and feedback unit is used to project the generated desired motion command into the safe motion space to generate the actual motion command.

5. The control system for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning according to claim 4, characterized in that, The input to the dynamic stability domain predictor unit is the concatenation vector of the physical state vector and the desired action command in the vehicle state information of the self-balancing two-wheeled vehicle. The safe action space output by the dynamic stability domain predictor unit defines the upper and lower limits of the linear acceleration and angular velocity that the self-balancing two-wheeled vehicle can execute.

6. The control system for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning according to claim 4, characterized in that, The action projection and feedback unit is also used to calculate the deviation between the expected action command and the actual action command to generate a stability margin penalty, which is then fed back to the high-level policy network unit for training and updating. The formula for calculating the stability margin penalty is as follows: ； in, and They represent the times at time 1 and 2 respectively. The lower and upper limits of linear acceleration in the safe action space output by the dynamic stability domain predictor unit; and They represent the times at time 1 and 2 respectively. The lower and upper limits of angular velocity in the safe action space output by the dynamic stability domain predictor unit; and They represent the times at time 1 and 2 respectively. The final actual linear acceleration and actual angular velocity are measured. and They represent the times at time 1 and 2 respectively. The desired linear acceleration and desired angular velocity output by the high-level policy network unit; To stabilize the punishment signal; It is a very small positive number that prevents the denominator from being zero.

7. The control system for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning according to claim 1, characterized in that, When the driving mode decision is narrow road U-turn mode, the motion planning and control module includes: The path splicing and smoothing unit is used to plan and generate the optimal spatiotemporal trajectory that meets the dynamic constraints and stability constraints based on the dynamic model. The path tracking control unit is used to calculate and generate actual action commands based on the optimal spatiotemporal trajectory generated by the planning and the vehicle state information of the self-balancing two-wheeled vehicle.

8. The control system for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning according to claim 1, characterized in that, The autonomous following mode includes autonomous following between vehicles and autonomous following between a vehicle and a person; The intelligent decision-making mode includes obstacle avoidance and detour, as well as U-turn on narrow roads.

9. The control system for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning according to claim 1, characterized in that, The execution module includes: The inverse kinematics unit is used to decompose the actual motion command into the target linear velocity of the left and right drive wheels; The underlying speed closed-loop controller is used to generate drive signals for the left and right drive wheels based on the target linear velocity.

10. A control method for a self-balancing two-wheeled vehicle based on autonomous following reinforcement learning, applied to the control system of the self-balancing two-wheeled vehicle based on autonomous following reinforcement learning as described in any one of claims 1-9, characterized in that, Includes the following steps: S100: Perform state perception and information fusion to obtain vehicle state information of the self-balancing two-wheeled vehicle and the motion state of the target being followed. S200 makes autonomous decisions on driving modes. Based on the vehicle status information of the self-balancing two-wheeled vehicle and the motion status of the target being followed, it decides on the current driving mode in autonomous following mode and intelligent decision mode. Autonomous following mode includes vehicle-to-vehicle following and vehicle-to-person following. Intelligent decision mode includes obstacle avoidance and U-turn on narrow roads. S300: When the driving mode is autonomous following mode, the desired action command is generated by the high-level policy network unit, the safe action space is predicted by the dynamic stability domain predictor unit, and the desired action command is projected into the safe action space to generate the actual action command. S400: When the driving mode is narrow road U-turn mode, an optimal spatiotemporal trajectory is planned based on the dynamic model of the self-balancing two-wheeled vehicle, and actual action commands are generated according to the optimal spatiotemporal trajectory. The S500 performs instruction parsing and physical execution, receives the generated actual action instructions, and converts them into drive signals for the self-balancing two-wheeled vehicle.

Images