A method for cross-vehicle adaptation of low-speed autonomous vehicles based on FPGA dynamic local reconstruction and an FPGA chip for low-speed autonomous vehicles.
By dividing the FPGA chip into static and dynamic areas and using the chassis task configuration parameters of the target vehicle model for online reconstruction, the underlying hardware adaptation problem in the cross-vehicle deployment of low-speed unmanned vehicles is solved, and efficient and safe cross-vehicle adaptation of autonomous driving system is achieved.
Patent Information
- Application Number
- CN202610789180.9
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2026-06-03
- Publication Date
- 2026-11-17
- Estimated Expiration
- 2046-06-03
AI Technical Summary
When existing low-speed autonomous vehicles are deployed across different vehicle models, the underlying hardware is difficult to dynamically adapt to the software for reconstruction, resulting in wasted resources, high computational power consumption, and difficulty in achieving large-scale commercial deployment in complex multi-tasking scenarios.
A dynamic local reconstruction method based on FPGA is adopted. By dividing the chip into static and dynamic areas, millisecond-level online reconstruction is performed using the chassis task configuration parameters of the target vehicle to achieve low-level adaptive control. This includes storing standardized vehicle motion intention vectors in the static area, reconstructing the circuit in reverse in the dynamic area, and achieving safe isolation and decoupling of logic module control through the boundary communication area.
It improves the versatility, portability, and safety of autonomous driving systems deployed across different vehicle models for low-speed unmanned vehicles, reduces the deployment cycle and computational power consumption across different vehicle models, and ensures the continuity and safety of the system.
Smart Images

Figure CN122332350B_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of autonomous driving and underlying computing hardware architecture technology. Specifically, it relates to a method for cross-vehicle adaptation of low-speed unmanned vehicles based on FPGA dynamic local reconstruction and an FPGA chip for low-speed unmanned vehicles. It is particularly suitable for solving the problem of deploying autonomous driving systems across vehicle models when low-speed unmanned vehicles are dealing with differentiated tasks. Background Technology
[0002] With the widespread application of low-speed autonomous vehicles, the business scenarios and tasks they undertake (such as logistics delivery, road cleaning, and security patrol) are becoming increasingly diversified. Currently, the development of low-speed autonomous vehicles generally adopts a model-specific approach, with each vehicle type corresponding to a dedicated set of underlying hardware and fixed software systems. This highly rigid model leads to significant resource waste and R&D barriers. To address the differentiated tasks undertaken by different vehicle types, the hardware system of the autonomous driving system needs to be dynamically reconfigured in conjunction with algorithm evolution to achieve efficient and low-cost deployment of a single autonomous driving system across different vehicle types.
[0003] It is particularly noteworthy that the need for dynamic adaptation across vehicle models and the engineering feasibility of underlying hardware reconfiguration have unique applicability in the field of low-speed autonomous vehicles. Compared to high-speed passenger vehicles, whose chassis forms are highly standardized and whose underlying protocols are strictly closed by OEMs, low-speed autonomous vehicles exhibit a high degree of diversity and customization in their chassis mechanical structures and actuator characteristics to meet extremely fragmented operating scenarios. This variable physical form makes frequent cross-vehicle deployments and rewriting of underlying logic a unique and extremely urgent industry pain point in low-speed autonomous driving scenarios.
[0004] However, existing underlying computing hardware lacks dynamic adaptability, making it difficult to quickly and seamlessly transform the expected motion planning results generated by the upper layer due to task changes into exclusive physical control commands that fit the unique mechanical dynamics characteristics of various target chassis. This technical contradiction not only leads to frequent rewriting of the underlying drive when deploying across different vehicle models, but also causes high computational power consumption and time costs, becoming a core bottleneck restricting the large-scale commercial deployment of low-speed autonomous vehicles in complex multi-task scenarios. Summary of the Invention
[0005] To address the challenge of dynamically adapting underlying hardware to software for reconfiguration when deploying low-speed autonomous vehicles across different vehicle models, this invention provides a method for cross-vehicle adaptation of low-speed autonomous vehicles based on FPGA dynamic local reconfiguration, along with an FPGA chip for low-speed autonomous vehicles. First, differentiated tasks are transformed into standardized motion intentions decoupled from the chassis at the upper layer, and a secure hardware architecture with physically isolated static working areas and dynamic reconfiguration areas is constructed within the underlying chip. Subsequently, millisecond-level online reconfiguration is performed based on this architecture, directly hardware decoding the standard intentions into dedicated physical control signals adapted to the target vehicle model, thereby achieving cross-vehicle underlying adaptive control while the system is running. The specific technical solution is as follows.
[0006] Firstly, this disclosure proposes a method for cross-vehicle adaptation of low-speed unmanned vehicles based on FPGA dynamic local reconfiguration. When performing cross-vehicle adaptation for autonomous driving, the method utilizes the bitstream corresponding to the chassis task configuration parameters of the target vehicle during the chassis control cycle. The dynamic area of the FPGA chip is physically reconstructed into an inverse calculation circuit that fits the dynamic characteristics of the target vehicle model. During autonomous driving, the standardized vehicle motion intention timing vector, which is independent of the chassis and consists of the target longitudinal acceleration, the target yaw rate, and the curvature of the target driving trajectory, obtained by the upper-level software, is stored in the static area of the FPGA chip. The dynamic area reads the standardized vehicle motion intention timing vector in the static area and uses the reconstructed inverse calculation circuit to perform calculations to obtain the physical control signal of the target vehicle model.
[0007] Secondly, this disclosure proposes an FPGA chip for low-speed autonomous vehicles, including a static area, a dynamic area, and a boundary communication area. The static area is configured to store, during autonomous driving operation, a standardized vehicle motion intention time-series vector obtained from upper-layer software, independent of the chassis and composed of the target longitudinal acceleration, target yaw rate, and target trajectory curvature. The dynamic area is configured to, during cross-vehicle adaptation for autonomous driving, utilize the bitstream corresponding to the chassis task configuration parameters of the target vehicle model under physical isolation during the chassis control cycle. The FPGA chip's dynamic zone is physically reconstructed into an inverse calculation circuit that matches the target vehicle's dynamic characteristics. During autonomous driving, the standardized vehicle motion intention timing vector in the static zone is read and calculated using the reconstructed inverse calculation circuit to obtain the target vehicle's physical control signals. A boundary communication zone is set at the communication boundary between the static and dynamic zones, and a hard-wired decoupling logic module is deployed across the clock domain in the boundary communication zone. When a cross-vehicle adaptation command is initiated in the static zone, the decoupling logic module is configured to latch the bus interface pins on the static zone side or force them low to a preset safety logic voltage. When the reconstruction is successful, the dynamic reconstruction zone issues a reconstruction success signal, and the decoupling logic module is configured to release the physical interface and open the data path.
[0008] The beneficial technical effects of this disclosure are: This solution helps to improve the versatility, portability, and safety of deploying autonomous driving systems across vehicle models for low-speed unmanned vehicles with differentiated tasks. Attached Figure Description
[0009] To more clearly illustrate the technical solutions in the embodiments of this application, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.
[0010] Figure 1 This is a schematic diagram of an autonomous driving cross-vehicle adaptation system architecture based on FPGA dynamic local reconstruction in one implementation. Detailed Implementation
[0011] To address the issue of frequent rewriting and system shutdown required for programming the underlying hardware of low-speed autonomous vehicles when deploying autonomous driving systems across different vehicle models, this disclosure proposes a method and system for cross-vehicle-model-based hardware adaptation of low-speed autonomous vehicle autonomous driving systems using dynamic local reconstruction based on Field-Programmable Gate Arrays (FPGAs). The method achieves efficient deployment of low-speed autonomous vehicles across vehicle models through a three-layer architecture: First, in the upper-layer end-to-end normalization and control module, the underlying mechanical parameters are completely shielded, and a standardized vehicle motion intent vector containing longitudinal acceleration, yaw rate, and curvature is uniformly output based on a cost function independent of the specific vehicle model. Second, the system adopts a heterogeneous architecture based on FPGAs, implementing strict physical isolation within the chip, dividing it into a static region for maintaining uninterrupted global operation and a dynamic region dedicated to local reconstruction, and preventing reconstruction deadlock through a communication decoupling mechanism. Finally, when a new vehicle model is added, the system, without interrupting global operation, calls a dedicated configuration bitstream online in milliseconds to physically reconstruct the dynamic region into an inverse calculation circuit matching the dynamics of that vehicle model. The reconstructed hardware can directly translate standard intent vectors into chassis physical control signals, enabling the underlying computing power to dynamically respond to upper-level regulation, significantly reducing cross-vehicle deployment cycle, control link latency, and computing power consumption.
[0012] The following will be combined with the appendix Figure 1 This application provides a clear and complete description of how the technical solution in this case is implemented. Obviously, the described implementation methods are only a part of the implementation methods in this case, not all of them. Based on the implementation methods in this case, all other implementation methods obtained by those skilled in the art without inventive effort are within the scope of protection of this application.
[0013] First level: Standardization of motion intention The motion intent of autonomous driving refers to the discrete driving behavior objectives (such as following, changing lanes, turning, and avoiding) output by the decision layer for a specific scenario. It guides the motion planning layer to generate a continuous, smooth, and executable spatiotemporal trajectory, and expresses it implicitly or explicitly to the outside world through vehicle dynamics and light signals.
[0014] When the motion intent of autonomous driving is converted into a planned trajectory, it must be done according to specific hardware instructions. Therefore, for the motion intent to be executed smoothly and safely, the chassis hardware design of autonomous vehicles needs to be considered. However, the underlying chassis hardware of autonomous driving for different purposes is specifically designed to perform specific tasks and varies significantly, as shown in Table 1.
[0015] Table 1
[0016] As shown in Table 1, the chassis designs for different vehicle types vary significantly, and the chassis hardware constitutes the vehicle's kinematic / dynamic constraints. For example, different front wheels result in different maximum steering angles; different motor / engine power leads to different maximum accelerations; and the tires, suspension, and center of gravity height determine the maximum lateral acceleration. Consequently, different vehicle models require different control algorithms, making it difficult to deploy the same algorithms across different models.
[0017] This application aims to completely shield the underlying mechanical parameters and, based on a cost function independent of specific vehicle models, uniformly output a standardized vehicle motion intention vector containing longitudinal acceleration, yaw rate, and curvature. This achieves hardware-software decoupling, transforming complex upper-level requirements into standardized motion intentions independent of specific vehicle models. Consequently, it provides a universal data input basis for the next level of hardware physical isolation and unified reception, eliminating the adaptation challenges to the underlying chassis hardware caused by differences in business tasks (such as logistics delivery, road cleaning, and security patrol) of different low-speed unmanned vehicles.
[0018] Specifically, in the end-to-end model or planning and control module deployed in the upper-layer software, a unified instruction format for the final output layer is enforced. First, a smooth trajectory with general physical feasibility is generated. To this end, the planning and control module, while masking the mechanical parameters of the target vehicle model, constructs an MPC (Model Predictive Control) trajectory optimization model based on preset general kinematic constraints, and optimizes the cost function of the MPC trajectory optimization model. Within the prediction time domain, based on the target cost function... A rolling optimization of the multidimensional spatial trajectory is performed. This cost function is completely independent of the vehicle's specific physical dynamic parameters, and its expression is:
[0019] in, For the current predicted speed, For reference speed, For prediction of the time domain (i.e., the endpoint time of trajectory optimization). For the target yaw rate, The rate of change of the curvature of the target trajectory (i.e., the derivative of curvature with respect to time). The target longitudinal acceleration. to These are weighting coefficients, and their adjustment logic is mainly used to balance trajectory following accuracy and vehicle ride smoothness. : Determines the penalty weight for speed following, used to ensure the efficiency and speed accuracy of task execution; and Excessive longitudinal acceleration and yaw rate are limited respectively to ensure ride comfort and chassis lateral and longitudinal stability. This is used to constrain drastic changes in steering wheel angle, ensuring smooth and consistent steering actions. The core objective of the algorithm is to minimize the aforementioned objective function while satisfying smoothness and obstacle avoidance constraints, thereby autonomously generating an optimal time-series trajectory polynomial decoupled from the underlying mechanical structure at the software level.
[0020] The core objective of the algorithm is to minimize the objective function while satisfying the constraints of smoothness and obstacle avoidance, thereby autonomously generating an optimal time-series trajectory polynomial decoupled from the underlying mechanical structure at the software level.
[0021] Based on the above optimal trajectory, the system uses the vehicle kinematics model to differentiate and solve it, transforming it into purely physical kinematic parameters, that is, outputting a time-varying parameter. Changing standardized vehicle motion intention time vector :
[0022] in, Represents the longitudinal acceleration of the target. Represents the target's yaw rate. This represents the curvature of the target driving trajectory. This standardized vehicle motion intent vector serves as the only universally applicable command across vehicle models throughout the entire system, triggering subsequent dynamic local reconstruction and dynamic analysis of the underlying hardware.
[0023] Second layer: Static and dynamic partitioning and secure communication This layer constructs the underlying physical hardware architecture that supports dynamic adaptation. It maps specific business logic through physical partitioning, implementing the decoupling of software and hardware for autonomous driving at the silicon chip level. The static area carries standardized intents unrelated to the chassis through rigid specifications, while the dynamic area is dedicated to solving chassis-specific inverse dynamics, realizing a computing power architecture with code that is universal across vehicle models.
[0024] To meet the dynamic control requirements for cross-vehicle platform adaptation, this solution adopts a heterogeneous system-on-a-chip architecture based on FPGA (see [link]). Figure 1 As shown in the diagram, strict silicon-level resource constraints are implemented within the chip. Through deep physical isolation of logic and timing, the boundaries between the two independent physical regions—static and dynamic—are clearly defined, and a secure cross-domain communication decoupling mechanism is constructed accordingly. This core design not only removes bus blocking obstacles for the secure delivery of standard instructions from the upper layer but also creates independent physical spaces that do not interfere with the online loading of dedicated hardware configurations at the lower layer. The specific physical region division and communication mechanism are as follows.
[0025] Static Area: This area houses the system's ARM main control core, AIE artificial intelligence engine, and basic communication bus. It remains permanently resident throughout the system's entire lifecycle, with its physical resource allocation always locked. Specifically, the system's main control center is the microprocessor, responsible for running the autonomous driving operating system, executing global task scheduling, and initiating dynamic local reconstruction instructions. The artificial intelligence engine here specifically handles the high-concurrency inference tasks of the neural network for the upper-layer multimodal end-to-end model, responsible for outputting environmental perception results and trajectory features. To achieve high-speed hardware-level interconnection between heterogeneous cores, the FPGA provides underlying static logic resources to instantiate and generate a global communication backbone based on the Advanced Dextensible Interface (AXI) interconnect matrix. To ensure efficient instruction and data distribution, this interconnect matrix is further divided into two types of buses: a lightweight control bus, serving as a low-bandwidth channel, is specifically responsible for transmitting configuration instructions and status polling signals issued by the microprocessor; the high-speed data stream bus, serving as a high-bandwidth channel, is responsible for continuously transmitting standardized vehicle motion intent vectors generated after computational processing by the artificial intelligence engine. .
[0026] Dynamic Region (DPR): This region is a reserved, flexible hardware gateway within the chip, built on an FPGA logic array, with physical reconfiguration capabilities. It is dedicated to receiving and instantiating vehicle inverse dynamics analysis circuits adapted to specific vehicle models. To meet the computational demands of high-frequency closed-loop control and complex multi-degree-of-freedom inverse dynamics calculations, this region integrates a large number of dynamically configurable underlying physical resources, including block memory for lookup table feedforward mapping, digital signal processing multiply-accumulate sequences for performing parallel matrix transformations, and configurable logic blocks for constructing customized control timing sequences. The high-frequency closed-loop control specifically refers to a low-level automatic correction control system running in the FPGA underlying hardware, with real-time sensor feedback, which dynamically controls the actuator output by continuously comparing and correcting real-time sensor data.
[0027] At the physical level, the boundary routing and logical netlist of this region are strictly constrained, and its physical resource topology model strictly follows the following constraints: , .in, For the overall FPGA logic resources, This is a static workspace resource (including hard core communication interface resources reserved for interaction between ARM and AIE). Resources for dynamic local reconstruction area.
[0028] Standard DPR (Dynamic Reconfiguration Processing) is highly susceptible to bus paralysis during reconstruction, which is disastrous for autonomous driving. To support the core requirement of uninterrupted global system operation, this system is designed with a boundary communication module and an anti-deadlock mechanism. Specifically, a boundary communication zone is set up at the communication boundary between the static working area and the dynamic local reconstruction area (such as at the AXI bus interface). A hard-wired decoupling logic module is deployed across the clock domain in the boundary communication zone. This decoupling logic module is controlled by the reconstruction controller. Each time the system is powered on or connected to the chassis, the static working area automatically reads the hard-coded ID on the chassis interface board. If the comparison finds that the current chassis ID is inconsistent with the previous one, the static area initiates a cross-vehicle adaptation command. The decoupling logic module receives the cross-vehicle adaptation command and immediately performs physical-level electrical isolation through a physical isolation signal switch, latching the bus interface pins on the static area side or forcibly pulling them low to a preset safe logic level. This action actively intercepts all out-of-bounds pending read / write transactions during reconstruction, preventing system-level deadlock caused by bus response timeouts in the processing cores (such as ARM or AIE) within the static area. Once the target local configuration bitstream is successfully injected through the internal configuration access port and verified as ready by the hardware state machine, the dynamic reconfiguration area issues a decoupling signal. The decoupling logic module synchronously releases the physical isolation, re-releases the physical interface, and reopens the data path. This mechanism ensures the secure continuity of the underlying control flow at the physical level, reducing the impact of cross-vehicle model reconfiguration on the overall system to zero.
[0029] Third layer: Millisecond-level chassis physical translation Building upon the standardized data interfaces and secure hardware isolation established in the first two layers, this layer serves as the core execution stage for cross-vehicle adaptive control. It aims to rapidly reconstruct the local logic gate array online, hardware-decoding the received standardized motion intent into low-level physical control signals specific to the target vehicle model, which are then executed by peripheral actuators. This process completely establishes the final physical execution link from the unified upper-layer algorithm to the specific vehicle control commands.
[0030] Offline pre-configuration triggering and online identification: To achieve accurate configuration invocation, an offline pre-configured vehicle model index library and a bitstream injection-based online loading and reconstruction are used in tandem. The vehicle model index library is stored in a non-volatile storage medium managed by the ARM main control core in the static area. The bitstream is injected into the dynamic area.
[0031] During the offline phase, corresponding target local configuration bitstreams are generated in advance for different vehicle identification sets and their corresponding chassis mechanical dynamics parameters and task configuration parameters (hereinafter collectively referred to as target configuration parameters). A mapping relationship of "vehicle identification - target configuration parameter - bitstream file" is pre-built to establish a vehicle index library. At the same time, for newly added vehicle identifications, an automated compilation process can be triggered offline to dynamically expand the index library. Specifically, the mechanical dynamics parameters include, but are not limited to, the chassis wheelbase, track width, maximum steering angle, and motor torque limit curve; the task configuration parameters include, but are not limited to, the maximum operating speed limit, maximum load threshold, and communication interface protocol of external equipment under specific business scenarios.
[0032] Before triggering dynamic partial reconstruction, the target vehicle model identifier is first established to form a complete reconstruction trigger. To this end, the present invention provides an acquisition mechanism: using the control core residing in the static working area, the vehicle model hard-coded ID information is automatically read from the non-volatile memory of the chassis interface board via the system bus.
[0033] After successfully acquiring the target vehicle model identifier, the static workspace initiates a cross-vehicle adaptation command based on the pre-defined vehicle model hard-coded information. The system uses this target vehicle model identifier as an index to search the vehicle model index library. By precisely matching the preset chassis mechanical features with the task configuration parameters, it retrieves the corresponding target local configuration bitstream from the storage medium and adds it to the buffer. Subsequently, the system performs a pre-emptive safety time assessment. After ensuring that the theoretical reconstruction time is within the chassis safety control cycle, it injects the bitstream into the dynamic local reconstruction area to initiate physical reconstruction.
[0034] Although low-speed autonomous vehicles are limited by their relatively low operating design speed, possessing a high degree of safety tolerance and fault tolerance boundaries, this system further incorporates a safety degradation strategy during dynamic local reconfiguration to ensure absolute safety of the chassis control. This safety degradation refers to the system actively depriving advanced autonomous driving control functions and reverting to basic safety actions when it detects a reconfiguration anomaly. This "controlled function concession" is defined as "degradation." This application designs the physical reconfiguration time of the dynamic reconfiguration zone during dynamic reconfiguration. It must be strictly limited by the chassis control cycle. This design ensures that even at low speeds, the millisecond-level reconstruction time can be fully accommodated by the chassis control cycle. Furthermore, even if an anomaly occurs during reconstruction, the system can execute an emergency safety stop through the degradation strategy of the underlying hardware, eliminating the risk of loss of control.
[0035] Here, the control cycle of the chassis This refers to the maximum permissible time interval between two consecutive discrete control messages received by the chassis drive-by-wire actuator. It represents the physical baseline time taken for the underlying hardware to complete a complete closed loop of state perception, feedforward calculation, and braking. The complete closed loop is a knowledge-decision-execution feedback loop.
[0036] To prevent the chassis from triggering the disconnection protection mechanism and losing control in the event of a command timeout, the constraint relationship between the physical reconfiguration time and the control cycle must be satisfied. : ,in, Configure the file size of the bitstream for the target locality. Configure the port's transmission rate.
[0037] See Figure 1 As shown, under the security degradation strategy, reconstruction is performed and the reconstruction process is monitored. The reconstruction time is calculated based on the file size of the target local configuration bitstream. If the system determines that the current reconstruction operation is within the chassis's tolerable safe time window, it will perform bitstream injection and load the reconstruction bitstream online; otherwise, if the injection fails, or the CRC check within the configuration frame reports an error, or if it is not within the chassis's control cycle... Refactoring completed within the time limit (actual refactoring time) If the injection fails, the system will immediately interrupt the current reconstruction process and trigger a safety rollback mechanism. Under this mechanism, the safety control logic in the static area will send a system reset command to the underlying bus matrix, forcing the reconstructed logic gate array in the dynamic area to zero through the reset signal, clearing any remaining invalid configuration states. Simultaneously, an emergency safety stop command will be issued. If the injection is successful, the dynamic area will send a decoupling signal to the static area, releasing the physical isolation of the static area, re-releasing the physical interface, and reopening the data path.
[0038] To further clarify the engineering feasibility of this constraint, since the dynamic local reconfiguration region only needs to instantiate the inverse dynamics solution and closed-loop control logic for a specific chassis, it occupies very few logic gate resources. Typically, the data size is only in the hundreds of kilobytes range. Relying on the high-speed configuration access ports within the heterogeneous chip, the actual physical reconfiguration time is minimal. Typically, the control cycle can be controlled in the 1 to 2 millisecond range. In comparison, the control cycle of mainstream low-speed autonomous vehicle drive-by-wire chassis... Typically, this takes between 10 and 20 milliseconds. Therefore, the actual time required for partial reconstruction is much shorter than the communication window during which the chassis receives commands. This process enables the partial reconstruction of the underlying logic gate array to be completed in milliseconds without interrupting the global system operation in the static area.
[0039] In addition, during the configuration reconfiguration loading process, if the system detects that the injection of the target local configuration bit stream fails or that the CRC check inside the configuration frame fails, the system will immediately interrupt the current reconfiguration process and trigger a safety rollback mechanism.
[0040] After reconstruction, the dynamic local reconstruction area has been completely transformed at the physical level into a dedicated inverse calculation circuit that highly matches the dynamic characteristics of the current target vehicle model, receiving the standardized vehicle motion intention vector transmitted from the upper layer. It performs pure circuit-level translation, using a dedicated reverse computation circuit for dynamic calculation / error correction to obtain the control signals for the actuator. First, in the feedforward dynamic calculation stage, the system utilizes the physical lookup table of the underlying block memory and hard-wired interpolation logic in parallel to translate the abstract trajectory curvature... The mechanical deflection angle feedforward is mapped to compensate for nonlinear characteristics such as gear backlash, and the longitudinal acceleration is transformed using a pipelined matrix transformation of a digital signal processing multiply-accumulate array. With yaw rate The parameters are precisely converted into target motor torque and wheel speed, generating feedforward commands. Subsequently, in the error correction stage, a kilohertz-level hardware high-frequency closed-loop control circuit instantiated in the dynamic reconfiguration area is used to read the feedback from the underlying sensors in real time and compare it with the feedforward commands to complete high-frequency error compensation. Finally, in the protocol encapsulation and delivery stage, the corrected actuation parameters are directly converted into precise control signals at the underlying level by the customized soft-core protocol logic, and then multiplexed through the routing matrix in the static area to map the control signals to fixed physical pins, ultimately outputting them to the peripheral drive-by-wire actuator.
[0041] Based on the above, this disclosure relates to a method for cross-vehicle adaptation of low-speed unmanned vehicles based on FPGA dynamic local reconfiguration. When performing cross-vehicle adaptation for autonomous driving, the method utilizes the bitstream corresponding to the chassis task configuration parameters of the target vehicle during the chassis control cycle. The dynamic area of the FPGA chip is physically reconstructed into an inverse calculation circuit that fits the dynamic characteristics of the target vehicle model. During autonomous driving, the standardized vehicle motion intention timing vector, which is composed of the target longitudinal acceleration, target yaw rate and target driving trajectory curvature obtained by the software and independent of the chassis, is stored in the static area of the FPGA chip. The dynamic area reads the standardized vehicle motion intention timing vector in the static area and uses the reconstructed inverse calculation circuit to calculate and obtain the physical control signal of the target vehicle model.
[0042] This disclosure also relates to an FPGA chip for low-speed unmanned vehicles, including a static area, a dynamic area, and a boundary communication area. The static area is configured to store, during autonomous driving operation, a standardized vehicle motion intention time vector consisting of the target longitudinal acceleration, target yaw rate, and target trajectory curvature obtained by software, independent of the vehicle model. The dynamic area is configured to, during cross-vehicle adaptation for autonomous driving, utilize the bit stream corresponding to the chassis task configuration parameters of the target vehicle model under physical isolation during the chassis control cycle. The FPGA chip's dynamic zone is physically reconstructed into an inverse calculation circuit that matches the target vehicle's dynamic characteristics. During autonomous driving, the standardized vehicle motion intention timing vector in the static zone is read and calculated using the reconstructed inverse calculation circuit to obtain the target vehicle's physical control signals. A boundary communication zone is set at the communication boundary between the static and dynamic zones, and a hard-wired decoupling logic module is deployed across the clock domain in the boundary communication zone. When a cross-vehicle adaptation command is initiated in the static zone, the decoupling logic module is configured to latch the bus interface pins on the static zone side or force them low to a preset safety logic voltage. When the reconstruction is successful, the dynamic reconstruction zone issues a reconstruction success signal, and the decoupling logic module is configured to release the physical interface and open the data path.
[0043] In summary, whether it is a method for cross-vehicle adaptation of low-speed unmanned vehicles based on FPGA dynamic partial reconfiguration or an FPGA chip for low-speed unmanned vehicles, they are both aimed at achieving the safe and efficient application of dynamic partial reconfiguration technology in cross-vehicle adaptation scenarios of low-speed unmanned vehicles. They innovatively adopt specific architectural constraints and secure communication mechanisms, as detailed below.
[0044] 1. Specific business mapping of physical partitions. Most existing technologies are general-purpose platforms, while this application advances the "software-hardware decoupling" of low-speed autonomous vehicles to the physical level of logic gates within a single chip. It clearly stipulates that the static area is specifically responsible for general logic, relying on its internally integrated ARM main control core and AIE artificial intelligence engine to generate and carry standardized motion intentions independent of the specific chassis, while the dynamic area is specifically responsible for solving the chassis-specific inverse dynamics, thereby constructing a computing power architecture with code that can be used across vehicle models.
[0045] 2. Boundary Isolation and Degradation Strategies for Autonomous Driving Functional Safety. Addressing the catastrophic risk of bus paralysis and vehicle loss of control during standard dynamic partial reconfiguration, this application goes beyond conventional physical layer isolation. It deeply integrates boundary communication decoupling mechanisms with the chassis safety control cycle. Based on actively latching the AXI bus interface during reconfiguration, this solution is tailored to the reconfiguration time... Less than chassis control cycle By addressing the timing constraints, a safety degradation strategy was designed, effectively solving the problem that existing DPR technology struggles to pass the stringent safety certifications for autonomous driving.
[0046] 3. Automated hardware-triggered closed loop based on chassis ID. Reconfiguration is not arbitrarily issued from the upper layer, but is triggered by automatically reading the chassis hard-coded ID, combined with a pre-built index library, to achieve millisecond-level plug-and-play functionality.
[0047] The various embodiments of the present invention have been described above. These descriptions are exemplary and not exhaustive, nor are they limited to the disclosed embodiments. Many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the described embodiments. The terminology used herein is chosen to best explain the principles, practical application, or technical improvements to the embodiments in the market, or to enable others skilled in the art to understand the embodiments disclosed herein. The scope of the invention is defined by the appended claims.
Claims
1. A method for cross-vehicle adaptation of low-speed unmanned vehicles based on FPGA dynamic local reconstruction, characterized in that: A boundary communication zone is set up at the communication boundary between the static and dynamic zones. A hard-wired decoupling logic module is deployed across the clock domain within this boundary communication zone. When a cross-vehicle adaptation command is initiated in the static zone based on preset vehicle model hard-coded information, the decoupling logic module is configured to latch the bus interface pins on the static zone side or force them low to a preset safe logic level to physically isolate the static and dynamic zones. Under this physical isolation, during autonomous driving cross-vehicle adaptation, the bit stream corresponding to the chassis task configuration parameters of the target vehicle is injected into the dynamic zone during the chassis control cycle. The bitstream is used to physically reconstruct the dynamic region of the FPGA chip into a reverse calculation circuit that conforms to the dynamic characteristics of the target vehicle model. When the reconstruction is successful, the dynamic reconstruction region sends a reconstruction success signal, and the decoupling logic module is configured to release the physical interface and open the data path. If the injection fails, the CRC check inside the configuration frame reports an error, or the chassis control cycle is not in progress, the signal will be lost. If the reconstruction is completed within the static area, the current reconstruction process will be interrupted and an emergency safety stop command will be issued. At the same time, the safety control logic in the static area will send a system reset command to the underlying bus matrix, and force the reconstructed logic gate array in the dynamic area to be reset to zero through the reset signal. During autonomous driving, a standardized vehicle motion intention timing vector, which is independent of the chassis and consists of the target longitudinal acceleration, the target yaw rate, and the curvature of the target driving trajectory, obtained by the upper-layer software, is stored in the static area of the FPGA chip. The dynamic area reads the standardized vehicle motion intention timing vector from the static area and uses the reconstructed inverse calculation circuit to perform calculations to obtain the physical control signal of the target vehicle model.
2. The method according to claim 1, characterized in that, Before physically reconstructing the dynamic region of the FPGA chip into a reverse calculation circuit that fits the dynamic characteristics of the target vehicle model, the physical reconstruction time is calculated based on the file size of the bitstream corresponding to the chassis task configuration parameters of the target vehicle model: , Configure the file size of the bitstream for the target locality. Configure the port's transmission rate; if the calculated physical reconfiguration time is less than the chassis's control cycle. If so, a refactoring will be performed, and the refactoring process will be monitored.
3. The method according to claim 1, characterized in that, The chassis configuration parameters of the target vehicle are obtained from the pre-set vehicle index library using the chassis ID of the target vehicle.
4. The method according to claim 3, characterized in that, The pre-set vehicle model index library is located in the static area.
5. The method according to claim 1, characterized in that, The target's longitudinal acceleration, yaw rate, and trajectory curvature are obtained using the cost function of the trajectory optimization model described below: in, For the current predicted speed, For reference speed, To predict the time domain, The rate of change of the curvature of the target trajectory. For the target longitudinal acceleration, to These are the weighting coefficients. The target yaw rate.
6. The method according to claim 1, characterized in that, The solution includes: During the feedforward calculation stage, the physical lookup table of the underlying block memory and hard-wired interpolation logic are used in parallel to determine the curvature of the target driving trajectory. Mapped to mechanical deflection angle, the longitudinal acceleration is transformed using a pipelined matrix transformation of a digital signal processing multiply-accumulate array. yaw rate of the target This is converted into target motor torque and wheel speed to generate feedforward commands; Subsequently, the feedback from the underlying sensors is read in real time using a hardware high-frequency closed-loop circuit, and compared with the feedforward command to complete the high-frequency error compensation and correction.
7. An FPGA chip for low-speed unmanned vehicles, comprising a static area, a dynamic area, and a boundary communication area, characterized in that, include: The static area is configured to store a standardized time-series vector of vehicle motion intent, which is independent of the chassis and consists of the target longitudinal acceleration, the target yaw rate, and the curvature of the target driving trajectory, obtained by the upper-layer software during autonomous driving. The dynamic area is configured to, during cross-vehicle adaptation for autonomous driving, inject the bitstream corresponding to the chassis task configuration parameters of the target vehicle under physical isolation, during the chassis control cycle. The bit stream is used to physically reconstruct the dynamic area of the FPGA chip into an inverse calculation circuit that fits the dynamic characteristics of the target vehicle model; during autonomous driving, the standardized vehicle motion intention time vector of the static area is read and the reconstructed inverse calculation circuit is used to calculate and obtain the physical control signal of the target vehicle model. The boundary communication zone is set at the communication boundary between the static zone and the dynamic zone. Hard-wired decoupling logic modules are deployed across the clock domain in the boundary communication zone. When a cross-model adaptation command is initiated in the static area based on pre-set vehicle model hard-coded information, the decoupling logic module is configured to latch the bus interface pins on the static area side, or force them low to a preset safe logic level, to physically isolate the static area and the dynamic area. Upon successful reconstruction, the dynamic reconstruction area issues a reconstruction success signal, and the decoupling logic module is configured to release the physical interface and open the data path. If injection fails, the CRC check within the configuration frame fails, or the control cycle of the chassis is not in effect, the decoupling logic module will not be activated. If the reconstruction is completed within the current area, the current reconstruction process will be interrupted and an emergency safety stop command will be issued. At the same time, the safety control logic in the static area will send a system reset command to the underlying bus matrix, and the reconstructed logic gate array in the dynamic area will be forced to zero through the reset signal.
Citation Information
Patent Citations
Chassis domain control system and vehicle
CN122078421A
Intelligent parking method and device
WO2025201028A1
