A multi-vehicle space-time joint trajectory planning method

Abstract

This invention belongs to the field of autonomous driving technology, and discloses a multi-vehicle spatiotemporal joint trajectory planning method, comprising the following steps: A01: Obtaining the current positions of multiple vehicles in the planning map, the start and end points of the delivery task, and the current vehicle status data; A02: Determining the initial trajectory results of multiple vehicles based on their path results; A03: Determining feasible and safe planning areas and collision-free feasible areas in the trajectory planning of multiple vehicles; A04: Constructing a spatiotemporal joint multi-vehicle trajectory optimization problem using the initial trajectory results; A05: Grouping the objects of the spatiotemporal joint multi-vehicle trajectory optimization problem based on the path results of multiple vehicles; A06: Solving the spatiotemporal joint multi-vehicle trajectory optimization problem for multiple vehicles in the same group; A07: Searching for collision-free path results for multiple vehicles. By constructing the spatiotemporal joint trajectory optimization in a convex space generated based on the environment map, vehicle paths can be planned more accurately.

Description

Technical Field

[0001] This invention relates to, but is not limited to, the field of autonomous driving technology, and specifically to a multi-vehicle spatiotemporal joint trajectory planning method. Background Technology

[0002] The new round of industrial manufacturing transformation and upgrading has placed new performance demands on unmanned transportation vehicles. Researching multi-vehicle trajectory optimization can help logistics companies complete more delivery tasks in a shorter time, improving transportation efficiency and service quality.

[0003] Researching collaborative trajectory planning technology for multiple unmanned transport vehicles can lead to the rational allocation of transportation and delivery tasks and a reduction in operating costs such as vehicle mileage and fuel consumption through optimized solutions.

[0004] Traditional multi-vehicle trajectory planning methods typically only consider time or space optimization, neglecting joint spatiotemporal optimization. This leads to significant deviations between the planned results and the optimization objectives, or inappropriate driving routes. Traditional multi-vehicle trajectory planning methods usually solve problems by constructing optimization problems and sampling, which often require substantial computational resources and time, resulting in low efficiency and inability to meet real-time requirements. Furthermore, different optimization problems often employ different optimization frameworks, making it difficult to unify optimization strategies and evaluation criteria, thus impacting the overall planning effectiveness.

[0005] Based on this research, a multi-vehicle spatiotemporal joint trajectory planning method was developed. Summary of the Invention

[0006] This invention provides a multi-vehicle spatiotemporal joint trajectory planning method. It constructs a spatiotemporal joint trajectory optimization within a convex space generated based on an environmental map, and groups vehicles with high distance coupling together, performing optimization on multiple vehicles within each group. This method can plan vehicle paths more accurately, avoid conflicts and congestion between vehicles, and improve the optimization degree of the paths. It solves the technical problems of existing multi-vehicle planning methods, such as low solution efficiency and inability to meet real-time requirements during the problem construction and solution process.

[0007] A multi-vehicle spatiotemporal joint trajectory planning method includes the following steps:

[0008] Step A01: Obtain the current location of multiple vehicles on the planning map, the start and end points of the delivery task, and the current vehicle status data;

[0009] Step A02: Based on the path results of multiple vehicles, determine the initial trajectory results of multiple vehicles, and process the initial guidance path results of multiple vehicles into segments as initial trajectory results.

[0010] Step A03: Determine the feasible and safe planning area in the trajectory planning of multiple vehicles. Based on the path results of multiple vehicles and combined with static or low-speed obstacle environment information, determine the collision-free feasible area.

[0011] Step A04: Using the initial trajectory results, construct a spatiotemporal joint multi-vehicle trajectory optimization problem;

[0012] Step A05: Group the objects of the spatiotemporal joint multi-vehicle trajectory optimization problem based on the path results of multiple vehicles;

[0013] Step A06: For multiple vehicles in the same group, solve the spatiotemporal joint multi-vehicle trajectory optimization problem;

[0014] Step A07: Search for collision-free paths for multiple vehicles.

[0015] Optionally, in step A02, the initial guidance path results of multiple vehicles are segmented into initial trajectory results, which includes the following steps: after obtaining the previous guidance path results and discretizing them into points, a learning-based speed decision method is used to obtain trajectory points, and then the trajectory points are fitted to a fifth-order polynomial.

[0016] Optionally, the learning-based speed decision method includes the following steps:

[0017] 1) Define the state space, action space, reward function, and state transition function;

[0018] 2) Use a POMDP-based reinforcement learning algorithm to estimate the state-action value and obtain the optimal action value function and optimal policy for the path;

[0019] 3) Using an online planner, such as a confidence tree-based planner, the optimal strategy and optimal action value function are used to plan based on the current state, obtain the optimal action in the current state, and execute it;

[0020] 4) The actions performed by the online planner and the rewards returned by the environment are used as experience and fed back to a reinforcement learning algorithm to update the policy and value function using the acquired experience;

[0021] 5) Repeat steps 1)-4) above, continuously updating the strategy and value function, and using the latest strategy and value function for planning and control, and finally complete the speed allocation and trajectory planning of the path.

[0022] Optionally, in step A04, the method for constructing the spatiotemporal joint multi-vehicle trajectory optimization problem using the initial trajectory results is as follows: based on the fitted fifth-degree polynomial, obtain each segment of the fifth-degree polynomial, then discretize each segment of the fifth-degree polynomial into N small segments, and integrate the time term of the square of the cubic derivative of each small segment.

[0023] Optionally, in step A03, the method for forming the collision-free feasible region includes the following steps:

[0024] 1) Discretize the continuous trajectory of the preceding guidance path, and break the preceding guidance path into segments of broken lines;

[0025] 2) For each broken line segment, with the broken line segment as the center line, construct two parallel line segments on both sides of the broken line segment. The parallel line segments extend to the left and right respectively, and the extension distance can be set to a fixed distance.

[0026] 3) Continue to divide the area divided by each parallel line segment into two sub-regions again according to the center line;

[0027] 4) Repeat the binary search operation in step 3) until the preset accuracy is reached. Once the preset accuracy is reached, the broken line segment is used as the dividing line between the obstructed and unobstructed areas.

[0028] 5) For each polyline segment, repeat steps 1)-4) above to obtain the discretized parallel line segments and the unobstructed convex rectangular regions divided by each parallel line segment.

[0029] 6) Connect all the unobstructed convex rectangular regions to form a collision-free path space.

[0030] Optionally, the energy cost term and time cost term of the fifth-order polynomial, together with the constraint penalty term that conforms to kinematic constraints, obstacle avoidance constraints, and running time constraints, constitute a joint multi-vehicle trajectory optimization problem.

[0031] Optionally, the term obtained by integrating the square of the cubic derivative at each segment point with respect to the time term is used as the energy cost term for controlling the vehicle.

[0032] Optionally, the overall trajectory time multiplied by the total time penalty coefficient represents the time cost term, and the energy cost term and time cost term for each vehicle can be expressed as follows:

[0033]

[0034] Where μ(t) represents p (m) The vehicle's acceleration control force, W is a diagonal matrix used to represent the penalty weight of the control quantity, w T Let T be the penalty weight for the total time T. After discretizing into piecewise points, this integral expression can be represented as the sum of the squares of the control force at each discrete trajectory point.

[0035] Optionally, the kinematic constraints can be expressed as

[0036]

[0037] in Let be the control variable for vehicle i, and j be the segment number of the trajectory. The spatial state variables and vehicle control variables in the kinematic state transition equations can be mutually converted. Through this conversion relationship, the variables relating to spatial state and control constraints can be uniformly represented in Cartesian coordinates using the vehicle's spatial position variables.

[0038] Optionally, in step A05, the method for grouping the objects of the optimization problem based on the path results of multiple vehicles is as follows: according to the results of the preceding guidance path, those with high degree of coupling between path points are grouped into the same group.

[0039] Optionally, in step A06, the method for solving the spatiotemporal joint multi-vehicle trajectory optimization problem for multiple vehicles in the same group is as follows: An optimization problem solver is used, or the constraint penalty term is transformed into a penalty function cost term. This method solves the multi-vehicle trajectory optimization problem for the same group using an unconstrained optimization problem. Compared with existing technologies, the beneficial effects of this invention are:

[0040] 1) The multi-vehicle spatiotemporal joint trajectory planning method described in this application comprehensively considers multiple objectives, such as time, distance, and energy, to generate trajectory schemes with higher quality.

[0041] 2) The solution described in this application transforms the problem into a convex optimization form by changing the expression of the unobstructed area around the vehicle body and the collision constraints, thereby improving the solution efficiency.

[0042] 3) The scheme described in this application considers the interaction and cooperative behavior between multiple vehicles to generate a trajectory scheme with higher cooperativeness, which enables multiple vehicles to run simultaneously in the same site and reach their respective destinations without conflict. It has more advantages in terms of accuracy, efficiency and scalability, and can better solve practical traffic management and scheduling problems. Attached Figure Description

[0043] Figure 1 This is a schematic diagram illustrating the implementation process of a multi-vehicle cooperative trajectory planning method in a closed environment, as provided in an embodiment of this application. Detailed Implementation

[0044] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the embodiments. The illustrative embodiments and descriptions of this invention are only for explaining this invention and are not intended to limit this invention.

[0045] Example:

[0046] like Figure 1 As shown, a multi-vehicle spatiotemporal joint trajectory planning method includes the following steps:

[0047] Step A01: Obtain the current location of multiple vehicles on the planning map, the start and end points of the delivery task, and the current vehicle status data;

[0048] Step A02: Based on the path results of multiple vehicles, determine the initial trajectory results of multiple vehicles, and process the initial guidance path results of multiple vehicles into segments as initial trajectory results.

[0049] Step A03: Determine the feasible and safe planning area in the trajectory planning of multiple vehicles. Based on the path results of multiple vehicles and combined with static or low-speed obstacle environment information, determine the collision-free feasible area.

[0050] Step A04: Using the initial trajectory results, construct a spatiotemporal joint multi-vehicle trajectory optimization problem;

[0051] Step A05: Group the objects of the spatiotemporal joint multi-vehicle trajectory optimization problem based on the path results of multiple vehicles;

[0052] Step A06: For multiple vehicles in the same group, solve the spatiotemporal joint multi-vehicle trajectory optimization problem;

[0053] Step A07: Search for collision-free paths for multiple vehicles.

[0054] The multi-vehicle spatiotemporal joint trajectory planning method described in this embodiment is mainly applied to unstructured environments, enabling multiple vehicles to run simultaneously in the same location and reach their respective destinations without conflict. It has higher solution efficiency and the results are close to the path with the shortest overall journey.

[0055] In step A01, based on information from vehicle sensors and a priori maps, the current positions of multiple vehicles on the planned map, the start and end points of the delivery task, and the current vehicle status data are obtained. The priori map refers to the point cloud map obtained from a target radar (excluding the radar to be calibrated) before radar extrinsic parameter calibration. Furthermore, in this embodiment, the start and end points of the delivery task refer to the start and end points of the logistics vehicle delivery task.

[0056] Optionally, in step A02, segmenting the initial guidance path results of multiple vehicles into initial trajectory results includes the following steps: after discretizing the obtained preliminary guidance path results into points, a learning-based speed decision method is used to obtain trajectory points, and then a fifth-order polynomial is fitted between the trajectory points. The method for discretizing the preliminary guidance path results into points is obtained through search, sampling, or a learning-based method.

[0057] The learning-based speed decision-making method includes the following steps:

[0058] 1) Define the state space, action space, reward function, and state transition function;

[0059] 2) Use a POMDP-based reinforcement learning algorithm to estimate the state-action value and obtain the optimal action value function and optimal policy for the path;

[0060] 3) Design an online planner, such as a confidence tree-based planner, that uses the optimal strategy and the optimal action value function to plan based on the current state, obtain the optimal action in the current state, and execute it.

[0061] 4) The actions performed by the online planner and the rewards returned by the environment are used as experience and fed back to a reinforcement learning algorithm to update the policy and value function using the acquired experience;

[0062] 5) Repeat steps 1)-4) above, continuously updating the strategy and value function, and using the latest strategy and value function for planning and control, and finally complete the speed allocation and trajectory planning of the path.

[0063] In practical implementation, the specific method for inputting data and the online planner is as follows:

[0064] 1) Create a large number of "negative samples" based on the dynamics and environmental information of multiple vehicles. These samples are caused by poor operations under the current policy decision, such as speeding and leaving the lane. Then, filter out negative examples from these samples to evaluate the performance of the current policy.

[0065] 2) In step 1), the samples created are all negative examples. Therefore, it is also necessary to create some "positive samples" to achieve strategies that meet safety standards, such as effectively controlling vehicle speed, maintaining lane position, and maintaining a safe distance from the vehicle in front. These "positive samples" can be generated using existing rule bases and human driving behavior data without the need for labeled data.

[0066] 3) Based on the above samples, including positive and negative samples, inverse reinforcement learning is used to derive the reward function of the optimal policy, bypassing the short-term manual rule base task and directly constructing the reward function for blocking the wall; it can provide equitable responsibility for the behavior of the car and encode goals such as safety and efficiency; in inverse reinforcement learning, one goal is to find a set of trust benchmarks and reward functions, so that the additional logic generalized from the data based on this function is more general than manually written rules;

[0067] 4) The newly learned safety strategy is directly applied to the online planner to generate a safe and effective path planning scheme, which is updated according to the gradient descent of the reward function.

[0068] Because this planner can monitor real-time changes in its surrounding environment and its own state to plan online and continuously update its strategy, it can achieve self-monitoring and improve its performance with the help of a reward function.

[0069] The local results of the online planner can be compared with historical data by the learner to generate new guidance strategies and value networks, thereby improving the online planner's results. Simultaneous execution of the online planner and learner ultimately generates results that meet threshold requirements, resulting in multiple discrete paths with reasonable speed allocations. These discrete trajectory points can be fitted with multiple segments of fifth-degree polynomials with respect to time t, serving as the initial solution to the joint spatiotemporal optimization problem. For the i-th segment of the polynomial, the coefficient vector can be represented as... Where m is set to 3, the segmentation can be represented as:

[0070]

[0071] β(t)=[1,t,t 2 ,...,t N ] T

[0072] Where t∈[0,bT], and i∈{1,2,...,M}.

[0073] In step A03, the method for forming the collision-free feasible region is as follows:

[0074] 1) Discretize the trajectory of the preceding guiding path continuity, and break the path into segments of broken lines;

[0075] 2) For each broken line segment, construct two parallel line segments on both sides of the broken line segment as the center. The parallel line segments extend to the left and right respectively, and the extension distance can be set to a fixed distance;

[0076] 3) Continue to divide the area divided by each parallel line segment into two sub-regions again according to the center line;

[0077] 4) Repeat the binary search operation until the preset accuracy is reached; once the preset accuracy is reached, the polyline segment can serve as the dividing line between areas with and without obstacles.

[0078] 5) For each polyline segment, repeat steps 1)-4) above to obtain the discretized parallel line segments and the unobstructed convex rectangular regions divided by each parallel line segment.

[0079] 6) Connect all the convex rectangular regions to form a collision-free path space.

[0080] Optionally, in step A04, the method for constructing the spatiotemporal joint multi-vehicle trajectory optimization problem using the initial trajectory results is as follows: Based on the fitted fifth-degree polynomial, obtain each segment of the fifth-degree polynomial, then discretize each segment of the fifth-degree polynomial into N smaller segments, and integrate the time term over the square of the cubic derivative of each smaller segment. This term serves as the energy cost term for controlling the vehicle. The overall trajectory time multiplied by the total time penalty coefficient represents the time cost term.

[0081] In implementation, the energy cost and time cost for each vehicle can be expressed as follows:

[0082]

[0083] Where μ(t) represents p (m) The vehicle's acceleration control force, W is a diagonal matrix used to represent the penalty weight of the control quantity, w T Let T be the penalty weight for the total time T. After discretizing into piecewise points, this integral expression can be represented as the sum of the squares of the control force at each discrete trajectory point.

[0084] Kinematic constraints can be expressed as

[0085]

[0086] in Let be the control variable for vehicle i, and j be the segment number of the trajectory. The spatial state variables and vehicle control variables in the kinematic state transition equations can be mutually converted. Through this conversion relationship, the variables relating to spatial state and control constraints can be uniformly represented in Cartesian coordinates using the vehicle's spatial position variables. The resulting higher-order polynomial for the segmented trajectory, such as a fifth-order polynomial, allows the position variables to be directly linked to the energy and time costs required by the optimization objective function.

[0087] The obstacle avoidance constraint involves simplifying its complex parts to create a series of easily solvable subproblems. For the safe obstacle avoidance constraint, a double-circular scheme is adopted.

[0088]

[0089] In the formula, P_f and P_r are the center coordinates of the front and rear centers of the enclosing vehicle, respectively. The subscripts i, j and x, y represent the x and y coordinates of the center coordinates of vehicles i and j in the Cartesian coordinate system. The scalar γ∈[0,1] is used to adjust the size of the two circles: when γ=0, the distance between the vehicles shrinks to two points with fixed relative positions; when γ=1, the distance reaches the limited limit of safety distance.

[0090] During the solution process, γ1 ∈ [0, 1) is first selected. Since γ1 is often set relatively small, subproblems with smaller safety distances can be solved successfully. If the solution fails, it means that the entire optimization problem has no solution after definition. If the solution succeeds, the successful solution is used as the initial solution, and the value of γ is gradually increased until it approaches the acceptable safety distance between vehicles.

[0091] Meanwhile, the vehicle's edge, enclosed by the double circles, must not collide with the boundaries of obstacles in the environment. In other words, the distance constraint min_dist between vehicle i and the nearest obstacle in the current environment can be established as follows:

[0092] κ+R i ≥min_dist,κ∈{P_f iy (t),P_r iy (t)},t∈[t0,t f ].

[0093] In summary, the multi-vehicle trajectory optimization problem is modeled as a nonlinear optimization problem. With minimizing the cost index as the objective function and vehicle pose as the optimization variable, solving this optimization problem can yield smoother paths that better conform to nonholonomic kinematic constraints.

[0094] The process of converting variables into convex feasible sets through continuous processing includes: using the two-dimensional planning plane as the base plane, abstracting the area occupied by the vehicle body into a regular shape with continuously closed edges, and defining the planning range of the vehicle body as multiple intersecting and connected rectangular frames extended from the path results generated by the graph search. Within the same time period, the frames of each vehicle path result will not intersect, so the vehicle planning results will not collide. The frames will also not intersect with obstacles in the two-dimensional planning environment.

[0095] In practice, centralized solutions to constrained optimization problems in multi-vehicle optimization suffer from reduced real-time performance as the number of vehicles increases. To address this issue, a priority-based trajectory grouping optimization method is proposed, incorporating a multi-vehicle trajectory priority approach. Vehicles are first divided into several groups with different priorities. Then, trajectory optimization subproblems are solved sequentially from the highest priority group to the lowest priority group. In each iteration, the trajectory of the current vehicle group is optimized while avoiding collisions with vehicles in higher priority groups. Using this decoupled framework, the trajectory optimization process runs relatively quickly. However, without determining reasonable priorities according to the algorithm, low-priority vehicles may lack feasible space, leading to infeasible subproblems. To address this problem, a novel grouping and priority allocation strategy is proposed, enabling the algorithm to find solutions with a higher probability.

[0096] The primary focus of priority trajectory optimization is the constraints between vehicles. This patent can analyze reference trajectories. Given the constraints among vehicles at each time step t, find three vehicles (i1, i2, i3) that satisfy the following conditions. t :

[0097]

[0098] in This is a threshold; each triple represents three vehicles that are relatively close to each other at a specific time step. In the discrete path planning phase, a scenario where all four vehicles simultaneously satisfy the above equation can only be found when they are simultaneously located at the four vertices of a square grid cell. This patent only analyzes this type of coupling and collision between three vehicles.

[0099] If three vehicles are assigned different priorities, the higher-priority vehicles will plan their trajectories first. In optimization problems involving lower-priority vehicles, these high-priority trajectories are considered hard constraints. In this case, high-priority constraints may hinder the optimization of lower-priority vehicles, potentially preventing the solution from reaching the final result. Therefore, the three vehicles in a triplet should be grouped together and assigned the same priority. The sequence of all triplets satisfying the condition is denoted by Tg.

[0100] If the number of vehicles to be planned in the workspace is large, Tg will contain a large number of elements, so a vehicle may be contained in different triples. In this case, vehicles cannot be directly grouped. Because each group represents the situation where three vehicles are close to each other, if a vehicle appears in multiple different Tgs, it means that the trajectories of the three vehicles are highly coupled, and the degree of coupling between paths will be used as one of the partitioning indicators.

[0101] First, the element with the highest frequency in Tg is selected as the first priority group. Since other vehicles in this new group may be contained in other Tgs, this patent removes them all from Tg, resulting in a new list Tg'. Therefore, each element in Tg' may contain three, two, or only one vehicle. Then, the most frequent element in the new list Tg' is selected as the second priority group. This process is repeated until the list is empty. Finally, the completeness of the algorithm is checked. If a vehicle has no coupling with other vehicles, it will not be included in the original list Tg. In this case, the vehicle is considered a single vehicle group and assigned the lowest priority. Using this algorithm, all vehicles in a centralized planning group can be grouped and assigned based on the frequency of coupling.

[0102] The foregoing results are based on the multi-machine path search in Chapter 3. Based on the estimated distance between reference trajectories between vehicles, those with high coupling are grouped into a high-priority group. Since the first-stage search also provides the estimated distance from the start point to the destination, a new dis_fitness metric is introduced to evaluate the trajectory. Specifically, in the multi-machine search stage, the length of the decision trajectory obtained by the vehicle path search can be determined by the step size of the single-machine sampling iteration during the search. More sampling times to reach the destination indicate a greater distance.

[0103] The paths generated by N vehicles are sorted according to the disfitness index, with two sorting schemes: ascending and descending. A higher disfitness indicates a lower task completion rate or a more complex path exploration process when following the decision trajectory. In sequential decision-making, prioritizing or deferring complex cases may yield better solutions; therefore, two orders are set. Combined with the aforementioned sorting scheme based on the coupling degree between multiple vehicles, three priority sorting schemes are obtained. Next, a superior sorting scheme is selected. The pseudocode of the grouped priority trajectory planning algorithm based on multi-path evaluation is shown below:

[0104]

[0105]

[0106] The energy cost term and time cost term of the fifth-order polynomial, along with the constraint penalty term conforming to kinematic constraints, obstacle avoidance constraints, and running time constraints, constitute a joint multi-vehicle trajectory optimization problem. In step A05, the method for grouping the optimization problem objects based on the path results of multiple vehicles is as follows: according to the preceding guidance path results, paths with high distance coupling are grouped into the same group. In step A06, the method for solving the spatiotemporal joint multi-vehicle trajectory optimization problem for multiple vehicles in the same group is as follows: using an optimization problem solver or transforming the constraint penalty term into an unconstrained optimization problem with a penalty function cost term to solve the multi-vehicle trajectory optimization problem for the same group.

[0107] The advantages of the execution method described in this embodiment are: it can consider multiple objectives simultaneously, such as time, distance, and energy consumption, and solve for high-quality trajectories with high efficiency; by transforming the expression of the unobstructed area around the vehicle and the collision constraints, the problem is transformed into a convex optimization form, thereby improving the efficiency of solving the optimization problem; it can consider the interaction and cooperative behavior between multiple vehicles to generate trajectory schemes with higher cooperativeness.

[0108] The specific embodiments described above further illustrate the purpose, technical solution, and beneficial effects of the present invention. It should be understood that the above description is only a specific embodiment of the present invention and is not intended to limit the scope of protection of the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A multi-vehicle spatiotemporal joint trajectory planning method, characterized in that: Includes the following steps, Step A01: Obtain the current location of multiple vehicles on the planning map, the start and end points of the delivery task, and the current vehicle status data; Step A02: Based on the path results of multiple vehicles, determine the initial trajectory results of multiple vehicles, and process the initial guidance path results of multiple vehicles into segments as initial trajectory results. Step A03: Determine the feasible and safe planning area in the trajectory planning of multiple vehicles. Based on the path results of multiple vehicles and combined with static or low-speed obstacle environment information, determine the collision-free feasible area. Step A04: Using the initial trajectory results, construct a spatiotemporal joint multi-vehicle trajectory optimization problem; Step A05: Group the objects of the spatiotemporal joint multi-vehicle trajectory optimization problem based on the path results of multiple vehicles; Step A06: For multiple vehicles in the same group, solve the spatiotemporal joint multi-vehicle trajectory optimization problem; Step A07: Search for collision-free paths for multiple vehicles.

2. The multi-vehicle spatiotemporal joint trajectory planning method according to claim 1, characterized in that: In step A02, the initial guidance path results of multiple vehicles are segmented and processed into initial trajectory results, which includes the following steps: after obtaining the previous guidance path results and discretizing them into points, a learning-based speed decision method is used to obtain trajectory points, and then the trajectory points are fitted to a fifth-order polynomial.

3. The multi-vehicle spatiotemporal joint trajectory planning method according to claim 2, characterized in that: The learning-based speed decision-making method includes the following steps: 1) Define the state space, action space, reward function, and state transition function; 2) Use a POMDP-based reinforcement learning algorithm to estimate the state-action value and obtain the optimal action value function and optimal policy for the path; 3) Using an online planner, the optimal strategy and optimal action value function are used to plan based on the current state, obtain the optimal action for the current state, and execute it; 4) The actions performed by the online planner and the rewards returned by the environment are used as experience and fed back to a reinforcement learning algorithm to update the policy and value function using the acquired experience; 5) Repeat steps 1)-4) above, continuously updating the strategy and value function, and using the latest strategy and value function for planning and control, and finally complete the speed allocation and trajectory planning of the path.

4. The multi-vehicle spatiotemporal joint trajectory planning method according to claim 2, characterized in that: Step A04, using the initial trajectory results, constructs the spatiotemporal joint multi-vehicle trajectory optimization problem as follows: based on the fitted fifth-degree polynomial, obtain each segment of the fifth-degree polynomial, then discretize each segment of the fifth-degree polynomial into N small segments, and integrate the time term of the square of the cubic derivative of each small segment.

5. The multi-vehicle spatiotemporal joint trajectory planning method according to claim 4, characterized in that: In step A03, the method for forming the collision-free feasible region includes the following steps: 1) Discretize the continuous trajectory of the preceding guidance path, and break the preceding guidance path into segments of broken lines; 2) For each broken line segment, with the broken line segment as the center line, construct two parallel line segments on both sides of the broken line segment. The parallel line segments extend to the left and right respectively, and the extension distance can be set to a fixed distance. 3) Continue to divide the area divided by each parallel line segment into two sub-regions again according to the center line; 4) Repeat the binary search operation in step 3) until the preset accuracy is reached. Once the preset accuracy is reached, use the polyline segment as the dividing line between the obstructed and unobstructed areas. 5) For each polyline segment, repeat steps 1)-4) above to obtain the discretized parallel line segments and the unobstructed convex rectangular regions divided by each parallel line segment; 6) Connect all the unobstructed convex rectangular regions to form a collision-free path space.

6. The multi-vehicle spatiotemporal joint trajectory planning method according to claim 2, characterized in that: The energy cost term and time cost term of the fifth-order polynomial, together with the constraint penalty term that conforms to kinematic constraints, obstacle avoidance constraints, and running time constraints, constitute a joint multi-vehicle trajectory optimization problem.

7. The multi-vehicle spatiotemporal joint trajectory planning method according to claim 6, characterized in that: The term obtained by integrating the square of the third derivative at each segment point with respect to the time term is used as the energy cost term for controlling the vehicle.

8. The multi-vehicle spatiotemporal joint trajectory planning method according to claim 7, characterized in that: The total trajectory time multiplied by the total time penalty coefficient represents the time cost term.

9. The multi-vehicle spatiotemporal joint trajectory planning method according to claim 2, characterized in that: In step A05, the method for grouping the objects of the optimization problem based on the path results of multiple vehicles is as follows: according to the results of the preceding guidance path, the path points with high coupling degree are divided into the same group.

10. The multi-vehicle spatiotemporal joint trajectory planning method according to claim 2, characterized in that: In step A06, the method for solving the spatiotemporal joint multi-vehicle trajectory optimization problem for multiple vehicles in the same group is to use an optimization problem solver or to transform the constraint penalty terms into penalty function cost terms to solve the multi-vehicle trajectory optimization problem in the same group.

Images