A robot environment detection method based on Bayesian kernel reasoning

Abstract

The present invention discloses a robot environment detection method based on Bayesian kernel reasoning, comprising the following steps: initializing a set of historical waypoints of the robot, an information gain threshold, and inference model training parameters; randomly sampling a set of kinematically reachable poses within the robot's current perception range as candidate control actions; explicitly calculating the mutual information of each candidate action and establishing a training sample set; continuing to sample a certain number of candidate actions as query samples, and predicting the information gain and uncertainty of the query samples using a Bayesian kernel reasoning method; deciding the optimal candidate action based on the predicted results of the query samples, and finally executing the optimal candidate action and updating the environment map until the detection is completed. The Bayesian kernel reasoning detection method of the present invention can predict the information gain and uncertainty of candidate points in the environmental space, has low computational complexity, and is suitable for robots to achieve online and safe detection in large-scale, cluttered, unknown environments.

Description

Technical Field

[0001] The present invention belongs to the technical field of robot environment detection, and mainly relates to a robot environment detection method based on Bayesian kernel reasoning. Background Art

[0002] Active robot exploration technology primarily involves proactively planning and selecting observation locations that maximize environmental information collection based on the robot's measurement data, without prior maps or planning. This technology meets multiple criteria, including accessibility, safety, and resource efficiency, while simultaneously completing tasks such as autonomous positioning and environmental mapping. The advantage of active robot exploration lies in its consideration of multiple constraints, including information gain, accessibility, and positioning accuracy. It provides real-time optimal decision-making and is well-suited for robotic missions in unknown environments, such as civilian applications like deep-sea exploration, pipeline and mine exploration, and field rescue, as well as unmanned reconnaissance missions in the military.

[0003] In recent years, some scholars have introduced theoretical results in information theory, such as Shannon entropy or mutual information (MI), into the field of robot environment detection technology, constructed information theory objective functions, and evaluated at what position and orientation the sensor can reduce the uncertainty of the map more, thereby naturally combining detection and optimization problems. For example, Julian et al. from MIT [1] used an occupancy grid map (OGM) and a narrow beam range sensor robot to prove that any controller that can maximize the mutual information reward function will eventually traverse an unexplored space.

[0004] In order to reduce the high computational cost of MI, some scholars have considered skeletonizing the known occupancy map or evaluating the information gain on a limited number of motion primitives, 3D viewpoints or occupancy grid maps to consider candidate trajectories [2]. For example, Bai et al. from Stevens Institute of Technology proposed an exploration method based on supervised learning to try to infer the information gain associated with candidate viewpoints, such as using Gaussian Process (GP) as a model or improving the sampled viewpoints through Bayesian Optimization (BO) based on GP priors [3].

[0005] However, robot environment detection methods based on supervised learning, especially those based on GP or BO with GP prior, are limited by the cubic complexity of their model update and sample prediction, and are difficult to apply to large-scale unstructured and cluttered environments, because in these environments, a large amount of information gain of candidate actions needs to be evaluated to obtain the best next executable action of the robot and ensure sufficient detection accuracy.

[0006] From this point of view, how to improve the reasoning speed of the robot environment detection method based on supervised learning without significantly losing the detection accuracy is a difficult problem that needs to be solved urgently.

[0007] [References]

[0008] [1]Julian BJ,Karaman S,Rus D.On mutual information-based control of range sensing robots for mapping applications[J].The International Journal ofRobotics Research,2014,33(10):1375-1392.

[0009] [2]Charrow B,Kahn G,Patil S,et al.Information-Theoretic Planning with Trajectory Optimization for Dense 3D Mapping[C] / / Robotics:Science andSystems.2015,11:3-12.

[0010] [3]Bai S,Wang J,Chen F,et al.Information-theoretic exploration withBayesian optimization[C] / / 2016IEEE / RSJ International Conference onIntelligent Robots and Systems(IROS).IEEE,2016:1816-1822. Summary of the Invention

[0011] The purpose of the present invention is to provide a robot environment detection method based on Bayesian kernel inference to address the problem of low robot detection efficiency in unknown environments. The Bayesian kernel inference technology (BKI) is applied to robot detection, so that the information gain of a large number of candidate actions can be quickly and efficiently predicted, and the final decision is made to obtain the next optimal detection action of the robot until the environmental detection task is completed.

[0012] In order to solve the above problems, the present invention provides the following technical solutions:

[0013] A robot environment detection method based on Bayesian kernel reasoning includes the following steps:

[0014] Step 1: Initialize the robot's historical action set x hist , information gain threshold I th , the number of training samples N that need to be calculated, the number of time steps N required for environmental detection s 、Number of samples to be queried q and the number of model training times N epoch ;

[0015] Step 2: During the kth iteration, the robot’s current position x k Can be obtained through independent robot positioning methods; according to the currently established map m k , randomly sample N kinematically reachable poses x={x1,…, i ,…, N},∈[1,N] as candidate actions, each candidate action includes the robot's target position coordinates and target heading angle;

[0016] Step 3: Explicitly evaluate each candidate action x based on the mutual information definition of the occupied grid map environment i ∈x corresponding to the information gain, establish a training sample set;

[0017] Specifically, traverse each candidate action x i ∈x, using the currently explored map m k , the virtual measurement Z′ is obtained by ray casting, and the MI corresponding to each candidate action is calculated according to the mutual information calculation formula: i =H(m k )-H(m k ; Z′), where H(·) is the information entropy function, and y i Store in MI set y; after traversal, obtain the training sample set

[0018] Step 4: Continue to randomly sample N within the current robot perception range q kinematically reachable candidate actions to obtain the input set x of query samples * , combined with the training sample set obtained in step 3, the MI corresponding to the query sample is predicted by the Bayesian kernel inference method to obtain the optimal candidate action index set x in each training best , and the corresponding optimal MI set I best ;

[0019] Specifically, initialize the optimal candidate action index set x best , optimal MI set I best ; For each training process, first according to x* and x to calculate the kernel function K(x,x * ); then according to K(x,x * ) and y to get x * Corresponding MI I * and the corresponding uncertainty

[0020] Given a training dataset Assume that there is an implicit mapping relationship from x to y in the data set, which can be expressed by the probability expression Description, assuming that the candidate action set to be queried is x * Also obeying this relationship, the corresponding MIy * It can be expressed as:

[0021]

[0022] where θ * For x * The relevant hyperparameters. Assume that θ * with y * There is a prior Gaussian likelihood relationship: where μ * N q dimensional unknown mean vector, Σ * N q ×N q dimensional known covariance diagonal matrix, whose diagonal values ​​are all known quantities σ 2 , where is the corresponding standard deviation constant, p(y * |θ * ) can also be represented by a Gaussian distribution: Where μ0 and ζ are initial confidence constant parameters. So given the training data set The sample input set x to be queried * The hyperparameter θ * The posterior distribution of can be expressed as:

[0023]

[0024] where μ i y i The corresponding sample mean, then the query sample input set x * The corresponding predicted value of MI I * and forecast uncertainty They are:

[0025]

[0026] Then, construct the objective function α∈(0,1) is used to balance the predicted MI and its uncertainty, and finally the candidate action with the largest objective function value is selected. If x s Already in the training sample x, then x s Add to set x best , and MIy s Add to collection I best If x s If it is not in the training sample x, then first explicitly calculate x s MIy s , then x s Add to set x best , and MIy s Add to collection I best , and {x s ,y s}Add to the training sample set In the process, until the number of training times reaches N epoch After the training is finished;

[0027] Step 5: If the currently predicted optimal MI set I best The maximum MI value is greater than the set information threshold I th , then the candidate action x corresponding to the maximum MI value best (MaxInfoIdx is the waypoint x that the robot will reach next. k+1 , where MaxInfoIdx is its best The index of x k+1 Add to the historical waypoint collection x hist ;

[0028] Step 6: If I best Not greater than the information threshold I th , then the robot waypoint x in the previous time step k-1 as the next waypoint x to be reached k+1 , while from x hist Delete x k-1 ;

[0029] Step 7: Robot Executes Waypoint X k+1 Instruction, using local path planning method to get from x k to x k+1 The local path P local While executing the path process, the robot uses the onboard sensors to continuously measure and update the occupied grid map, and obtains the environment map m at time k+1 k+1 ;

[0030] Step 8: Repeat steps 2 to 7 until the number of iterations reaches N.s End the detection process.

[0031] Compared with the prior art, the present invention has the following beneficial effects:

[0032] This paper implements a robot environment detection method based on Bayesian kernel reasoning, constructs a continuous alternative evaluation model between robot control actions and mutual information, and can efficiently predict the information gain brought by robot control actions to the surrounding environment. Compared with the current mainstream methods, the time complexity of the MI prediction process of this invention is only Much smaller than the robot environment detection method based on BO [3] It greatly speeds up the reasoning speed and is more suitable for large-scale unknown environment detection that requires a large number of candidate action queries. It proposes an objective function consisting of the predicted MI value and its uncertainty, which can balance the robot's own risk and information gain when deciding the best action for each step, thereby achieving safer robot detection. BRIEF DESCRIPTION OF THE DRAWINGS

[0033] Figure 1 Schematic diagram of a single-step decision-making scenario in the robot detection process in the method of the present invention.

[0034] Figure 2 Schematic diagram of the scene and robot detection trajectory in an embodiment of the method of the present invention.

[0035] Figure 3 Schematic diagram of the mutual information of the environment after the exploration is completed in an embodiment of the method of the present invention.

[0036] Figure 4 Schematic diagram comparing changes in map entropy in an embodiment of the method of the present invention.

[0037] Figure 5 Schematic diagram comparing changes in map exploration coverage in an embodiment of the method of the present invention. DETAILED DESCRIPTION

[0038] The following describes in detail the implementation of the method of the present invention with reference to the accompanying drawings.

[0039] refer to Figure 1 — Figure 5 , a robot environment detection method based on Bayesian kernel reasoning, the specific steps include:

[0040] Step 1: Initialize the robot's historical waypoint set x hist , information gain threshold I th , the number of training samples N that need to be calculated, the number of time steps N required for environmental detection s 、Number of samples to be queried q and the number of model training times Nepoch ;

[0041] Step 2: During the kth iteration, the robot’s current position x k It can be obtained through independent robot positioning methods (such as Gmapping positioning algorithm, etc.); Figure 1 As shown, according to the currently established map m k In order to avoid obstacles in the environment and areas that the robot cannot reach, N kinematically reachable poses x={x1,…,x i ,…,x N}, i∈[1,N] is a candidate action, each candidate action includes position coordinates and heading angle;

[0042] Step 3: Explicitly evaluate each candidate action x based on the mutual information definition of the occupied grid map environment i ∈x corresponding to the information gain, establish a training sample set;

[0043] Specifically, traverse each candidate action x i ∈x, using the currently explored map m k , the virtual measurement Z′ is obtained by ray casting, and the MI corresponding to each candidate action is calculated according to the mutual information calculation formula: i =H(m k )-H(m k ; Z′), where H(·) is the information entropy function, and y i Store in MI set y; after traversal, obtain the training sample set

[0044] Step 4: Continue to randomly sample N within the current robot perception range q kinematically reachable candidate actions to obtain the input set x of query samples * , combined with the training sample set obtained in step 3, the MI corresponding to the query sample is predicted by the Bayesian kernel inference method to obtain the optimal candidate action index set x in each training best , and the corresponding optimal MI set I best ;

[0045] Specifically, initialize the optimal candidate action index set x best , optimal MI set I best ; For each training process, first according to x * and x to calculate the kernel function K(x,x * ); then according to K(x,x * ) and y to get x * Corresponding MI I *and the corresponding uncertainty

[0046] Preferably, this embodiment uses a non-stationary Matern kernel function:

[0047]

[0048] Where l is a constant that adjusts the smoothness of the kernel function. Compared to the commonly used Gaussian kernel function, this non-stationary kernel function is less smooth and is more suitable for cluttered unknown environments where information gain is prone to sudden changes.

[0049] Given a training dataset Assume that there is an implicit mapping relationship from x to y in the data set, which can be expressed by the probability expression Description, assuming that the candidate action set to be queried is x * Also obeying this relationship, the corresponding MIy * It can be expressed as:

[0050]

[0051] where θ * For x * The relevant hyperparameters. Assume that θ * with y * There is a prior Gaussian likelihood relationship: where μ * N q dimensional unknown mean vector, Σ * N q ×N q dimensional known covariance diagonal matrix, whose diagonal values ​​are all known quantities σ 2 , where is the corresponding standard deviation constant, p(y * |θ * ) can also be represented by a Gaussian distribution: Where μ0 and ζ are initial confidence constant parameters. Generally, ζ is 0, which means no prior knowledge, and infinity, which means sufficient prior knowledge. In this embodiment, a value close to 0 means that there is basically no prior knowledge in the unknown area. So given a training data set The sample input set x to be queried * The hyperparameter θ * The posterior distribution of can be expressed as:

[0052]

[0053] where μ i y i The corresponding sample mean, then the query sample input set x *The corresponding predicted value of MI I * and forecast uncertainty They are:

[0054]

[0055] Then, construct the objective function α∈(0,1) to balance the predicted MI and its uncertainty, and finally select the candidate action with the largest objective function value If x s Already in the training sample x, then x s Add to set x best , and MIy s Add to collection I best If x s If it is not in the training sample x, then first explicitly calculate x s MIy s , then x s Add to set x best , and MIy s Add to collection I best , and {x s ,y s}Add to the training sample set In the process, until the number of training times reaches N epoch After the training is finished;

[0056] Step 5: If the currently predicted optimal MI set I best The maximum MI value is greater than the set information threshold I th , then the candidate action x corresponding to the maximum MI value best (MaxInfoIdx is the waypoint x that the robot will reach next. k+1 , where MaxInfoIdx is its best The index of x k+1 Add to the historical waypoint collection x hist ;

[0057] Step 6: If I best Not greater than the information threshold I th , then the robot waypoint x in the previous time step k-1 as the next waypoint x to be reached k+1 , while from x hist Delete x k-1 ;

[0058] Step 7: Robot Executes Waypoint X k+1 Instruction, using local path planning method to get from x k to x k+1 The local path Plocal While executing the path process, the robot uses the onboard sensors to continuously measure and update the occupied grid map, and obtains the environment map m at time k+1 k+1 ;

[0059] Preferably, the local path planning method of this embodiment adopts the well-known A* algorithm, which can achieve local path optimization and obstacle avoidance.

[0060] Step 8: Repeat steps 2 to 7 until the number of iterations reaches N. s End the detection process.

[0061] The above implementation process is applied to a large-scale robot exploration task, where the robot uses a lidar as a distance sensor. The BO method in reference [3] is used as a comparison to verify the superiority of the BKI method of the present invention.

[0062] In this embodiment, if Figure 2 The robot is located in an indoor environment and is equipped with a two-dimensional laser radar that can emit 20 evenly distributed beams within a horizontal viewing angle range of ±π / 3rad, with a maximum ranging range of 4 meters, a map size of 24 meters × 14 meters, and a map resolution of 0.2 meters per grid. Table 1 shows the relevant parameters used in the embodiment of the present invention. In order to verify the computing performance advantages of the present invention under large sample sizes and large-scale training, N and N epoch When N=30, N epoch =15, and N = 60 epoch =30, so each method is labeled with two different N values: BO 30, BO 60, BKI 30, and BKI 60. In this example, 50 Monte Carlo experiments were conducted. Table 2 shows a comparison of the mean and standard deviation of the single-step time consumption for each method in this example. It can be seen that the BKI method proposed in this invention significantly improves single-step computational performance compared to the BO method, with a 25% improvement for N = 30 and a 53% improvement for N = 60, demonstrating the computational efficiency advantage of the method proposed in this invention.

[0063] From the perspective of detection effect, the final planned path and the explored map of this embodiment are as follows: Figure 2 As shown, the final generated environment MI is as follows Figure 3 As shown, Figure 4 and Figure 5 The map entropy changes and map coverage changes of this embodiment are shown in Figure 2. It can be seen that the map entropy decreases faster and the coverage changes faster in the early stage of the method proposed in the present invention, and the final map entropy and coverage are better than the BO method.

[0064] In summary, the robot environment detection method based on Bayesian kernel reasoning proposed in the present invention has higher computational efficiency and stronger exploration performance, which demonstrates the effectiveness and superiority of the present invention.

[0065] Table 1 Basic parameter settings

[0066]

[0067]

[0068] Table 2 Comparison of single-step time consumption of each method

[0069] Comparison Method Time consumed (unit: seconds) BO 30 0.5203±0.0554 BO 60 1.0528±0.1124 BKI 30 0.3903±0.1175 BKI 60 0.4955±0.1775

[0070] The embodiments of the present invention are described above in conjunction with the accompanying drawings, but the present invention is not limited to the above-mentioned specific embodiments. The above-mentioned specific embodiments are merely illustrative and not restrictive. Under the guidance of the present invention, ordinary technicians in this field can also make many forms without departing from the scope of protection of the present invention and the claims, all of which fall within the scope of protection of the present invention.

Claims

1. A robot environment detection method based on Bayesian kernel reasoning, characterized in that: The following steps are involved: Step 1: Initialize the robot's historical action set x hist , information gain threshold I th , the number of training samples N that need to be calculated, the number of time steps N required for environmental detection s 、Number of samples to be queried q and the number of model training times N epoch ; Step 2: During the kth iteration, the robot’s current position x k Can be obtained through independent robot positioning methods; according to the currently established map m k , randomly sample N kinematically reachable poses x={x1,…,x i ,…,x N }, i∈[1,N] is a candidate action, each candidate action includes the robot's target position coordinates and target heading angle; Step 3: Explicitly evaluate each candidate action x based on the mutual information definition of the occupied grid map environment i ∈x corresponding to the information gain, get the MI set y, and establish the training sample set; Step 4: Continue to randomly sample N within the current robot perception range q kinematically reachable candidate actions to obtain the input set x of query samples * , combined with the training sample set obtained in step 3, the MI corresponding to the query sample is predicted by the Bayesian kernel inference method to obtain the optimal candidate action index set x in each training best , and the corresponding optimal MI set I best ; Step 5: If the currently predicted optimal MI set I best The maximum MI value is greater than the set information threshold I th , then the candidate action x corresponding to the maximum MI value best (MaxInfoIdx) is the waypoint x that the robot will reach next. k+1 , where MaxInfoIdx is its best The index of x k+1 Add to the historical action set x hist ; Step 6: If I best Not greater than the information threshold I th , then the robot action x in the previous time step is k-1 As the next action to be performed x k+1 , while from x hist Delete x k-1 ; Step 7: The robot performs action x k+1 Instruction, using local path planning method to get from x k to x k+1 The local path P local While the robot is traveling along the path, it uses the onboard sensors to continuously measure and update the occupied grid map, and obtains the environment map m at time k+1 k+1 ; Step 8: Repeat steps 2 to 7 until the number of iterations reaches N. s End the detection process.

2. A robot environment detection method based on Bayesian kernel reasoning according to claim 1, characterized in that: The process of establishing the training sample set in step 3 is specifically as follows: Traverse each candidate action x i ∈x, using the currently explored map m k , the virtual measurement Z′ is obtained by the ray casting method, and the MI corresponding to each candidate action is calculated according to the mutual information calculation formula: i =H(m k )-H(m k ; Z′), where H(·) is the information entropy function, and y i Store in MI set y; after traversal, obtain the training sample set 3. The robot environment detection method based on Bayesian kernel reasoning according to claim 1 is characterized in that: The specific steps of predicting the MI corresponding to the query sample by the Bayesian kernel reasoning method in step 4 are as follows: Initialize the optimal candidate action index set x best , optimal MI set I best ; For each training process, first according to x * and x to calculate the kernel function K(x,x * ); then according to K(x,x * ) and y to get x * Corresponding MII * and the corresponding uncertainty Given a training dataset Assume that there is an implicit mapping relationship from x to y in the data set, which can be expressed by the probability expression Description, assuming that the candidate action set to be queried is x * Also obeying this relationship, the corresponding MI y * It can be expressed as: where θ * For x * Regarding the hyperparameters, assume that θ * with y * There is a prior Gaussian likelihood relationship: where μ * N q dimensional unknown mean vector, Σ * N q ×N q dimensional known covariance diagonal matrix, whose diagonal values ​​are all known quantities σ 2 , where is the corresponding standard deviation constant, p(y * |θ * ) can also be represented by a Gaussian distribution: Where μ0 and ζ are initial confidence constant parameters, so given the training data set The sample input set x to be queried * The hyperparameter θ * The posterior distribution of can be expressed as: where μ i y i The corresponding sample mean, then the query sample input set x * The corresponding predicted value of MI I * and forecast uncertainty They are: Then, construct the objective function α∈(0,1) is used to balance the predicted MI and its uncertainty, and finally the candidate action with the largest objective function value is selected. If x s Already in the training sample x, then x s Add to set x best , and MIy s Add to collection I best If x s If it is not in the training sample x, then first explicitly calculate x s MIy s , then x s Add to set x best , and MIy s Add to collection I best , and {x s ,y s }Add to the training sample set In the process, until the number of training times reaches N epoch Then the training ends.

4. The robot environment detection method based on Bayesian kernel reasoning according to claim 3 is characterized in that: The kernel function adopts the non-stationary Matern kernel function: in is a constant that adjusts the smoothness of the kernel function.

5. The robot environment detection method based on Bayesian kernel reasoning according to claim 1 is characterized in that: The airborne sensors include lidar, depth vision camera, ultrasonic array, and sonar.

6. The robot environment detection method based on Bayesian kernel reasoning according to claim 1 is characterized in that: The local path planning method adopts the A* algorithm.

Images