AI edge computing host system in industrial control scene

Abstract

The invention relates to the technical field of artificial intelligence, in particular to an AI edge computing host system in an industrial control scene. The system comprises a mobile robot end and an edge computing server end. The mobile robot end is responsible for collecting environment data and a motion state, and the edge computing server end performs obstacle detection and global path planning by using a deep neural network. When the robot receives a lagging path planning result returned by the edge end, searching a matched historical pose state by analyzing an original acquisition timestamp, and calculating a relative transformation matrix between the historical pose state and the current real-time pose state; based on the matrix, the system maps a lagged path point sequence to a current robot coordinate system to generate a real-time tracking path. According to the scheme, the powerful computing power of the edge end is utilized to plan a complex route, the control deviation caused by network lag is effectively eliminated through time delay compensation, and autonomous navigation with low cost, low power consumption and high precision is realized.

Description

Technical Field

[0001] This invention relates to the field of artificial intelligence technology, specifically to an AI edge computing host system for industrial control scenarios. Background Technology

[0002] With the rapid development of smart logistics and the service robot industry, mobile robots have been widely used in scenarios such as warehousing and logistics, hospital delivery, intelligent inspection, and last-mile delivery in cities. In order to operate safely and efficiently in unstructured and highly dynamic environments, mobile robots must have strong environmental perception, localization and navigation, semantic understanding, and path planning capabilities.

[0003] Traditional mobile robot navigation systems typically employ an "onboard computing" architecture, where all sensor data processing, SLAM mapping, obstacle detection, and path planning tasks are performed on an embedded computer mounted on the robot itself. However, with advancements in deep learning and computer vision technologies, the computational resource demands of high-precision object detection algorithms and gradient-based path optimization algorithms are growing exponentially. Running these algorithms on mobile devices requires high-performance GPUs or TPUs, which significantly increases robot manufacturing costs, boosts system power consumption, reduces battery life, and necessitates additional heat dissipation designs. Furthermore, onboard computing units have long hardware upgrade cycles, making them ill-suited for the rapidly iterating development of artificial intelligence algorithms.

[0004] To address these challenges, edge computing architecture has emerged. This architecture offloads computationally intensive tasks (such as LiDAR point cloud processing, visual semantic segmentation, and global path planning) to network edge servers or base stations. Edge servers are capable of running larger-scale neural network models, providing a higher level of intelligent decision-making.

[0005] However, edge computing architecture introduces a new problem—latency. In edge-assisted navigation, the total latency includes not only the uplink transmission latency of sensor data and the downlink transmission latency of control commands, but also the computational latency required for edge servers to perform deep learning inference.

[0006] While 5G networks and Wi-Fi technologies have significantly reduced communication latency, inference latency is often rigid and fluctuating. For example, performing panoramic segmentation on a single high-resolution image frame can take 50ms to 150ms. During this time, the actual position and posture of a high-speed moving robot have already changed. If the robot directly executes path instructions calculated by the edge server based on previous sensor data, it may lead to problems such as trajectory tracking errors and obstacle avoidance failures. Furthermore, latency can cause phase lag in the control system, reducing phase margin and potentially causing oscillations or even divergence in the control loop.

[0007] In existing technologies, some solutions employ a "go-stop-go" strategy to mitigate latency effects, but this significantly reduces operational efficiency. Therefore, a system is needed that can fully utilize the advanced inference capabilities of edge computing while effectively compensating for inference latency and network latency. Summary of the Invention

[0008] This invention provides an AI edge computing host system for industrial control scenarios, including a mobile robot terminal and an edge computing server. The mobile robot terminal includes a data acquisition module, a local control module, and a clock synchronization module; The data acquisition module is used to collect environmental data and motion status data of the mobile robot; The local control module is used to perform latency compensation on the path planning results returned by the edge computing server and generate a real-time tracking path that is aligned with the current physical environment. The clock synchronization module is used to maintain time synchronization with the edge computing server using a precise time protocol; The edge computing server includes an inference module and a global path planning module, and communicates with the mobile robot terminal via a wireless network. The inference module receives environmental data from the mobile robot and uses a deep neural network to detect obstacles. The global path planning module performs global path planning based on the results generated by the inference module, and generates the optimal path point sequence.

[0009] Preferably, the inference module receives environmental data from the mobile robot and uses a deep neural network for obstacle detection, specifically including: Receive the perception data packet at time t, parse out the environmental data frame, and perform denoising, normalization and size scaling on it to generate tensor data that meets the input requirements of the pre-trained model; The tensor data is input into the backbone network of a pre-trained deep neural network, which includes multiple convolutional layers or attention mechanism layers to extract low-level texture features and high-level semantic features of the environmental data layer by layer, generating multi-scale feature maps. The feature map is input into the segmentation head of the model, upsampled and fused, and the pixel semantic mask is output. By combining the sensor intrinsic and extrinsic parameter matrices, the aforementioned semantic mask is projected onto the robot's top-down coordinate system to construct a local semantic occupancy grid map at time t, where each grid cell stores the probability of the obstacle's presence and semantic category information.

[0010] Preferably, the global path planning module performs global path planning based on the results generated by the inference module to generate an optimal path point sequence, specifically including: Based on the semantic category information in the grid map, different expansion radii and cost values ​​are assigned to different types of obstacles; for the semantic labels of dynamic attributes, a Gaussian-distributed prediction cost field is generated around the obstacle. Starting from the robot's position at time t and ending with the projection of the target point on the local map, a search algorithm is run on the cost map to generate an optimal path sequence that avoids high-cost regions. The optimal path point sequence is repackaged with the original timestamps to generate navigation control commands, which are then sent back to the mobile robot.

[0011] Preferably, the step of performing latency compensation on the path planning results returned by the edge computing server to generate a real-time tracking path aligned with the current physical environment specifically includes: During operation, the local control module reads motion data from the robot's odometry and inertial measurement unit at a fixed frequency, and stores the state vector containing timestamps, global position, heading angle, linear velocity and angular velocity into a circular state buffer. When the navigation path packet sent back by the edge computing server is received, the original collection timestamp carried in the data packet is parsed. A binary search or linear interpolation is performed in the circular state buffer to retrieve the robot's historical pose state that is closest to the original timestamp. Obtain the robot's real-time pose state at the current moment, and calculate the robot's relative pose transformation matrix from the historical moment to the current moment based on the historical pose state and the real-time pose state. Based on the relative positional transformation matrix, each point in the path point sequence sent by the edge computing server is transformed into a target point relative to the current robot body coordinate system, thereby generating a real-time tracking path aligned with the current physical environment.

[0012] Preferably, the local control module further includes a timeliness verification unit, which is used for: After parsing and obtaining the original acquisition timestamp carried in the navigation path packet, calculate the lag time of the navigation path packet from the current time. If the lag time exceeds the preset safety threshold, the edge calculation result of the frame is determined to be invalid, the path data is discarded, and the local control logic is triggered to control the robot to perform deceleration or restricted local obstacle avoidance actions until path data that meets the requirements is received.

[0013] A method for AI edge computing host in industrial control scenarios includes the following steps: The mobile robot triggers the data acquisition module to collect environmental data at time T, and records the robot's global pose at that moment; the timestamp T is used as a unique identifier to label the environmental data and the global pose data respectively, and the two are stored together. The environmental data and timestamp T are encapsulated and sent to the edge server via a wireless network; After receiving the data packet, the edge server inputs the environmental data into a deep convolutional neural network to perform semantic segmentation or object detection, generating a local occupancy grid map containing the distribution of dynamic and static obstacles. The edge server runs a search algorithm on the locally occupied grid map, starting from the robot's pose at data acquisition time T, to plan the optimal path point sequence to avoid obstacles. The generated path point sequence is packaged together with the original timestamps into navigation control instructions and sent back to the mobile robot. After receiving navigation control instructions from the edge server, the mobile robot calculates the time difference between the current time and the data acquisition time, and calculates the displacement vector and rotation angle based on the integration of local odometry and inertial measurement unit data over the time interval. Apply inverse rigid body transformation to the received optimal path point sequence to generate a new path aligned with the robot's current pose; The robot is driven by speed control commands generated based on the newly generated path.

[0014] Compared with the prior art, the advantages of this invention are: By using an explicit time delay compensation algorithm, the phase lag caused by inference time is eliminated, enabling the robot to accurately execute complex paths sent from the edge side even at high speeds.

[0015] By using edge servers to run complex deep neural networks for refined semantic segmentation and obstacle detection, the system can not only distinguish the geometric location of obstacles, but also identify their semantic categories (such as roads, grass, and pedestrians). Based on this, it constructs a refined semantic cost map that includes static cost, inflation cost, and dynamic risk cost, thereby planning an optimal path that is more in line with human logic and safer.

[0016] The timeliness verification unit calculates the lag time of path data before performing latency compensation. If the lag time exceeds a preset safety threshold, the system will automatically determine that the frame data is invalid and immediately trigger a local degradation strategy (deceleration or local obstacle avoidance). This effectively prevents the robot from executing outdated instructions with extreme errors due to network congestion or service crashes. Attached Figure Description

[0017] Figure 1 This is a system architecture diagram of an AI edge computing host system for industrial control scenarios proposed in this invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. All other embodiments obtained by those skilled in the art based on the embodiments of the present invention without creative effort are within the scope of protection of the present invention.

[0019] The present invention provides an AI edge computing host system for industrial control scenarios, which mainly consists of a mobile operation robot and an edge computing server, and the two are connected through a low-latency wireless network.

[0020] The mobile robot terminal includes a data acquisition module, a local control module, and a clock synchronization module.

[0021] The data acquisition module is used to collect environmental data and motion status data of the mobile robot.

[0022] The local control module is used to perform latency compensation on the path planning results returned by the edge computing server and generate a real-time tracking path that is aligned with the current physical environment.

[0023] The clock synchronization module is used to maintain time synchronization with the edge computing server using a precise time protocol.

[0024] The edge computing server includes an inference module and a global path planning module, and communicates with the mobile robot terminal via a wireless network.

[0025] The inference module receives environmental data from the mobile robot and uses a deep neural network to detect obstacles.

[0026] The global path planning module performs global path planning based on the results generated by the inference module, and generates the optimal path point sequence.

[0027] The mobile robot's sensor module uses an equipped RGB-D depth camera to collect environmental data around the robot and transmits the collected data to an edge computing server via a wireless connection.

[0028] After receiving the data packet, the inference module of the edge computing server first performs time-series alignment of the data based on the timestamp. Then, it performs distortion correction, histogram equalization, and size normalization operations on the received image data to generate a tensor. This generated tensor is then input into the inference module.

[0029] The inference module includes a pre-trained deep neural network model, which is a convolutional neural network with an encoder-decoder architecture.

[0030] The received tensor data first enters the backbone network of the convolutional neural network. After multi-scale features are extracted by ResNet-101 or MixTransformer, a set of feature maps is generated, which includes low-level features such as environmental texture and shape, as well as semantic features of object categories.

[0031] Subsequently, the semantic segmentation head of the deep neural network upsamples and fuses the feature maps, outputting a pixel-level semantic mask. , where C is the total number of semantic categories (such as roads, pedestrians, vehicles, grass, etc.). Each pixel stores the probability distribution of which category it belongs to.

[0032] For navigation, the aforementioned forward-looking semantic mask needs to be transformed into a top-down coordinate system. Specifically: First, pre-calibrated extrinsic parameters are used to align the pixels in the semantic mask with the depth values ​​in the depth map to the same physical space point.

[0033] By using the camera intrinsic parameter matrix, each pixel on the image plane and its corresponding depth value are mapped back to the 3D camera coordinate system, generating a set of 3D point clouds with semantic labels.

[0034] The point cloud is transformed from the "camera coordinate system" to the "robot body coordinate system". During this process, points that are higher than the robot's own height (suspended objects) or lower than the ground threshold (ground noise) are removed, and only obstacle points and ground points that affect navigation are retained.

[0035] Create a local scrolling grid map centered on the robot, with dimensions of L×W (e.g., 10m×10m) and a resolution of r (e.g., 0.05m).

[0036] The preserved 3D points are projected onto the raster map plane, and their indices in the raster map are calculated.

[0037] Since multiple visual pixels may be projected into the same grid cell and may carry different semantic prediction results, the system employs a probabilistic fusion mechanism. Each grid cell maintains a probability vector of length C (number of categories). Whenever a point belonging to category k falls into the grid, the confidence of that category within the grid is increased.

[0038] Finally, all rasters are traversed, and the semantic category with the highest probability is selected as the final attribute of that raster.

[0039] Through the above steps, the visual image from the forward-looking perspective is "transmitted" onto a local top-down map that the robot can understand, with semantic labels such as "road, people, vehicles, and obstacles".

[0040] The global path planning module receives the semantic raster map described above and generates the optimal path from the current location to the target point. Specifically: S1. Generate a semantic cost map: For the semantic raster map described above, the "travel cost" of each raster is calculated. The formula for calculating the travel cost is as follows: ; in: Let (x, y) be the cost of travel.

[0041] The static semantic cost is assigned based on the semantic category of the raster. For example: The road cost is assigned a value of 1.0 (low cost, priority passage).

[0042] The cost of grassland is set to 5.0 (medium cost, viable but not recommended).

[0043] The cost of an obstacle is assigned as positive infinity (impassable).

[0044] To compensate for the expansion cost, the robot decays outwards from the obstacle in a Gaussian distribution, ensuring that the robot maintains a safe distance from the obstacle.

[0045] For dynamic risk costs, for dynamic objects identified as "pedestrians" or "other robots", additional risk costs are added to the areas they may pass through in the future, based on their predicted velocity vectors.

[0046] Based on the semantic cost map generated above, a hybrid A* algorithm is used for path search.

[0047] The original path generated by the search often contains a polyline. The module uses gradient descent to optimize the sequence of path points and finally outputs a trajectory vector containing position, heading, and recommended speed.

[0048] The optimal path point sequence is repackaged with the original timestamps to generate navigation control commands, which are then sent back to the mobile robot.

[0049] The aforementioned reasoning and planning process is computationally intensive, typically taking between 50ms and 200ms. Therefore, the generated path is already "outdated" when transmitted back to the robot. Consequently, this path, planned based on the old timeline map, needs to be transformed to the current coordinate system for the robot to execute correctly.

[0050] The delay compensation unit in the local control module performs delay compensation based on the returned path planning results, generating a real-time tracking path aligned with the current physical environment. Specifically: During operation, the local control module reads motion data from the robot's odometry and inertial measurement unit at a fixed frequency, and stores the state vector containing timestamps, global position, heading angle, linear velocity, and angular velocity into a circular state buffer.

[0051] When the navigation path packet sent back by the edge computing server is received, the original collection timestamp carried in the data packet is parsed.

[0052] A binary search or linear interpolation is performed in the circular state buffer to retrieve the robot's historical pose state that is closest to the original timestamp.

[0053] The pose increment is calculated by integrating high-frequency data from the odometry and inertial measurement unit: ; In the formula , and These represent the corresponding coordinates and heading angle increments, respectively, where v(t) is the linear velocity. Angular velocity, This is the original data collection time. This is the current time.

[0054] The set of path points distributed from the edge is relative to the robot's coordinate system at the time of data acquisition. Therefore, it is necessary to use the aforementioned positional increments to transform them to the current robot coordinate system. ; in For the path points before compensation, This refers to the compensated target path point relative to the current robot position.

[0055] The local controller generates control commands based on the compensated path points, driving the robot to travel along the compensated route.

[0056] In a preferred embodiment, the local control module further includes a timeliness verification unit, which is used for: After parsing and obtaining the original acquisition timestamp carried in the navigation path packet, the lag time of the navigation path packet from the current time is calculated.

[0057] If the lag time exceeds a preset safety threshold (e.g., 200ms), the edge calculation result of the frame is determined to be invalid, the path data is discarded, and the local control logic is triggered to control the robot to perform deceleration or restricted local obstacle avoidance actions until path data that meets the requirements is received.

[0058] A method for AI edge computing host in industrial control scenarios includes the following steps: The mobile robot triggers the data acquisition module to collect environmental data at the local clock time T, and records the robot's global pose at that moment. The timestamp T is used as a unique identifier to label the environmental data and the global pose data respectively, and the two are stored together.

[0059] The environmental data and timestamp T are encapsulated and sent to the edge server via a wireless network.

[0060] After receiving the data packet, the edge server inputs the environmental data into a deep convolutional neural network to perform semantic segmentation or object detection, generating a local occupancy grid map containing the distribution of dynamic and static obstacles.

[0061] The edge server runs a search algorithm on the locally occupied grid map, starting from the robot's pose at data acquisition time T, to plan the optimal path point sequence to avoid obstacles.

[0062] The generated path point sequence is packaged together with the original timestamps into navigation control instructions and sent back to the mobile robot.

[0063] After receiving navigation control instructions from the edge server, the mobile robot calculates the time difference between the current time and the data acquisition time, and calculates the displacement vector and rotation angle based on the integration of local odometer and IMU data within the time interval.

[0064] Apply an inverse rigid body transformation to the received optimal path point sequence to generate a new path aligned with the robot's current pose.

[0065] The robot is driven by speed control commands generated based on the newly generated path.

[0066] In the description of this specification, references to terms such as "an embodiment," "example," "specific example," etc., indicate that a specific feature, structure, material, or characteristic described in connection with that embodiment or example is included in at least one embodiment or example of the invention. In this specification, illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0067] The preferred embodiments of the present invention disclosed above are merely illustrative of the invention. These preferred embodiments do not exhaustively describe all details, nor do they limit the invention to the specific implementations described. Clearly, many modifications and variations can be made based on the content of this specification. This specification selects and specifically describes these embodiments to better explain the principles and practical applications of the invention, thereby enabling those skilled in the art to better understand and utilize the invention. The invention is limited only by the claims and their full scope and equivalents.

Claims

1. An AI edge computing host system for industrial control scenarios, characterized in that, This includes mobile robot terminals and edge computing servers; The mobile robot terminal includes a data acquisition module, a local control module, and a clock synchronization module; The data acquisition module is used to collect environmental data and motion status data of the mobile robot; The local control module is used to perform latency compensation on the path planning results returned by the edge computing server and generate a real-time tracking path that is aligned with the current physical environment. The clock synchronization module is used to maintain time synchronization with the edge computing server using a precise time protocol; The edge computing server includes an inference module and a global path planning module, and communicates with the mobile robot terminal via a wireless network. The inference module receives environmental data from the mobile robot and uses a deep neural network to detect obstacles. The global path planning module performs global path planning based on the results generated by the inference module, and generates the optimal path point sequence.

2. The AI ​​edge computing host system for industrial control scenarios according to claim 1, characterized in that, The inference module receives environmental data from the mobile robot and uses a deep neural network for obstacle detection, specifically including: Receive the perception data packet at time t, parse out the environmental data frame, and perform denoising, normalization and size scaling on it to generate tensor data that meets the input requirements of the pre-trained model; The tensor data is input into the backbone network of a pre-trained deep neural network, which includes multiple convolutional layers or attention mechanism layers to extract low-level texture features and high-level semantic features of the environmental data layer by layer, generating multi-scale feature maps. The feature map is input into the segmentation head of the model, upsampled and fused, and the pixel semantic mask is output. By combining the sensor intrinsic and extrinsic parameter matrices, the aforementioned semantic mask is projected onto the robot's top-down coordinate system to construct a local semantic occupancy grid map at time t, where each grid cell stores the probability of the obstacle's presence and semantic category information.

3. The AI ​​edge computing host system for industrial control scenarios according to claim 2, characterized in that, The global path planning module performs global path planning based on the results generated by the inference module, generating an optimal path point sequence, specifically including: Based on the semantic category information in the grid map, different expansion radii and cost values ​​are assigned to different types of obstacles; for the semantic labels of dynamic attributes, a Gaussian-distributed prediction cost field is generated around the obstacle. Starting from the robot's position at time t and ending with the projection of the target point on the local map, a search algorithm is run on the cost map to generate an optimal path sequence that avoids high-cost regions. The optimal path point sequence is repackaged with the original timestamps to generate navigation control commands, which are then sent back to the mobile robot.

4. The AI ​​edge computing host system for industrial control scenarios according to claim 1, characterized in that, The process of performing latency compensation on the path planning results returned by the edge computing server to generate a real-time tracking path aligned with the current physical environment specifically includes: During operation, the local control module reads motion data from the robot's odometry and inertial measurement unit at a fixed frequency, and stores the state vector containing timestamps, global position, heading angle, linear velocity and angular velocity into a circular state buffer. When the navigation path packet sent back by the edge computing server is received, the original collection timestamp carried in the data packet is parsed. A binary search or linear interpolation is performed in the circular state buffer to retrieve the robot's historical pose state that is closest to the original timestamp. Obtain the robot's real-time pose state at the current moment, and calculate the robot's relative pose transformation matrix from the historical moment to the current moment based on the historical pose state and the real-time pose state. Based on the relative positional transformation matrix, each point in the path point sequence sent by the edge computing server is transformed into a target point relative to the current robot body coordinate system, thereby generating a real-time tracking path aligned with the current physical environment.

5. The AI ​​edge computing host system for industrial control scenarios according to claim 1, characterized in that, The local control module further includes a timeliness verification unit, which is used for: After parsing and obtaining the original acquisition timestamp carried in the navigation path packet, calculate the lag time of the navigation path packet from the current time. If the lag time exceeds the preset safety threshold, the edge calculation result of the frame is determined to be invalid, the path data is discarded, and the local control logic is triggered to control the robot to perform deceleration or restricted local obstacle avoidance actions until path data that meets the requirements is received.

Images