AI-empowered intelligence in industrial robotics: technologies, challenges, and emerging trends

2.1. Dexterous manipulation: grasping in clutter

A long-standing challenge in industrial automation is “Bin Picking”. This involves robots grasping objects from containers filled with cluttered, randomly oriented items^[41,42]. Traditional industrial robots typically perform highly structured tasks from fixed fixtures. This is due to a lack of sophisticated perception capabilities to handle occlusion and pose uncertainty. Therefore, empowering robots with visual autonomy in unstructured scenarios is a core problem^[43]. Therefore, enabling robots to achieve visual autonomy in unstructured environments represents the core challenge in expanding their industrial applications.

Integrating AI-powered vision systems has become a primary driver for overcoming this limitation. It transforms industrial robots from pre-programmed automated equipment into adaptive executors. By equipping robots with Red-Green-Blue-Depth (RGB-D) sensors and deep learning (DL) models, they gain a fundamental ability. This ability is to identify objects and estimate their graspable positions in cluttered bins^[44,45]. This method has proven effective in various industrial sectors. In construction automation, Wu et al. used the You Only Look Once version 5 (YOLOv5) model^[29]. They addressed the high-precision localization and placement of embedded parts in precast concrete components. In high-throughput logistics sorting, Nguyen et al. employed You Only Look Once version 7 (YOLOv7) and combined it with generative adversarial networks (GANs) to synthesize training data, thereby enhancing the robot’s adaptability to novel items^[30].

To achieve more precise manipulation, AI models enable robots to evolve from simple object localization to full 3D spatial pose understanding. For scenarios with complex stacking or severe occlusion, Wei et al. constructed a multi-stage DL model that integrates a graspability estimation module, point feature aggregation, and six degrees of freedom (6-DoF) parameter prediction^[31]. This model enables the stable and efficient generation of grasp strategies for robotic manipulation. Besides precision, operational efficiency is also a critical requirement in industrial environments. Song et al. developed a hierarchical fusion network that enables robots to complete the entire cycle from perception to prediction with real-time performance, a critical factor in maintaining production throughput^[46]. This level of performance also extends to tasks requiring high-precision handling of minute components, such as the assembly of electronic components^[47].

Beyond optimizing single grasping actions, AI is also enabling robots to learn multi-stage manipulation strategies. Huang et al. proposed a cooperative push-and-grasp method that combines image masking with DRL^[32]. This method allows the robot to actively sort through disordered piles to isolate target objects, subsequently performing stable industrial grasping tasks.

2.2. Automated inspection: defect detection

Industrial robots are increasingly being employed as automated inspection platforms within QC processes^[48,49]. However, significant challenges remain, including reliably identifying diverse and often microscopic surface defects^[35], alongside addressing uncertainties arising from variations in products, lighting conditions, and background textures^[50]. Such issues often lead to missed defects or elevated false-alarm rates, thereby limiting the reliability of automated inspection.

Under supervised learning (SL), robots can be trained for precise defect localization. For instance, Wang et al. developed a lightweight 3D-convolutional neural networks (3DCNN) architecture enabling robots to accurately classify multiple internal welding defects^[33]. While achieving high classification accuracy, the model possesses significantly fewer parameters than traditional 3D models, rendering it suitable for deployment on edge devices. Similarly, in the surface inspection of composite materials, a hybrid model combining convolutional neural network (CNN) and support vector machine was used. This allowed robots to perform highly accurate online assessments of drilling quality^[48].

The application of SL in QC has expanded to more diverse industries. In the traditionally labor-intensive textile industry, Hassan et al. employed an enhanced deep CNN [You Only Look Once version 8 nano (YOLOv8n)] to enable robots to identify up to 13 distinct fabric defects in real time^[34]. This approach significantly enhanced detection performance while maintaining both real-time capability and operational safety. Similarly, within the automotive supply chain, Cheng et al. developed a system integrating 3D vision with DL, enabling robots to autonomously scan tire sidewalls^[51]. This achieves exceptionally precise detection of minute embossed text anomalies on multi-specification tires without requiring computer aided design (CAD) models.

However, SL relies heavily on large volumes of labeled data. Consequently, model-driven and unsupervised learning approaches are addressing the issue of insufficient samples in industrial settings. Zhou et al. developed the robotic camera (RoboCam) system^[35]. It integrates CAD model rendering, dual-domain pose error tuning, and visual comparison techniques, achieving the first high-precision robotic automatic detection of micro-mesh defects under unsupervised conditions. Moreover, lightweight and efficient AI models are crucial to meet real-time inspection demands on production lines. For dynamic quality sorting tasks, Lin et al. integrated a lightweight You Only Look Once version 4 (YOLOv4)-tiny model with a cascaded neural network onto a six-axis robot^[52]. This system enables robots to perform multi-stage visual tasks in real time, demonstrating the potential of computationally efficient models within complex industrial inspection workflows.

2.3. Robust perception: multimodal fusion

In many industrial environments, factors such as smoke, dust, or strong light severely affect the performance of visual systems^[36]. This makes single visual perception a critical point of failure. To address this challenge, researchers are employing multimodal fusion techniques. By intelligently combining data from diverse sensor types, these approaches enable robots to maintain reliable environmental awareness even under visually constrained conditions^[53].

For equipment condition monitoring in extremely harsh environments such as mining sites, Rocha et al. developed the robot operating system inspection (ROSI) inspection robot system integrating multiple sensors, including vision, acoustics, and thermal imaging^[36]. By employing algorithms such as CNN and random forests for multi-source data analysis, the system achieves reliable detection and assessment of various failure modes in critical equipment components^[36].

Furthermore, researchers are exploring the integration of visual data with other physical signals for process monitoring and control. For instance, in robotic grinding operations, the condition of abrasive belts constitutes a critical quality-determining parameter. Research by Surindra et al. demonstrates that robots can precisely monitor belt wear by fusing data from accelerometers and force sensors^[37]. Utilizing machine learning (ML) models such as decision trees and random forests, this approach achieves high-accuracy, low-latency multimodal wear recognition, providing valuable insights for intelligent industrial maintenance^[37].

Multimodal perception is also extensively applied to enhance the naturalness and efficiency of human-robot collaboration in industrial settings. Effective collaboration demands that robots comprehend a range of human commands and physical cues, evolving from basic emergency stop functions to a more nuanced understanding of operator intent^[54]. To this end, integrating multiple modalities such as speech and vision is considered a key technological pathway towards achieving this objective. For example, Mendez et al. developed a collaborative assembly system integrating three independent CNN models, each tasked with processing voice commands, tracking hand movements, or identifying components^[55]. This multimodal approach enables operators to guide the robot through mechanical part assembly using natural language and gestures, achieving voice-controlled, contactless interaction within the collaborative assembly process.

In collaborative tasks involving substantial physical contact, a core challenge lies in managing inherent physical uncertainty, which has driven research into the fusion of vision with force/torque sensing. Bartyzel proposed a reinforcement learning (RL)-based multimodal variational deep Markov decision process (DeepMDP) approach for precision insertion tasks^[56]. By learning dynamic representations through the fusion of visual and force/torque feedback, this method enables robots to perform precision assembly more robustly, enhancing the model’s generalization capabilities across varying background disturbances and object types.

2.4. In-process metrology: 3D modeling and measurement

The growing demand for in-situ, sub-millimeter precision measurement in precision manufacturing is driving the evolution of industrial robots from mere manipulator arms towards intelligent metrology platforms^[57,58] As conventional offline inspection methods prove inefficient and unable to meet modern production rhythms, researchers are leveraging AI technologies to address this challenge by enhancing sensor accuracy, enabling dynamic process modeling, and optimizing measurement strategies^[38].

To enhance the raw measurement accuracy of sensors, Yang et al. combined industrial 3D simulation, AI training, and multi-algorithm preprocessing to learn and correct sensor-specific error patterns^[27]. This approach improved the positioning accuracy of time-of-flight cameras in industrial mobile robot vision inspection, facilitated robotic repositioning, and enhanced the stability and consistency of 3D visual inspection. This external vision-based measurement principle was also applied to the robot’s own state assessment. Simoni et al. developed the double branch semi-perspective decoupled heatmaps (D-SPDH) system, which employs depth cameras and CNNs to perform high-frequency, high-precision 3D pose estimation of robotic arms without relying on internal encoders^[59]. This provides a crucial capability for safe monitoring in industrial human-robot collaboration.

In the field of robotic machining, Ni et al. developed a digital twin-based system where ML models predict final contour errors based on the numerical control (NC) code trajectory generated by robotic planning^[39]. This enables pre-compensation of the path, significantly reducing machining errors. Such predictive analysis complements real-time feedback monitoring during the process. Chew et al. proposed a real-time 4D reconstruction system that employs multi-camera fusion technology to achieve object tracking and high-precision modelling in dynamic environments, providing robust support for online quality monitoring^[38].

The role of industrial robots has expanded beyond passively executing measurement tasks to actively planning and optimizing measurement strategies themselves. Li et al. proposed the 3D scanning coverage prediction framework 3D scanning coverage prediction (3DSCP)-Net, which predicts the coverage of a structured light scanner on a workpiece prior to actual scanning, enabling efficient viewpoint planning^[57]. Furthermore, Roos-Hoefgeest et al. proposed an end-to-end surface scanning trajectory optimization method incorporating the proximal policy optimization (PPO) algorithm^[58]. This approach generates RL-driven transferable scanning paths based on CAD models for industrial contour measurement tasks, enabling high-quality surface acquisition.

2.5. Adaptive learning: generalization to novelty

Under the flexible manufacturing model of HMLV, traditional SL incurs high costs and time due to the frequent introduction of new products requiring extensive data collection, annotation, and model retraining^[60]. Consequently, enabling industrial robots to rapidly generalize to new objects and tasks with minimal data represents a critical research challenge for achieving agile manufacturing.

To mitigate reliance on real-world data, researchers have turned to synthetic data generation techniques. Robots can be trained on large, automatically annotated virtual datasets rather than solely depending on real-world images. Mangat et al. successfully trained a You Only Look Once version 3 (YOLOv3) model for a pick-and-place task using only synthetic images generated from 3D models^[60].

Beyond generating additional data, another research direction focuses on endowing models with inherent generalization capabilities. Gao et al. employed few-shot learning methods to achieve high-precision translational manipulation of novelly shaped objects under unknown friction conditions^[61]. In increasingly prevalent open and dynamic industrial settings, robots frequently need to recognize and handle novel objects or categories unseen during training. Within visual systems research for assembly tasks, Chen and Lai proposed a zero-shot learning (ZSL) framework for semantic segmentation^[40]. This enhances preprocessing efficiency for industrial robot operations in human-robot interaction scenarios, enabling efficient perception of unseen object categories within industrial human-robot interaction contexts.

3.1. Reactive motion planning

Deploying robots within human-centric collaborative environments presents primary safety and efficiency challenges in enabling secure navigation around dynamic and unpredictable obstacles^[67]. Conventional path planning algorithms are typically computationally intensive and operate offline, struggling to respond to sudden changes^[75]. DRL offers a viable solution by training an end-to-end policy that directly maps real-time sensor inputs to underlying control commands, thereby achieving responsive and fluid motion^[65].

Multiple studies have demonstrated the efficacy of DRL in endowing industrial robots with robust dynamic obstacle avoidance capabilities. Wu et al. developed a hybrid DRL model integrating the strengths of deep deterministic policy gradient (DDPG) and soft actor-critic (SAC) algorithms, enabling a robot to achieve high success rates in real-world tests involving moving obstacles^[65]. Beyond validating dynamic obstacle avoidance efficacy, the study further examined the quality and efficiency of DRL-generated paths. Lindner and Milecki proposed a DRL-based obstacle avoidance algorithm integrating DDPG and hindsight experience replay^[66]. This approach not only achieved sub-millimeter positioning accuracy but also generated paths shorter than traditional path planning and manipulation algorithms. Furthermore, it ensured robotic behavior appeared safe and predictable to human collaborators. To this end, DRL is being employed to learn strategies that actively minimize risk, rather than relying solely on hard-coded safety rules. Liu et al. introduced an “intrinsic reward” function into the DDPG algorithm, enabling dynamic adjustment of motion trajectories based on human arm positions^[67]. This allows industrial robots to simultaneously satisfy safety requirements and task efficiency, significantly enhancing the robustness and generalizability of the strategy.

3.2. Contact-rich task planning

The complexity of industrial robot decision-making is particularly evident in task and motion planning problems, especially within precision assembly tasks involving physical contact and multi-step sequences^[63].

Traditional programming approaches lack the flexibility to handle such uncertainties, while standard DRL often proves inefficient due to sparse reward issues^[76-79]. Consequently, recent research has focused on methods such as IL, hierarchical reinforcement learning (HRL), and policy transfer, aiming to enhance robotic autonomy and dexterity in assembly tasks^[41,71,80].

The foundation of complex assembly tasks lies in robotic solutions for precise positioning under high degree of freedom (DoF). Honelign et al. employed DDPG to train a seven-DoF robotic arm policy network, enabling autonomous planning and execution of end-effector movements from arbitrary initial states to target positions under continuous control^[68]. This demonstrated the capability of DRL to master complex high-dimensional control without explicit kinematic modelling. This learning-based approach has also extended to optimizing underlying inverse kinematics (IK) solvers. Yu and Tan combined PPO with Damping Least Squares to enhance the accuracy and efficiency of IK solving for multiple industrial robots^[81].

The complexity of tasks significantly increases when robots transition from free-space positioning to contact-rich physical interaction. For the classic “plug-in” assembly task, Ji et al. proposed a DRL approach integrated with a non-diagonal compliant controller, enabling the robot to learn how to continuously adjust its stiffness and modify its trajectory during insertion without switching between different control modes^[69]. To enhance the generalizability of these learned skills, Men et al. proposed the policy fusion transfer algorithm, which enables robots to transfer acquired assembly knowledge to novel tasks by fusing source and target task policies^[70]. This approach effectively enhances both generalization capability and safety in unfamiliar insertion scenarios.

Beyond task-level generalization, addressing complexity within individual tasks is equally crucial. Zhou and Lin employed the multiple goal-conditioned hierarchical learning framework to decompose complex operations into multiple subtasks^[71]. Each subtask is learned by a unified policy network, with exploration guided by intermediate objectives to enhance convergence speed and robustness. This approach effectively resolves the training instability inherent in standard RL under sparse rewards. Complementing RL, IL circumvents RL’s inefficient exploration by directly learning expert policies, offering a more data-efficient pathway for complex motion planning. The path planning and collision checking network framework proposed by Ghafarian Tamizi generates paths for disordered grasping by imitating expert planners, utilizing neural networks for rapid collision detection on each path segment^[41]. This approach significantly reduces the planning time required by traditional sampling algorithms.

3.3. High-level strategic planning

The decision-making capabilities of industrial robots are evolving from fundamental motion control towards higher-level task planning and strategic coordination, particularly within the framework of human-robot collaboration^[82,83]. A central challenge in this evolution is enabling robots to comprehend abstract, natural language instructions from human operators^[68,73]. Another is empowering robots to engage in dynamic, efficient task allocation within human-robot teams^[63,74].

The advent of large language models (LLMs) has opened new avenues for robots to directly interpret high-level human commands. The robot operating system generative pre-trained transformer (ROSGPT) framework developed by Koubaa et al. integrates generative pre-trained transformer (GPT)-4 with robot operating system 2 (ROS2), enabling efficient conversion from natural language to robotic control commands^[72]. Moving beyond simple instruction translation, Gupta et al. proposed a novel robotic action contextualization framework combining LLMs with dynamical system (DS) controllers^[73]. This allows robots to fine-tune action parameters and correct execution errors based on task contexts, thereby enhancing autonomy and adaptability. This capability has also extended to mobile platforms. Wang et al. employed GPT-3.5-turbo to generate Python code from natural language instructions, controlling automated guided vehicles to execute navigation tasks within an intelligent workshop^[84]. By integrating semantic annotation with the Pathfinder algorithm for path planning and execution, they achieved natural human-machine interaction and multimodal intelligence fusion.

However, effectively applying these general LLMs in industrial environments requires addressing their lack of domain knowledge. To this end, Li et al. constructed the inaugural industrial robot Wizard-of-Oz dialoguing dataset (IRWoZ) dialogue dataset, specifically tailored for industrial tasks such as assembly and material handling^[85,86]. This provides an indispensable foundation for fine-tuning and evaluating LLM capabilities within manufacturing-specific scenarios.

Beyond interpreting instructions, higher-level decision-making involves optimizing task allocation between robots and humans to maximize team efficiency and fluidity. Faroni et al. proposed the task and motion planning framework^[63]. This framework dynamically plans and assigns assembly subtasks to humans and robots by integrating timeline task modelling with multi-objective action search, effectively enhancing the execution efficiency and robustness of multi-agent systems. Furthermore, RL has been applied to this challenge. For instance, Hou et al. developed a DRL-based model for mobile collaborative robots in automotive assembly scenarios^[74]. This model decomposes tasks through a hierarchical task network and utilizes the revival double deep Q-network algorithm to optimize task allocation, achieving exceptionally high success rates and efficiency in complex final assembly tasks.

4.1. Dexterous manipulation: precision force control

A primary operational challenge for industrial robots lies in maintaining precise force control during contact-intensive tasks such as grinding, polishing, and precision component assembly^[92]. Conventional position-controlled robots cannot respond compliantly to surface variations or contact forces, posing risks of damaging both workpieces and themselves. Meanwhile, classical force control methods often struggle to model complex contact dynamics accurately. To address this challenge, researchers have begun employing DRL to directly learn end-to-end force control strategies. Ma et al. proposed a robot flexible assembly strategy system based on SAC, achieving high-precision, low-contact-force autonomous assembly of flexible components such as flexible printed circuits under simulated real-world physical constraints^[93].

In industrial practice, DRL is also employed to optimize mature classical controllers, enhancing their adaptive capabilities. For instance, addressing challenging assembly tasks such as inclined holes, Zhang et al. utilized the twin delayed DDPG algorithm to optimize the virtual damping parameters of a robotic admittance controller online^[94]. This approach improved mating assembly precision while reducing contact forces. To further enhance controller responsiveness in unknown or rapidly changing environments, Hu et al. combined Gaussian processes with DDPG for optimizing admittance control parameters^[95]. By introducing an initial strategy to accelerate the training process, they achieved rapid convergence and high robustness in force control for industrial robots operating within uncertain environments.

The foundation of force control applications relies upon an accurate understanding of the robot’s dynamic model. By substituting analytical models with learned neural networks, robots can achieve more precise torque prediction. Deng et al. introduced an energy-to-neural network, embedding the structural form of the dynamic equations within the network itself^[91]. This approach combines the interpretability of physical models with the flexibility of ML, enhancing the precision and robustness of robots in parameter identification and trajectory fitting.

Furthermore, research by Shan and Pham demonstrated that neural networks can learn to estimate external forces and torques directly from motor currents and joint states within the robot^[90]. This breakthrough enables high-precision force control and sensitive manual guidance without any external force/torque sensors, significantly enhancing the autonomy and cost-effectiveness of industrial robots in precision manipulation scenarios.

4.2. Safe collaboration: risk management and fluent interaction

In human-robot collaboration, the challenges facing a robot’s executive intelligence extend beyond physical collision avoidance in dynamic path planning^[89]. A safe and effective collaborative system not only requires robots to avoid physical contact but must also ensure their behavior appears fluid and predictable to human collaborators, as hesitant or unstable movements undermine human-robot trust and diminish team efficiency^[54]. Consequently, recent research focuses on developing robotic execution systems capable of actively assessing risks and autonomously learning safe interaction strategies from data^[96-98].

In safe human-robot physical interactions, industrial robots must simultaneously address external interaction risks stemming from human behavior and internal errors within their own perception systems. To mitigate external risks, Zheyuan et al. proposed the advanced human-robot collaborative model for industrial collaborative settings^[94]. This model predicts interaction risks based on human behavior and semantic scene graphs, adjusting action strategies to optimize team task completion. Furthermore, ensuring safe execution necessitates accounting for uncertainties within the robot’s own perception systems. Xin et al. proposed a runtime compositional model verification framework^[97]. This framework calculates trust metrics based on sensor data, updates state transition probabilities to predict system failure risks, thereby enabling runtime model verification and dynamic safety assurance.

To enable robots to autonomously learn safety policies from data, researchers are turning to safety-oriented RL (SRL). However, a core challenge for SRL in industrial applications lies in handling the common presence of outliers in data and the distribution shift between deployment environments and training data. Addressing this issue, Hickman et al. developed a SRL designed to learn policies that maximize task rewards while satisfying safety constraints^[98]. By employing a student-t process instead of the standard Gaussian process, this framework better accommodates data anomalies during safety policy learning, thereby achieving a superior balance between task performance and safety constraints.

4.3. Robust deployment: sim-to-real transfer

Numerous studies have actively introduced uncertainty during the training phase to compel AI strategies to learn generalization capabilities insensitive to differences between virtual and real environments. A representative technique is domain randomization, which deliberately alters the physical and visual parameters of simulations throughout training. For instance, Amaya and Von Arnim randomized dynamic parameters such as friction within the simulation environment when training a spiking neural network controller for precision hole-punching tasks^[100]. Strategies trained in this manner could be directly deployed onto physical robots controlled by Loihi neuromorphic chips without any real-world fine-tuning, while maintaining high success rates.

Beyond randomizing environments, unifying simulation and reality through “scene representation” proves equally effective. Zhang et al. proposed an end-to-end approach that avoids direct use of raw images, instead converting them into a unified 3D bounding box representation insensitive to object shape and appearance^[103]. This significantly enhanced the transfer success rate of obstacle avoidance strategies. Furthermore, robustness can be bolstered by refining the learning algorithms themselves. Considering sensor noise as a primary cause of transfer failure, Wang et al. proposed the Density estimation-SAC algorithm^[106]. This technique integrates density estimation into the DRL framework, enabling explicit modelling and adaptation to noisy measurements during training.

Complementary domain adaptation techniques explicitly aim to minimize feature-level discrepancies between simulated and real-world data. For instance, when investigating multimodal grasping stability assessment, Zhang et al. encountered the challenge of stark stylistic differences between simulated and real tactile images^[104]. They employed a conditional GAN specifically trained to convert real tactile images into the simulated style. Following this processing, an evaluation network trained entirely on simulated data could directly process real tactile data from the physical world, achieving successful zero-shot transfer.

At the system level, high-fidelity digital twins are emerging as powerful system-level tools to bridge this gap. Rather than directly transferring strategies from general-purpose simulators to reality, an alternative approach involves utilizing digital twins as an intermediary layer. Liu et al. developed a transfer learning framework wherein grasping strategies pre-trained in simulation are first fine-tuned within a digital twin highly synchronized with the physical robot^[105]. This enables final optimization of the strategy within an environment closer to reality, significantly enhancing its ultimate grasping success rate on the actual robot.

5.1. Learning and generalization

The generalization capability and data efficiency of existing AI models constitute the primary constraints limiting their adaptation to HMLV production patterns. Whilst paradigms such as meta-learning offer avenues for rapid task adaptation^[107], strategy formulation under conditions of extremely sparse interactive data remains a core challenge in industrial settings. Future research must develop more robust self-SL frameworks to leverage vast amounts of unlabeled data in industrial environments for pre-training general-purpose world models. Moreover, as models grow increasingly large, efficient knowledge transfer from cloud-based large models to resource-constrained edge controllers emerges as a critical issue. Strategy distillation stands as one of the promising techniques in this field; however, establishing a mature theoretical framework remains necessary to address the issue of insufficient behavioral diversity and generalization in teacher models when strategy optimization is inadequate^[108].

5.2. Robustness and reliability

The long-term robustness and self-maintenance capabilities of systems are pivotal for transitioning technology from laboratory to factory settings. Current research predominantly focuses on immediate success at the task level, with less attention given to performance degradation in robots during prolonged continuous operation due to mechanical wear or gradual environmental changes^[109,110]. A significant future direction lies in developing robotic systems capable of proactive fault prediction, diagnosis, and autonomous recovery. Whilst initial explorations exist in online fault diagnosis, translating concepts such as reversible execution into robust, generalizable error recovery strategies feasible within the physical world holds substantial value for minimizing unplanned downtime on industrial production lines^[111].

5.3. Cognition and interaction

Generative AI is providing novel technical pathways for long-standing challenges in robotics, such as high-level task planning and sparse perception data. At the perception level, researchers are employing models such as conditional GANs to address the simulation-to-reality domain gap. Examples include enhancing the transfer performance of perception models by learning style transfer between simulated and real tactile images^[104], or directly generating kinematic solutions for robotic motions^[112]. At the decision-making level, research centered on LLMs has begun exploring the autonomous decomposition of natural language instructions into robot-executable task sequences^[72,73]. However, applying these technologies to industrial manufacturing, where safety and reliability demands are exceptionally high, still presents fundamental challenges. For LLMs, core issues requiring resolution in task planning include inference latency, lack of logical consistency, and insufficient understanding of physical world constraints. Regarding data generation, ensuring the fidelity and diversity of synthetic data to cover extreme conditions and long-tail problems in real industrial environments remains a key research topic requiring in-depth investigation. Consequently, future research will focus on establishing effective mechanisms to validate and constrain the outputs of generative models. It will also explore how generative models can be leveraged at scale to synthesize high-quality, diverse multimodal training data, thereby reducing the dependence of industrial AI applications on costly manual data collection and annotation.