Augmented Reality (AR)-based operational guidance has been increasingly adopted in manual assembly scenarios due to the advantages in improving operator efficiency and reducing cognitive load. Since users become increasingly reliant on AR interaction, trust management in AR guidance has become a critical determinant of collaborative safety and overall task performance. Over-trust may lead to blind reliance on AR-generated instructions, whereas under-trust can cause hesitation and inefficient coordination. To address these challenges, this study investigates trust variations within AR-assisted assembly tasks and examines how different instruction designs influence trust. Three AR guidance modes were evaluated, including pure instruction mode, instruction with reliability mode, and instruction with explanation mode. Data collected across different modes revealed significant variation in user trust, with pure instruction mode leading to the highest behavioral reliance but lowest subjective trust, while explanation mode encouraged the most balanced user behavior and highest subjective trust scores. Trust calibration error captures deviations between user trust and system capability, which provides empirical support for adaptive trust calibration strategies in AR-based collaboration and offers practical insights for designing information interaction systems that maintain appropriate trust levels, thereby ensuring safety and efficiency.
With the increasing demand for mass customization, assembly scheduling problems have become increasingly complex and fraught with uncertainty. Traditional scheduling methods often struggle to balance real-time responsiveness with solution quality when facing dynamic disturbances such as machine breakdowns and urgent orders. Furthermore, optimizing multiple conflicting objectives simultaneously, such as makespan and energy consumption, remains a significant challenge. To address these issues, this paper proposes a multi-objective assembly scheduling optimization framework based on Deep Reinforcement Learning. First, the dynamic assembly scheduling problem is modeled as a Markov Decision Process, fully considering assembly precedence and resource constraints. Second, an improved multi-objective reward function is designed to guide the agent in learning Pareto-optimal strategies, balancing production efficiency and operational costs. Finally, experiments conducted in a simulated flexible assembly environment demonstrate that the proposed method achieves millisecond-level real-time responsiveness in dynamic settings and outperforms traditional genetic algorithms and heuristic rules in comprehensive multi-objective metrics.
Aiming at the pain points such as inefficient path planning of coordinate measuring machines (CMMs) when dealing with revolving bodies in the engine production and inspection process, this paper proposes an automatic measurement path planning method based on OCCT to achieve the goal of assembly process optimization. Firstly, the method parses IGES format CAD files, extracts the geometric and topological information of parts, and converts it into Brep representation to realize accurate redrawing and feature recognition of the model. Subsequently, it adaptively plans the number and distribution mode of measurement points according to the type of assembly surface (cylindrical surface, toroidal surface, etc.), and generates the initial measurement path based on the principles of collision-free and shortest path. It identifies interference risks through scene-specific collision detection (probe moving/workpiece stationary, probe stationary/workpiece rotating, synchronous movement of both) and optimizes the path using an obstacle avoidance point insertion strategy. Finally, it outputs MOV and MVC standardized motion commands to directly guide the execution of measurement equipment.
Proactive Human-Robot Collaboration (HRC) in flexible manufacturing relies heavily on the robot’s ability to accurately interpret human actions. However, traditional supervised action recognition methods face significant limitations in this setting, as they demand extensive, task-specific annotated datasets and generalize poorly to unseen procedures. While Large Vision-Language Models (VLMs) offer a promising zero-shot alternative, their generalist priors often lead to ungrounded reasoning that overlooks fine-grained, task-critical human-object interactions. To address this, we introduce Scene-Grounded VLM (SG-VLM), a novel framework that systematically anchors VLM inference within an explicit model of the physical scene. Specifically, we leverage open-vocabulary object detection to construct a Structured Scene Representation (SSR)—a dynamic, textual abstraction of task-relevant tools and parts across the assembly timeline. This SSR is then integrated into a context-augmented prompt that constrains the VLM to analyze specific interactions within the workspace, effectively coupling language-based reasoning with physical scene structure. We evaluate the approach on a complex engine assembly task. Experimental results demonstrate that SG-VLM significantly outperforms unguided counterparts in zero-shot recognition accuracy. This work establishes the viability of VLMs for industrial domains and highlights structured scene grounding as a critical mechanism for achieving the robustness required in practical assembly applications.
Precision positioning measurement technology is widely used in the field of micro-assembly. Although pixel interpolation can algorithmically improve image resolution, phase variations in images caused by factors such as illumination conditions and imaging noise can interfere with positioning measurements. Therefore, in research on the resolution of positioning measurement based on microscopic vision, close attention must be paid to the aforementioned factors. In this paper, based on three-dimensional and two-dimensional microscopic images of ultra-precision diamond-turned surfaces, the correlation between surface topography and its corresponding images is analysed. Furthermore, a study on nanoscale positioning measurement resolution is conducted by incorporating image interpolation and registration methods. Experimental results demonstrate that this approach can effectively resolve nanoscale changes in the position of microstructures. It is also found that phase shifts in images induced by variations in illumination affect measurement resolution, thereby introducing positioning measurement errors. Additionally, the study indicates that registered images have minimal impact on measurement resolution.
