A data engine which gathers data and improves its performance while executing the task, and demonstrates the ability to learn and reduce the need for expensive verifications over time, while staying within the set error-rate.
Abstract
Flexible manufacturing requires rapid deployment of solutions and minimal setup time to remain competitive. An essential attribute is the ability to control error levels, as failures can range from minor performance degradation to severe equipment damage. However, conventional deployment often involves extensive setup, data collection, model training or parameter tuning, and system testing, resulting in significant delays that hinder commercial feasibility. We propose a data engine which gathers data and improves its performance while executing the task. The data engine consists of two classifiers, a fast model prediction and expensive verification. First, a model prediction is performed and based on the confidence level of the prediction, the expensive verification can be used. By adjusting the confidence level, users can control the level of tolerable error. Our method is implemented on a real-world robotic insertion task, which uses force data for the model prediction. The system applies UMAP dimensionality reduction and uses Wilson-Score to compute the confidence bounds of the prediction. Results demonstrate the ability to learn and reduce the need for expensive verifications over time, while staying within the set error-rate. The results highlight the potential of confidence bounds in self-improving models to enhance reliability in robotic classification task.
To maintain high efficiency and reduce operational downtime in industrial manufacturing, effective Predictive Maintenance (PdM) for robotic manipulators is essential. Although combining Model-Agnostic Meta-Learning (MAML) with digital twin technology offers a solid basis for quickly identifying faults, conventional methods often face challenges regarding parameter sensitivity and generalizing to new domains. To mitigate these issues, we introduce an ensemble-based metalearning framework that combines MAML with majority voting and operational grouping. This methodology improves generalization, stabilizes performance across diverse conditions, and strengthens few-shot learning capabilities. We validated the framework using a synthetic vibration dataset generated via a digital twin to simulate various robotic arm faults. Our findings demonstrate that this method achieves 93.8% accuracy and 93.1% precision in the ten-shot regime, outperforming the MAML baseline by 11.1%, across a broad range of mechanical faults, showing strength in cross-domain few-shot (CDFS) scenarios. Comparisons with established frameworks - including Reptile, Protonet, and ANIL, confirm the effectiveness of our model. By employing ensemble learning, we attain greater robustness and classification accuracy, establishing the method as a viable solution for industrial PdM. Furthermore, the integration of digital twins bridges the gap between simulation and real-world deployment, reducing data dependency and enabling effective fault classification even in dynamic environments with limited labeled data.
Mainak Mallick, Seung-Kyum Choi· 2026 6th International Confe...· 0 citations
Injection molding is a widely used process in the plastics industry, yet its quality depends heavily on precise adjustment of interdependent process parameters, which has traditionally relied on trial-and-error by skilled engineers. Such reliance is inefficient, time-consuming, and difficult to scale, particularly as the number of controllable parameters increases. This study investigates reinforcement learning (RL) as a means of supporting process optimization in injection molding by constructing an environmental model that reflects the severe class imbalance and the requirement for continuous control. The environmental model for RL-based agents was built using the Injection Molding AI Dataset, where weight balancing was applied to classifiers and anomaly detection methods were separately employed to improve reliability under imbalanced conditions. On this basis, two RL-based agents, Deep Q-Network (DQN) and Multi-Agent Deep Deterministic Policy Gradient (MADDPG), were designed and evaluated for their ability to optimize key process parameters such as injection time, filling time, and screw positions. The comparative analysis showed that while DQN offered advantages in terms of lightweight architecture, stable training behavior, and ease of implementation in discretized settings, MADDPG achieved greater effectiveness in continuous action spaces with multiple interdependent variables. Importantly, the MADDPG-based agent consistently identified non-defective production conditions in a single step, within the simulated environment, suggesting promising industrial applicability in future injection molding systems.
Daeyoung Kang, Keonwoo Nam, Seongrae Kim et al.· International Journal of Pre...· 0 citations
This paper presents a novel contamination-aware robotic manipulation framework for real-time IV bag quality assessment with adaptive manipulation policies. The system employs dual S0-ARM101 low-cost robot arms with eye-in-hand RGB vision for IV bag handling, coordinated through an open-source ROS2 architecture. Using a leader-follower architecture, the system learns visual-motor mappings from just 50 expert demonstrations, achieving 96% and 98% success rates for loading and unloading operations, respectively. The contamination-aware approach demonstrates 94% success to detected contaminated IV bags, and 90% adaptation to new scenarios, with zero safety incidents during testing. Statistical validation with 95% confidence intervals confirms the system’s reliability, while the integration of quality feedback into manipulation decision-making creates a responsive system that adapts to contamination detection results in real-time. This work demonstrates that learned visual-motor mappings with contamination awareness can significantly enhance pharmaceutical quality control while maintaining safety, establishing an economically viable framework for intelligent manufacturing automation.
The transition toward high-mix low-volume manufacturing demands flexibility in robotic manipulation. However, conventional vision systems remain a bottleneck, requiring extensive data collection and model retraining whenever a new object is introduced to the production line. To overcome this rigidity, we present xperception, a zero-shot 6D pose estimation technology that eliminates the need for object-specific fine-tuning and laborious data annotation. By directly utilizing typical CAD models and integrating the rich semantic features of foundation models (e.g. DINOv2, GeDi), xperception achieves millimeter-accurate 6D pose estimation. xperception showed robustness against severe occlusions in industrial tasks like bin picking and is engineered for deployment on industrial edge hardware, such as NVIDIA Jetson Thor. Validated at a TRL of 6, the core methodology behind xperception is based on the FreeZe algorithm, which won the international BOP Challenge 2024, paving the way for scalable, plug-and-play robotic automation in unstructured high-mix low-volume manufacturing industries.
M. Bortolon, Andrea Caraffa, Alice Fasoli et al.· 0 citations
The positional repeatability of industrial robots is a critical yet state-dependent performance metric, highly sensitive to thermal conditioning and mechanical loading. This study develops a data-driven framework for predicting repeatability of FANUC LR Mate 200iD (FANUC, Oshino-mura, Japan) and KUKA KR 6 R700 Sixx (KUKA AG, Augsburg, Germany) robots under varying operational conditions. ISO 9283-compliant experiments using a TriCal system (TRI-CAL Ltd., Montreal, QC, Canada) were conducted across three warm-up durations, three payload levels, and five poses. Ensemble artificial neural networks with 10 independently trained networks were developed for each robot. The FANUC model achieved R2 = 0.9922, RMSE = 0.004231 mm, and MAE = 0.002979 mm, while the KUKA model achieved R2 = 0.9926, RMSE = 0.002919 mm, and MAE = 0.002215 mm. Prediction interval coverage was 93.3% for FANUC and 100% for KUKA. Per-pose R2 ranged from 0.9588 to 0.9966 for KUKA. Response surface analysis identified thermal stabilization as the dominant factor affecting repeatability, with improvements of 86% for FANUC and 84% for KUKA after 4 h of warm-up. The KUKA robot demonstrated superior robustness and lower variability compared to the FANUC robot. The framework provides a practical tool for predicting repeatability, supporting process planning, uncertainty budgeting, and precision manufacturing optimization.
B. Louhichi, Mohamed Slamani, I. Bonev et al.· Machines· 0 citations
This study presents a data-driven approach to improve the absolute positional accuracy of a MABI MAX-100-2.25-P industrial robot. To quantify positioning errors, the robot was programmed to move through a dense grid of points with spacing of 25 mm in two planes above and across a working table. A training dataset was acquired by pulling commanded positions from the robot controller and pairing them with tool centre point values measured by a Leica AT960-LR laser tracker. The end effector orientation was fixed programmatically throughout the experiment.Four predictive models were trained on the dataset to learn the error patterns between the commanded and measured tool centre point positions. These models were Linear Regression, Random Forest, Gradient Boosting, and Artificial Neural Networks. Whilst training and validation metrics initially indicated that Random Forest and Gradient Boosting offered the lowest mean squared errors (as low as 0.0 mm2), real-world testing revealed significant discrepancies in model generalisation. The models were evaluated on three distinct test sets: hold-out grid points, interpolated cell midpoints, and randomly generated points. In the tests, Linear Regression delivered the most robust performance, achieving Euclidean error reductions of up to 85% across all sets. In contrast, the tree-based models generalised poorly when given unseen or interpolated locations, sometimes even amplifying positional errors by over 3000%. Artificial Neural Network models demonstrated moderate generalisation capability, performing better than tree-based models on non-grid data but not surpassing Linear Regression. In the most widely applicable test, random points, Linear Regression achieved an average Euclidean error of 0.12 mm which corresponds to a repeatability to accuracy ratio of 1:2.4, an improvement on the 1:3 ratio from the robot datasheet.
Zaid Alhgam, Flavin Lee John, Christopher Burkinshaw et al.· 2026 IEEE/ASME International...· 0 citations