R. Sabzevari
info
Please Note
<p>This page displays the records of the person named above and is not linked to a unique person identifier. This record may need to be merged to a profile.</p>
10 records found
1
A Framework for the Analysis, Transformation, and Validation of Rule-Based Parcel Sorting Plans
Towards improved maintainability, interpretability, and behavioural assurance
Rule-based parcel sorting plans are difficult to maintain because their effective decision logic is distributed across configuration objects, executable rule representations, routing tables, and run-specific context. This paper presents an evidence-grounded pipeline for analysing, conservatively transforming, behaviourally validating, and explaining recurring maintainability faults in such plans. The pipeline reconstructs sorting behaviour from abstract syntax trees, routing tables, lookup data, and run-level metadata, and combines canonical structural comparison, SMT-supported satisfiability and implication checks, routing-aware scope analysis, staged transformation in a copied working representation, synthetic behavioural validation against the original plan, and a local retrieval-supported explanation layer over the resulting evidence. The evaluation covers nine selected industrial configuration sets that expand to fourteen run-level evaluations. A controlled synthetic setup with 100 injected primary faults confirmed coherent end-to-end behaviour, with 21 out of 21 pipeline checks and 4 out of 4 validation checks passing. In the industrial case, Duplicate Sorting Class, Unsatisfiable Rule, Overshadowed Rule, and Shared Routing Overlap findings were reduced to zero in the copied working representation, while routing rows were reduced from 1,206,652 to 618,131. Behavioural validation separated quiet cleanup from concentrated review-sensitive differences, while the explanation layer helped reviewers inspect the evidence without becoming the decision authority. The behavioural claims remain bounded to the generated workload and do not establish full equivalence for all possible parcels or deployment conditions.
...
Rule-based parcel sorting plans are difficult to maintain because their effective decision logic is distributed across configuration objects, executable rule representations, routing tables, and run-specific context. This paper presents an evidence-grounded pipeline for analysing, conservatively transforming, behaviourally validating, and explaining recurring maintainability faults in such plans. The pipeline reconstructs sorting behaviour from abstract syntax trees, routing tables, lookup data, and run-level metadata, and combines canonical structural comparison, SMT-supported satisfiability and implication checks, routing-aware scope analysis, staged transformation in a copied working representation, synthetic behavioural validation against the original plan, and a local retrieval-supported explanation layer over the resulting evidence. The evaluation covers nine selected industrial configuration sets that expand to fourteen run-level evaluations. A controlled synthetic setup with 100 injected primary faults confirmed coherent end-to-end behaviour, with 21 out of 21 pipeline checks and 4 out of 4 validation checks passing. In the industrial case, Duplicate Sorting Class, Unsatisfiable Rule, Overshadowed Rule, and Shared Routing Overlap findings were reduced to zero in the copied working representation, while routing rows were reduced from 1,206,652 to 618,131. Behavioural validation separated quiet cleanup from concentrated review-sensitive differences, while the explanation layer helped reviewers inspect the evidence without becoming the decision authority. The behavioural claims remain bounded to the generated workload and do not establish full equivalence for all possible parcels or deployment conditions.
Object-level maps represent objects in an environment as a set of persistent instances; they support questions such as which objects are present, what they are, where they are, and how large they are. This thesis presents an online system that builds such a map for indoor mobile robots from RGB-D images and 2D LiDAR scans under imperfect detection, localization uncertainty, and changing object presence. Each object is represented by recursive beliefs over pose, physical extent, semantic class, and existence. Measurement uncertainty is calibrated empirically per class and expressed in a direction dependent, sensor-aligned frame. Observations are associated one-to-one by Mahalanobis distance for robustness to localization and measurement noise, and a viewpoint-novelty gate restricts geometric updates to sufficiently unique viewpoints, reducing overconfidence caused by treating repeated same-viewpoint observations as independent. Object existence is updated as an evidence-driven log-odds belief: detections provide positive evidence, while LiDAR free-space ray tracing provides negative evidence only when the predicted object location is confirmed empty. Evaluated on held-out recordings against motion-capture ground truth, the system maintains the correct object set under increasing synthetic Gaussian localization noise while achieving centimeter-level geometric accuracy and approximately consistent position covariances in the tested indoor setting.
...
Object-level maps represent objects in an environment as a set of persistent instances; they support questions such as which objects are present, what they are, where they are, and how large they are. This thesis presents an online system that builds such a map for indoor mobile robots from RGB-D images and 2D LiDAR scans under imperfect detection, localization uncertainty, and changing object presence. Each object is represented by recursive beliefs over pose, physical extent, semantic class, and existence. Measurement uncertainty is calibrated empirically per class and expressed in a direction dependent, sensor-aligned frame. Observations are associated one-to-one by Mahalanobis distance for robustness to localization and measurement noise, and a viewpoint-novelty gate restricts geometric updates to sufficiently unique viewpoints, reducing overconfidence caused by treating repeated same-viewpoint observations as independent. Object existence is updated as an evidence-driven log-odds belief: detections provide positive evidence, while LiDAR free-space ray tracing provides negative evidence only when the predicted object location is confirmed empty. Evaluated on held-out recordings against motion-capture ground truth, the system maintains the correct object set under increasing synthetic Gaussian localization noise while achieving centimeter-level geometric accuracy and approximately consistent position covariances in the tested indoor setting.
Language-conditioned local navigation requires a robot to infer a nearby traversable target location from its current observation and an open-vocabulary, relational instruction. Existing vision-language spatial grounding methods usually rely on vision–language models (VLMs) to reason in image space, producing 2D predictions tied to visible pixels. As a result, they struggle to infer target locations in occluded regions, typically caused by furniture or moving humans. To address this issue, we propose BEACON, which predicts an ego-centric Bird’s-Eye View (BEV) affordance heatmap over a bounded local region including occluded areas. Given an instruction and surround-view RGB-D observations from four directions around the robot, BEACON predicts the BEV heatmap by injecting spatial cues into a VLM and fusing the VLM’s output with depth-derived BEV features. Using an occlusion-aware dataset built in the Habitat simulator, we conduct detailed experimental analysis to validate both our BEV space formulation and the design choices of each module. Our method improves the accuracy averaged across geodesic thresholds by 22.74 percentage points over the state-of-the-art image-space baseline on the validation subset with occluded target locations.
...
Language-conditioned local navigation requires a robot to infer a nearby traversable target location from its current observation and an open-vocabulary, relational instruction. Existing vision-language spatial grounding methods usually rely on vision–language models (VLMs) to reason in image space, producing 2D predictions tied to visible pixels. As a result, they struggle to infer target locations in occluded regions, typically caused by furniture or moving humans. To address this issue, we propose BEACON, which predicts an ego-centric Bird’s-Eye View (BEV) affordance heatmap over a bounded local region including occluded areas. Given an instruction and surround-view RGB-D observations from four directions around the robot, BEACON predicts the BEV heatmap by injecting spatial cues into a VLM and fusing the VLM’s output with depth-derived BEV features. Using an occlusion-aware dataset built in the Habitat simulator, we conduct detailed experimental analysis to validate both our BEV space formulation and the design choices of each module. Our method improves the accuracy averaged across geodesic thresholds by 22.74 percentage points over the state-of-the-art image-space baseline on the validation subset with occluded target locations.
Reproducing Matrix Capsule Networks with EM Routing
A Modular and Faithful TensorFlow 2 Implementation
Matrix Capsule Networks with EM Routing, introduced by Hinton et al. (2018a), offer a powerful way to model part-whole relationships and pose information in neural networks. However, their broader adoption has been limited, possibly due in part to the complexity of the model and inconsistencies between the original paper and its official implementation. In this thesis, we present a faithful and modular reimplementation of the Matrix Capsule Network in TensorFlow 2, addressing missing and unclear aspects of the original description. Our implementation achieves competitive results on the SmallNORB dataset and is designed to be accessible, adaptable, and well-documented. By clarifying key architectural details and providing open-source code, we aim to support further research and lower the barrier to working with capsule-based models.
...
Matrix Capsule Networks with EM Routing, introduced by Hinton et al. (2018a), offer a powerful way to model part-whole relationships and pose information in neural networks. However, their broader adoption has been limited, possibly due in part to the complexity of the model and inconsistencies between the original paper and its official implementation. In this thesis, we present a faithful and modular reimplementation of the Matrix Capsule Network in TensorFlow 2, addressing missing and unclear aspects of the original description. Our implementation achieves competitive results on the SmallNORB dataset and is designed to be accessible, adaptable, and well-documented. By clarifying key architectural details and providing open-source code, we aim to support further research and lower the barrier to working with capsule-based models.
This paper presents a method which is capable of creating an object centered world description based upon consecutive measurements about an unknown number of static objects using Bayesian inference. The objects are represented by a two dimensional position, with the aim of adding more attributes in future works. This objective is reformulated into a clustering problem which is then solved using a structure learning method. It is implemented using RxInfer which uses the message passing algorithm in combination with factor graphs to perform Bayesian inference. The results indicate a promising performance of the structure learning model, but also show signs that the object representation has been over simplified. The future works section provides guidance on how the model complexity can be increased by adding additional attributes in order to improve performance.
...
This paper presents a method which is capable of creating an object centered world description based upon consecutive measurements about an unknown number of static objects using Bayesian inference. The objects are represented by a two dimensional position, with the aim of adding more attributes in future works. This objective is reformulated into a clustering problem which is then solved using a structure learning method. It is implemented using RxInfer which uses the message passing algorithm in combination with factor graphs to perform Bayesian inference. The results indicate a promising performance of the structure learning model, but also show signs that the object representation has been over simplified. The future works section provides guidance on how the model complexity can be increased by adding additional attributes in order to improve performance.
To ensure the safety of a spacecraft, operators collect thousands of telemetry signals and monitor them for anomalies, which is both expensive and time-consuming. Space agencies have been researching Machine Learning (ML)-based Time Series Anomaly Detection (TSAD) methods to improve automation, but this is hindered by a lack of high-quality benchmarks. This thesis explores the application of ML-based TSAD on instrument telemetry data for the XMM-Newton space telescope. The data was explored to find several unique challenges, such as recurring eclipses and a high volume of missing data. A methodology was then developed to pre-process the raw data into an ML-compatible format, detect anomalies using a semi-supervised forecasting approach and post-process the detections into a benchmark-suitable format. The method yielded over 40 detections, which were partially validated in discussion with instrument engineers. A refined methodology, incorporating areas for improvement, is presented to proceed with an eventual XMM-Newton anomaly benchmark.
...
To ensure the safety of a spacecraft, operators collect thousands of telemetry signals and monitor them for anomalies, which is both expensive and time-consuming. Space agencies have been researching Machine Learning (ML)-based Time Series Anomaly Detection (TSAD) methods to improve automation, but this is hindered by a lack of high-quality benchmarks. This thesis explores the application of ML-based TSAD on instrument telemetry data for the XMM-Newton space telescope. The data was explored to find several unique challenges, such as recurring eclipses and a high volume of missing data. A methodology was then developed to pre-process the raw data into an ML-compatible format, detect anomalies using a semi-supervised forecasting approach and post-process the detections into a benchmark-suitable format. The method yielded over 40 detections, which were partially validated in discussion with instrument engineers. A refined methodology, incorporating areas for improvement, is presented to proceed with an eventual XMM-Newton anomaly benchmark.
Open-world object manipulation has emerged as a popular research frontier in robotics. While recent advances in vision-language-action (VLA) models have achieved impressive results, they typically rely on large amounts of task-specific action data for training. This thesis aims to enable a manipulator to perform open-world object manipulation tasks without any action demonstrations. Instead of learning direct action mappings, we focus on understanding object dynamics.
To this end, we propose a novel framework that builds an explicit world model for open-world object manipulation. The framework integrates open-set segmentation and grasping, 3D digital twin reconstruction, and simulation-based strategy sampling within a unified framework. At the core of our approach lies the construction of a physically grounded digital twin of the environment, which enables the framework to simulate and evaluate diverse interaction strategies before real-world execution.
Experimentally, the proposed framework is able to perform multiple open-set manipulation tasks, such as “put the banana into the basket”, “stack the green cube onto the yellow cube”, and “place the blue cup upside on the wooden box”. These results are obtained without any task-specific action demonstrations, demonstrating strong generalization and autonomy compared to existing closed-set or imitation-based systems. ...
To this end, we propose a novel framework that builds an explicit world model for open-world object manipulation. The framework integrates open-set segmentation and grasping, 3D digital twin reconstruction, and simulation-based strategy sampling within a unified framework. At the core of our approach lies the construction of a physically grounded digital twin of the environment, which enables the framework to simulate and evaluate diverse interaction strategies before real-world execution.
Experimentally, the proposed framework is able to perform multiple open-set manipulation tasks, such as “put the banana into the basket”, “stack the green cube onto the yellow cube”, and “place the blue cup upside on the wooden box”. These results are obtained without any task-specific action demonstrations, demonstrating strong generalization and autonomy compared to existing closed-set or imitation-based systems. ...
Open-world object manipulation has emerged as a popular research frontier in robotics. While recent advances in vision-language-action (VLA) models have achieved impressive results, they typically rely on large amounts of task-specific action data for training. This thesis aims to enable a manipulator to perform open-world object manipulation tasks without any action demonstrations. Instead of learning direct action mappings, we focus on understanding object dynamics.
To this end, we propose a novel framework that builds an explicit world model for open-world object manipulation. The framework integrates open-set segmentation and grasping, 3D digital twin reconstruction, and simulation-based strategy sampling within a unified framework. At the core of our approach lies the construction of a physically grounded digital twin of the environment, which enables the framework to simulate and evaluate diverse interaction strategies before real-world execution.
Experimentally, the proposed framework is able to perform multiple open-set manipulation tasks, such as “put the banana into the basket”, “stack the green cube onto the yellow cube”, and “place the blue cup upside on the wooden box”. These results are obtained without any task-specific action demonstrations, demonstrating strong generalization and autonomy compared to existing closed-set or imitation-based systems.
To this end, we propose a novel framework that builds an explicit world model for open-world object manipulation. The framework integrates open-set segmentation and grasping, 3D digital twin reconstruction, and simulation-based strategy sampling within a unified framework. At the core of our approach lies the construction of a physically grounded digital twin of the environment, which enables the framework to simulate and evaluate diverse interaction strategies before real-world execution.
Experimentally, the proposed framework is able to perform multiple open-set manipulation tasks, such as “put the banana into the basket”, “stack the green cube onto the yellow cube”, and “place the blue cup upside on the wooden box”. These results are obtained without any task-specific action demonstrations, demonstrating strong generalization and autonomy compared to existing closed-set or imitation-based systems.
Humans are our best example of the ability to learn a structure of the world through observation of environmental regularities. Specifically, humans can learn about different objects, different classes of objects, and different class-specific behaviors. Fundamental to these human abilities is evolved sensory hardware and automatic pattern recognition systems thought to be powered in part by the leading neuroscience theory of predictive coding. Artificial intelligence research is often inspired by neuroscience and algorithms already exist that implement predictive coding. In this paper, we seek to evaluate a leading predictive coding video-prediction algorithm, PredNet, for its ability to perform human-like learning of the types mentioned. By successfully training PredNet on a custom Simple Shape Motion (SSM) video dataset that explicitly requires structure learning to occur in order to accurately predict the next frame, we establish that PredNet is capable of rudimentary structure learning. We investigate PredNet filters and feature maps but find scant evidence of truly symbolic knowledge, and propose instead that PredNet performs semi-symbolic learning. We perform ablation studies that reveal the aspects of PredNet that critically contribute to its structure learning ability. Finally, we detail a set of modifications made to PredNet to allow object-centric processing as a promising step change towards human-like structure learning. Evaluation results and investigations are provided. Performance was slightly worse than Baseline, likely due to a noted implementation flaw. Code and instructions to reproduce dataset creation and model training / evaluation are available at https://github.com/ofSingularMind/parallel_prednet.
...
Humans are our best example of the ability to learn a structure of the world through observation of environmental regularities. Specifically, humans can learn about different objects, different classes of objects, and different class-specific behaviors. Fundamental to these human abilities is evolved sensory hardware and automatic pattern recognition systems thought to be powered in part by the leading neuroscience theory of predictive coding. Artificial intelligence research is often inspired by neuroscience and algorithms already exist that implement predictive coding. In this paper, we seek to evaluate a leading predictive coding video-prediction algorithm, PredNet, for its ability to perform human-like learning of the types mentioned. By successfully training PredNet on a custom Simple Shape Motion (SSM) video dataset that explicitly requires structure learning to occur in order to accurately predict the next frame, we establish that PredNet is capable of rudimentary structure learning. We investigate PredNet filters and feature maps but find scant evidence of truly symbolic knowledge, and propose instead that PredNet performs semi-symbolic learning. We perform ablation studies that reveal the aspects of PredNet that critically contribute to its structure learning ability. Finally, we detail a set of modifications made to PredNet to allow object-centric processing as a promising step change towards human-like structure learning. Evaluation results and investigations are provided. Performance was slightly worse than Baseline, likely due to a noted implementation flaw. Code and instructions to reproduce dataset creation and model training / evaluation are available at https://github.com/ofSingularMind/parallel_prednet.
This thesis presents an approach to monocular depth estimation for Unmanned Aerial Vehicles (UAVs). Monocular depth estimation is a critical perception task for UAVs, enabling them to infer depth information from visual data without relying on heavy or power-consuming sensors such as LiDAR or stereo cameras. Given the operational constraints of UAVs, such as limited payload and energy resources, robust and efficient depth estimation methods are required to facilitate safe navigation and environmental interaction. The proposed methodology in this thesis integrates visual data from a monocular camera with inertial measurements from an Inertial Measurement Unit (IMU) sensor. This combination aims to address challenges such as scale ambiguity in the depth estimates and inaccuracies in dynamic environments that are common in aerial operations. The integration of IMU data with a differentiable camera-centric Extended Kalman Filter (EKF) allows for better ego-motion estimation, effectively calibrating the visual information with drone dynamics. The method further incorporates depth map frame prediction, leveraging initial depth estimates along with temporal dynamics to predict future depth maps. This predictive capability improves efficiency by reducing the need for full depth estimation in every frame, allowing robotic agents to anticipate environmental changes. The evaluation on simulated and real-world datasets shows that while the algorithm performs well over short forecast horizons, accumulating errors from IMU data and the assumption of a static environment limit its long-term accuracy. The future depth map prediction algorithm reduced the need for DynaDepth from 10 runs per second to 2, and on the Mid-Air dataset, from 25 to 5. Additionally, this study provides a foundation for future work, including the integration of an object-oriented frame prediction algorithm.
...
This thesis presents an approach to monocular depth estimation for Unmanned Aerial Vehicles (UAVs). Monocular depth estimation is a critical perception task for UAVs, enabling them to infer depth information from visual data without relying on heavy or power-consuming sensors such as LiDAR or stereo cameras. Given the operational constraints of UAVs, such as limited payload and energy resources, robust and efficient depth estimation methods are required to facilitate safe navigation and environmental interaction. The proposed methodology in this thesis integrates visual data from a monocular camera with inertial measurements from an Inertial Measurement Unit (IMU) sensor. This combination aims to address challenges such as scale ambiguity in the depth estimates and inaccuracies in dynamic environments that are common in aerial operations. The integration of IMU data with a differentiable camera-centric Extended Kalman Filter (EKF) allows for better ego-motion estimation, effectively calibrating the visual information with drone dynamics. The method further incorporates depth map frame prediction, leveraging initial depth estimates along with temporal dynamics to predict future depth maps. This predictive capability improves efficiency by reducing the need for full depth estimation in every frame, allowing robotic agents to anticipate environmental changes. The evaluation on simulated and real-world datasets shows that while the algorithm performs well over short forecast horizons, accumulating errors from IMU data and the assumption of a static environment limit its long-term accuracy. The future depth map prediction algorithm reduced the need for DynaDepth from 10 runs per second to 2, and on the Mid-Air dataset, from 25 to 5. Additionally, this study provides a foundation for future work, including the integration of an object-oriented frame prediction algorithm.
Supporting machine vision system design for quality control in manufacturing
An automotive case study
Master thesis
(2024)
-
L.W. van Keulen, A. Napoleone, R. Leite Patrão, A. Andersson, Y. Jacquet, B. Atasoy, R. Sabzevari
Quality control is considered an important process in manufacturing to minimise waste related to the manufacturing process. A way of performing quality control is with help of machine vision. Understanding all decisions that are to be made when designing a machine vision system for quality control is essential, but challenging. This research presents a method that aids decision-makers in the process of designing these machine vision systems for quality control so the efficiency of the decision-making process can be increased. The method results in a knowledge base and a tool that can both be considered as useful to decision-makers. The method was designed by following a Design Science Research approach and the method was applied on a use case in the automotive industry to prove that it addresses its goal.
...
Quality control is considered an important process in manufacturing to minimise waste related to the manufacturing process. A way of performing quality control is with help of machine vision. Understanding all decisions that are to be made when designing a machine vision system for quality control is essential, but challenging. This research presents a method that aids decision-makers in the process of designing these machine vision systems for quality control so the efficiency of the decision-making process can be increased. The method results in a knowledge base and a tool that can both be considered as useful to decision-makers. The method was designed by following a Design Science Research approach and the method was applied on a use case in the automotive industry to prove that it addresses its goal.