Artificial Intelligence & Robotics (Current Research)

Optimization and Learning-based Approaches to Extending Autonomy in Resource-Constrained Marine Environments

Approximately 40% of the global population resides within 100 kilometers of coastal or estuarine environments. As population density and economic activities in these regions increase, so do the stressors on coastal ecosystems. Consequently, there has been growing attention on the application of intelligent autonomous systems (IAS), including autonomous underwater vehicles (AUVs) and autonomous surface vehicles (ASVs), for the continuous monitoring of aquatic environments. 

My current research directions aim to address the theoretical and technical challenges with creating a comprehensive framework for autonomous sampling, navigation, and human-robot collaboration in dynamic marine environments.

To achieve these objectives, we propose a research agenda that will be focused on the following research topics:

• Research Topic 1: Autonomous Robotic Information Gathering Systems

Overview: Robotic Information Gathering (RIG) is a process of optimizing an information-theoretic metric from efficient exploration of a region of interest by robots with motion constraints while considering inferences from a probabilistic model within a limited mission time. Effective data collection in collaborative information-gathering systems relies heavily on maintaining uninterrupted connectivity. Yet, real-world communication disruptions often pose challenges to information-gathering processes. 

Contributions: In our recent contribution [1], we propose an approach—a decentralized information-gathering system with limited communication designed for multiple robots to explore environmental phenomena represented as unknown spatial fields. The approach utilizes quadtree structures to achieve thorough workspace coverage and facilitate efficient exploration. Unlike conventional systems that require global and synchronized communication, this method allows robots to share localized information within restricted transmission ranges and coordinate tasks through pairwise, asynchronous interactions. 

The information estimation process employs a Gaussian Process with an Attentive Kernel, enabling the adaptive identification of critical patterns and behaviors in the data. This approach is validated through simulations in dynamic environments, where multiple robots explore and map non-stationary scalar fields. The method is supported by theoretical guarantees establishing the convergence of distributed area coverage and providing regret bounds for online distributed scalar field mapping.

Related publications: 

• [1] A. Redwan, P. Padrao, J. Fuentes, T. Alam, G. Govindarajan, L. Bobadilla. LCD-RIG: Limited Communication Decentralized Robotic Information Gathering Systems. IEEE Robotics and Automation Letters, 2024

• Research Topic 2: Learning-based Approaches for Autonomous Environmental Monitoring

Overview: Over the past few decades, advances in machine learning and data processing have significantly contributed to the exploration and sampling of aquatic environments using autonomous robots. However, predicting and estimating phenomena of interest in these environments remain challenging due to their complex spatio-temporal dynamics. In response, this part of my research introduces novel frameworks that address these challenges by leveraging learning techniques and adaptive navigation for improved field estimation and mapping [2, 3].

Contributions: Recently, I have been working on learning-based adaptive navigation for scalar field mapping and feature tracking. Scalar field features such as extrema, contours, and saddle points are essential for applications in environmental monitoring, search and rescue, and resource exploration. Traditional navigation methods often rely on predefined trajectories, leading to inefficient and resource-intensive mapping. In [4], we introduce a new adaptive navigation framework that leverages learning techniques to enhance exploration efficiency and effectiveness in scalar fields, even under noisy data and obstacles.

Related publications: 

• [2] P. Padrao, A. Dominguez, L. Bobadilla, and R. N. Smith. Towards learning ocean models for long-term navigation in dynamic environments. IEEE Oceanic Engineering Society - OCEANS 2022.

• [3] P. Padrao, J. Fuentes, L. Bobadilla, and R. N. Smith. Estimating spatio-temporal fields through reinforcement learning. Frontiers in Robotics and AI, 9, 2022.

• [4] J. Fuentes, P. Padrao, A. Redwan, L. Bobadilla, Scalar Field Feature Seeking with Learning-based Adaptive Robot Navigation. ICRA 2025, IEEE International Conference on Robotics and Automation 

• Research Topic 3: Digital Twins as Robotic Training Tools and Human-Robot Interface Design

Overview: We explore how virtual environments—powered by digital twins and extended reality (XR) technologies—can support the development of natural and effective robotic training scenarios. In this context, the design of intuitive, robust, and efficient Human-Robot interfaces (HRI) can enable a new generation of applications in critical areas such as precision agriculture, automated construction, rehabilitation, and environmental monitoring [5, 6]. These interfaces, which aim to foster seamless and natural collaboration between humans and robots, rely on key properties like consistency, linearity, and continuity.

Contributions: Building upon recent work [7], we focus on exploring mappings between human and robot spaces and developing an optimization-based framework for designing natural human-robot interfaces. This aligns with recent advancements in HMD-based immersive teleoperation interfaces, human perception-optimized planning, and strategies that leverage optimal control for teleoperated robots. In [8], we introduced a cloud-based platform for underwater robotics education, integrating virtual reality for teleoperation and a digital twin for real-time interaction with physical equipment.

 Related publications: 

• [5] Kaarlela, T.; Padrao, P.; Pitkäaho, T.; Pieskä, S.; Bobadilla, L., 2023. Digital Twins Utilizing XR-Technology as Robotic Training Tools. Machines 2023, 11, 13.

• [6] Kaarlela, T.; Pitkäaho, T.; Pieskä, S.; Padrão, P.; Bobadilla, L.; Tikanmäki, M.; Haavisto, T.; Blanco Bataller, V.; Laivuori, N.; Luimula, M. Towards Metaverse: Utilizing Extended Reality and Digital Twins to Control Robotic Systems. Actuators 2023, 12, 219.

• [7] P. Padrao, J. Fuentes, T. Kaarlela, A. Bayuelo, L. Bobadilla., Towards Optimal Human-Robot Interface Design Applied to Underwater Robotics Teleoperation. 2024 IEEE CASE - International Conference on Automation Science and Engineering

• [8] D. Correa, P. Padrao, J. Fuentes, T. Kaarlela, L. Bobadilla., A Web-based Interactive Digital Twin for Marine Robotics Education

• Research Topic 4: Knowledge-Guided Behavior Forecasting and Uncertainty-Aware Navigation 

1. Uncertainty-Aware Navigation: We build on the GUIDEd Agents framework [9], which enables robots to incorporate task-specific uncertainty representations directly into their navigation policies via Task-Specific Uncertainty Maps (TSUMs). In aquatic environments, we adapt this framework to model spatially-varying uncertainty across different regions, allowing underwater robots to dynamically adjust their localization precision based on task-critical requirements. This approach helps the robot reason about where high accuracy is essential, improving task performance while conserving computational and energy resources. 

2. Knowledge-Driven Behavior Forecasting: In [10], we introduce TRACE (Tree-of-thought Reasoning And Counterfactual Exploration), an inference framework that couples tree-of-thought generation with domain-aware feedback to refine behavior hypotheses over multiple rounds. In this context, a Vision-Language Model (VLM) first proposes candidate trajectories for the agent; a counterfactual critic then suggests edge-case variations consistent with partial observations, prompting the VLM to expand or adjust its hypotheses in the next iteration. This creates a self-improving cycle where the VLM progressively internalizes edge cases from previous rounds, systematically uncovering not only typical behaviors but also rare or borderline maneuvers, ultimately yielding more robust trajectory predictions from minimal sensor data. 

3. Data-Driven Environmental Models for State Estimation: Motivated by the need for prompt solutions to detect and respond to harmful water events, such as algal blooms or fish kills, our recent work [11] aimed to improve the state estimation of underwater robots by incorporating detailed 3D maps of underwater regions, known as Satellite-Derived Underwater Environments (SDUEs). Our methodology leverages remote sensing imagery from the Sentinel-2 (S2) and Landsat 8-9 (L8-9) satellites with in-situ water measurements from Biscayne Bay, Florida.

Related publications: 

• [9] G. Puthumanaillam, P. Padrao, J. Fuentes, L. Bobadilla, M. Ornik. GUIDEd Agents: Enhancing Navigation Policies through Task-Specific Uncertainty Abstraction in Localization-Limited Environments. arXiv preprint arXiv:2410.15178, 2024. 

• [10] G. Puthumanaillam, P. Padrao, J. Fuentes, P. Thangeda, W. E. Schafer, J. H. Song, K. Jagdale, L. Bobadilla, M. Ornik. TRACE: A Self-Improving Framework for Robot Behavior Forecasting with Vision-Language Models. arXiv preprint arXiv:2503.00761, 2025. 

• [11] C. Rojas, P. Padrao, J. Fuentes, G. Reis, A. Albayrak, B. Osmanoglu, L. Bobadilla. Combining Multi-Satellite Remote and in-situ Sensing for Unmanned Underwater Vehicle State Estimation. Elsevier Journal of Ocean Engineering, 2024

Fault Detection & Diagnosis

Sensor Fault Detection and Diagnosis Approaches applied to an Autonomous Solar-powered Aircraft

Developed by Elektra Solar, a spin-off of the DLR Institute for Robotics and Mechatronics (DLR-RMC), the Elektra 2 Solar is a solar-powered autonomous aircraft designed for long-distance and high-altitude missions. Elektra 2 Solar uses a simple limit-checking approach for certain measurements, such as angular velocities and attitude angles. However, this oversimplified detection method can hide faulty system behaviors, potentially leading to missed fault alarms and posing serious risks to the aircraft's overall system.

The motivation of this work is to compare various IMU sensor fault detection and diagnosis (FDD) approaches that can be applied to the Elektra 2 aircraft.

The first proposed FDD approach is based on decoupled lateral and longitudinal linear models of the aircraft, combined with Kalman filters for residual generation. An adaptive threshold technique (ATLMS) is used for fault detection. 

The second approach utilizes the aircraft's kinematic model in conjunction with an extended Kalman filter for residual generation, with a decision table based on alarm flag activation sequences for fault diagnosis. 

The third FDD approach is model-free, relying on principal component analysis (PCA), where contributions to squared prediction errors are used for fault diagnosis. Finally, a fourth FDD strategy is developed, leveraging the key advantages of the previous approaches.

Simulations were performed across various flight scenarios, applying additive faults to measurements of roll and pitch rates as well as longitudinal acceleration. Real flight data from nominal operations were used for validation purposes.

Related publications:

• Padrao, P., Hsu, L., Vilzmann, M. and Kondak, K., 2019, December. A Comparative Study of Sensor Fault Detection Approaches applied to an Autonomous Solar-powered Aircraft. In 2019 19th International Conference on Advanced Robotics (ICAR) (pp. 761-766). IEEE. 

• Padrao, P., Hsu, L., Vilzmann, M. and Kondak, K., 2019, October. Model-based sensor fault detection in an autonomous solar-powered aircraft. In FT2019. Proceedings of the 10th Aerospace Technology Congress, October 8-9, 2019, Stockholm, Sweden (No. 162, pp. 247-254). Linköping University Electronic Press. 

Optimal Threshold Synthesis for State-of-Health Classification and Evaluation of a Tribological System

The primary objective of this project was to develop a method to enhance state-of-health classification performance based on experimental data from a tribological system. The study was conducted at the Chair of Dynamics and Control, University of Duisburg-Essen. We introduced two performance metrics to evaluate system behavior: the Switching State Rate and Regeneration Time. Based on the analysis of experimental data, four distinct operational states were defined: 

• State 1 (stable and error-free operation)

• State 2 (stable with minor surface condition changes) 

• State 3 (acceptable surface condition changes)

• and State 4 (significant surface condition changes).

To facilitate optimization across different datasets, we employed an in-house Optimization Graphical Interface, which streamlined problem setup and enabled numerical optimization through MATLAB algorithms.

Related publications:

• Rothe, S., Padrao, P., Leite, A., and Söffker, D., 2015. Improvement and comparison of wear-oriented state-of-health classification methods using optimization techniques. Structural Health Monitoring 2015. 

• Padrao, P., Rothe, S., Leite, A. and Soeffker, D. (2015). Optimal Threshold Synthesis for State-of-Health Classification and Evaluation of a Tribological System. In: 17th International Symposium on Dynamic Problems of Mechanics. Brazilian Society of Mechanical Sciences and Engineering.

STEM Education