Publications
* denotes equal contribution and joint lead authorship.
2026
Uncertainty-Aware Probabilistic Forecasting of Non-Cooperative Dynamic Obstacles for Safe Autonomous Aerial Navigation
Ph.D. Dissertation, Texas A&M University-Corpus Christi, Corpus Christi, Texas
@phdthesis{syedphddissertation, author = {Ullah, Syed Izzat}, title = {Uncertainty-Aware Probabilistic Forecasting of Non-Cooperative Dynamic Obstacles for Safe Autonomous Aerial Navigation}, school = {Texas A\&M University-Corpus Christi}, year = {2026}, address = {Corpus Christi, Texas}, type = {Ph.D. dissertation} }Unmanned Aerial Vehicles (UAVs) are increasingly deployed in complex, dynamic environments for applications such as urban air mobility, surveillance, emergency response, and package delivery. Safe autonomous navigation in airspace shared with non-cooperative dynamic obstacles, such as birds, unregistered drones, and other uninformed aerial agents, remains a fundamental challenge in aerial robotics. Existing motion planning methods are constrained by significant limitations that undermine their effectiveness in real-world scenarios. First, they often rely on linear or constant-velocity motion models that do not capture the nonlinear and context-dependent behavior of dynamic obstacles, which can lead to delayed or inadequate avoidance maneuvers. Second, most trajectory planning frameworks lack modularity, treating obstacle prediction and planning as separate processes, which hinders proactive decision-making and limits the integration of advanced prediction models. Third, the absence of standardized benchmarking infrastructure to characterize the platform-level uncertainties such as actuator nonlinearities, estimator drift, and battery-dependent thrust degradation on resource-constrained aerial systems, that affect the perception-prediction-planning pipeline.
This dissertation addresses these limitations through four contributions within a modular framework for uncertainty-aware UAV trajectory planning. First, POF+MADER integrates a decentralized Kalman filter-based Probabilistic Obstacle Filter with optimization-based trajectory planners, replacing the assumption of perfect obstacle trajectory knowledge with real-time uncertainty-aware forecasts embedded as dynamic constraints. Evaluation across 800 simulation runs yields a 38.75% improvement in success rate with a 0.8% increase in navigation time, and hardware experiments on Crazyflie 2.1 UAVs show a 25% reduction in collision rate under realistic sensing conditions. Second, SynTraG introduces a parametric trajectory generator that produces configurable three-dimensional corpora of non-cooperative aerial obstacle motion using randomized kinematic primitives with heteroscedastic noise, and releases 47,894 trajectories to address the lack of public training data in this domain. Building on this data foundation, the third contribution, AeroCast, combines a Pre-LN Transformer encoder with a Mixture Density Network output head to predict per-timestep Gaussian mixture distributions over future three-dimensional displacements. On a hybrid real-and-synthetic corpus of 90,116 sequences spanning nine motion categories, AeroCast reduces Average Displacement Error and Final Displacement Error by approximately 50% relative to the strongest recurrent baseline, achieves the lowest negative log-likelihood among all compared methods, and completes inference in 0.1 ms per sample. Finally, NanoBench introduces a multitask benchmark for the Crazyflie 2.1 nano-quadrotor comprising over 170 flight trajectories with synchronized Vicon ground truth, raw IMU measurements, onboard EKF estimates, PID controller internals, and motor PWM commands at 100 Hz. NanoBench defines standardized protocols for system identification, controller benchmarking, and state estimation, and establishes the first open dataset to jointly expose actuator-level and estimator-level signals with millimeter-accurate external ground truth on a commercially available nano-scale platform.
These contributions show that uncertainty in autonomous UAV navigation must be addressed across the full pipeline rather than within any single module. In this dissertation, uncertainty is treated at the planning level through probabilistic trajectory generation, at the data level through synthetic corpus construction with realistic localization noise, at the prediction level through calibrated multi-modal forecasting, and at the platform level through benchmarked characterization of actuation and estimation effects on resource-constrained hardware.
2026
AeroCast: Probabilistic 3D Trajectory Prediction for Non-Cooperative Aerial Obstacles via Transformer-MDN Architecture
IEEE Transactions on Aerospace and Electronic Systems (IEEE TAES) (Under Review)
@misc{ullah2026aerocastprobabilistic3dtrajectory, title={AeroCast: Probabilistic 3D Trajectory Prediction for Non-Cooperative Aerial Obstacles via Transformer-MDN Architecture}, author={Syed Izzat Ullah and Jose Baca}, year={2026}, eprint={2606.25122}, archivePrefix={arXiv}, primaryClass={cs.RO}, url={https://arxiv.org/abs/2606.25122}, }Existing aerial-robotics benchmarks target vehicles from hundreds of grams to several kilograms and typically expose only high-level state data. They omit the actuator-level signals required to study nano-scale quadrotors, where low-Reynolds-number aerodynamics, coreless DC motor nonlinearities, and severe computational constraints invalidate models and controllers developed for larger vehicles. We introduce \emph{NanoBench}, an open-source multi-task benchmark collected on the commercially available Crazyflie~2.1 nano-quadrotor (takeoff weight \SI{27}{\gram}) in a Vicon motion capture arena. The dataset comprises 172~flight recordings spanning multi-frequency excitation, geometric trajectory tracking at three speed regimes, and long-duration battery-drain hover. Each recording provides synchronized Vicon ground truth, raw IMU data, onboard extended Kalman filter (EKF) estimates, PID controller internals, and motor PWM commands at \SI{100}{\hertz}, alongside battery telemetry at \SI{10}{\hertz}, cross-aligned by gyroscope-based correlation on a \SI{1}{\milli\second} search grid. NanoBench defines standardized evaluation protocols, train/test splits, and open-source baselines for three tasks: nonlinear system identification, closed-loop controller benchmarking, and onboard state estimation assessment. To our knowledge, it is the first public dataset to jointly provide actuator commands, controller internals, and estimator outputs with millimeter-accurate ground truth on a commercially available nano-scale aerial platform.
2026
NanoBench: A Multi-Task Benchmark Dataset for Nano-Quadrotor System Identification, Control, and State Estimation
IEEE International Conference on Robotics \& Automation (IEEE ICRA 2027) (Under Review)
@misc{ullah2026nanobenchmultitaskbenchmarkdataset, title={NanoBench: A Multi-Task Benchmark Dataset for Nano-Quadrotor System Identification, Control, and State Estimation}, author={Syed Izzat Ullah and Jose Baca}, year={2026}, eprint={2603.09908}, archivePrefix={arXiv}, primaryClass={cs.RO}, url={https://arxiv.org/abs/2603.09908}, doi={10.48550/arXiv.2603.09908} }Existing aerial-robotics benchmarks target vehicles from hundreds of grams to several kilograms and typically expose only high-level state data. They omit the actuator-level signals required to study nano-scale quadrotors, where low-Reynolds number aerodynamics, coreless DC motor nonlinearities, and severe computational constraints invalidate models and controllers developed for larger vehicles. We introduce NanoBench, an open-source multi-task benchmark collected on the commercially available Crazyflie 2.1 nano-quadrotor (takeoff weight 27 g) in a Vicon motion capture arena. The dataset contains over 170 flight trajectories spanning hover, multi-frequency excitation, standard tracking, and aggressive maneuvers across multiple speed regimes. Each trajectory provides synchronized Vicon ground truth, raw IMU data, onboard extended Kalman filter estimates, PID controller internals, and motor PWM commands at 100 Hz, alongside battery telemetry at 10 Hz, aligned with sub-0.5 ms consistency. NanoBench defines standardized evaluation protocols, train/test splits, and open-source baselines for three tasks: nonlinear system identification, closed-loop controller benchmarking, and onboard state estimation assessment. To our knowledge, it is the first public dataset to jointly provide actuator commands, controller internals, and estimator outputs with millimeter-accurate ground truth on a commercially available nano-scale aerial platform.
2026
SynTraG: A Synthetic Trajectory Generator for Non-Cooperative Dynamic Obstacles in UAV Navigation
Advances in Automation and Robotics Research, Springer.
@ARTICLE{Ullah2025SYNTRAG, author={Izzat Ullah, Syed and Baca, José}, journal={Advances in Automation and Robotics Research, Springer}, title={SynTraG: A Synthetic Trajectory Generator for Non-Cooperative Dynamic Obstacles in UAV Navigation}, year={2026} }Unmanned Aerial Vehicles (UAVs) require robust navigation in dynamic environments to avoid collisions with non-cooperative obstacles like birds or other aerial objects, whose trajectories are non-linear, irregular, and not governed by road or air-traffic rules. Data-driven forecasting models (RNNs, Transformers) have shown strong performance in trajectory forecasting; however, progress in the UAV domain is constrained by the absence of publicly available data capturing such unstructured aerial motions. In this paper, we present SynTraG, a parametric syn- thetic trajectory generator that produces a configurable and extensive corpus of 3D trajectories by mixing kinematic primitives that emulate common non-cooperative motions: linear flight, oscillatory weaving, complex looping maneuvers, and ver- tical undulations. The generator is formulated as a mixture of closed-form kine- matic primitives with randomized parameters and explicit controls over path length, speed, acceleration, curvature, and spectral content. Heteroscedastic Gaussian noise is incorporated to model localization uncertainty(e.g., GPS/IMU errors) to enable probabilistic trajectory forecasting. We provide an open-source implementation of SynTraG and a sample dataset of 47,894 diverse trajectories, which enables robust training and benchmarking of forecasting models for UAV collision avoidance with non-cooperative dynamic obstacles. This work enables the development and stan- dardized benchmarking of next-generation forecasting models for safer autonomous aerial navigation.
2025
POF+MADER: Trajectory Planner in Multiagent and Dynamic Environments with Improved Collision Avoidance
IEEE Access, Vol 13 2025.
@ARTICLE{Ullah2025POFMADER, author={Izzat Ullah, Syed and Baca, José and Rangel, Pablo and Chu, Tianxing and Rubio-Medrano, Carlos}, journal={IEEE Access}, title={POF+MADER: Trajectory Planner in Dynamic Environments With Improved Collision Avoidance}, year={2025}, volume={13}, number={}, pages={215533-215549}, keywords={Trajectory;Vehicle dynamics;Heuristic algorithms;Collision avoidance;Dynamics;Autonomous aerial vehicles;Uncertainty;Real-time systems;Trajectory planning;Probabilistic logic;Collision avoidance;multi-robot systems;unmanned aerial vehicles (UAVs);autonomous navigation;motion planning;trajectory planning;obstacle tracking}, doi={10.1109/ACCESS.2025.3646978} }Trajectory planning in dynamic environments is crucial for safe and efficient navigation of UAVs, particularly in scenarios involving unknown dynamic obstacles. The Trajectory Planner in Multiagent and Dynamic Environments (MADER) algorithm has been successfully used for this purpose; however, it assumes perfect knowledge of the future trajectories of the dynamic obstacles. In practical scenarios, the future trajectory of the obstacles may not be known in advance. To address this limitation, we propose POF+MADER, an enhanced version that integrates a probabilistic obstacle filter with adaptive prediction horizons based on obstacle velocity. Our approach integrates a decentralized real-time Kalman Filter-based trajectory estimation with MADER’s planning core, eliminating the need for perfect a priori knowledge. We validate our approach through simulations with sensor noise (σ = 0.2 − 0.5m ) to simulate real- world localization uncertainty, with performance benchmarked against the MADER, Robust-MADER, and Ego-Swarm planner baselines. To further demonstrate its practical applicability, we conducted real-world experiments on the Crazyflie UAV platform, comparing its performance to the original MADER algorithm. Results show a 38.75% collision rate reduction in simulation and a 25% reduction in real-world experiments, without a significant increase in average navigation time. Furthermore, even in the presence of real-time sensor-induced noise in obstacle position measurements, our approach consistently outperforms the baseline planners, demonstrating its robustness in trajectory planning under perception uncertainty.
2023
Coaxial Modular Aerial System and the Reconfiguration Applications
In Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) 2023.
@INPROCEEDINGS{baca2023CMAS, author={Baca, José and Ullah, Syed Izzat and Rangel, Pablo}, booktitle={2023 IEEE International Conference on Robotics and Automation (ICRA)}, title={Coaxial Modular Aerial System and the Reconfiguration Applications}, year={2023}, volume={}, number={}, pages={11929-11935}, keywords={Connectors;Three-dimensional displays;Automation;Navigation;Trajectory;Task analysis;Payloads}, doi={10.1109/ICRA48891.2023.10161064} }This paper presents a coaxial modular aerial system (CMAS) formed by homogeneous modules driven by their center of mass. CMAS is designed to perform independent and cooperative flight with or without payload. Properties of the modularity concept allow the system to adapt to different situations and/or tasks by adding/removing modules to/from a configuration. The CMAS module is based on a coaxial motor and a two degree-of-freedom mechanism that transfers its center of mass from one side to another to make the module navigate around. The magnetic-based connector mechanism allows the module to be attached to other modules and to different metallic surfaces. A decentralized and asynchronous 3D path planning algorithm is implemented to avoid the trajectories of other modules/obstacles and ensures safe reconfiguration of the modules. Simulations within various environments show the applicability of the reconfiguration algorithm.
2023
Autonomous Navigation and Mapping of Snake Robots for Urban Search and Rescue.
In Proceedings of the IEEE International Conference on Robotics and Automation in Industry 2023.
@INPROCEEDINGS{syed2023snakerobot, author={Ullah, Syed Izzat and Mahmood, Tallat and Anayatullah}, booktitle={2023 International Conference on Robotics and Automation in Industry (ICRAI)}, title={Autonomous Navigation and Mapping of Snake Robot for Urban Search and Rescue}, year={2023}, volume={}, number={}, pages={1-8}, keywords={Three-dimensional displays;Simultaneous localization and mapping;Navigation;Service robots;Linux;Simulation;Snake robots;Snake robot;navigation;Mapping;Path Planning;ROS;CoppeliaSim;USAR;RRT*}, doi={10.1109/ICRAI57502.2023.10089544} }In an Urban Search and Rescue (USAR) situation, under extreme time pressure, rescue workers have to locate and extract the trapped people in collapsed structures. Due to the lack of medical treatment, food, and water, the victim's mortality rate dramatically increases over time. Rescue operations for both rescue workers and victims might be as dangerous as the initial event. For such situations, snake robots which are inspired by their biological counterparts, are shown to be a good option in the literature, to help the rescue workers in positioning the victims or delivering life-saving drugs to extend the life of the victims for some time. However, current research mainly focuses on mechanical design, control mechanisms, and gait generation. To alleviate this concern, we have integrated state-of-the-art methods to develop an autonomous snake robot that can navigate in an unknown environment while also generating a 3D map, to provide a better idea of the environment to the rescue workers. A simulated maze environment is implemented and demonstrated by using the CoppeliaSim simulation, running on Robot Operation System (ROS) and Linux OS. The simulation result shows the effectiveness of the proposed autonomous navigation system for the snake robot to plan an obstacle-free path from the robot's current position to the goal position without an apriori knowledge of the environment.
2023
Autonomous Navigation and Mapping of Water Channels in a Simulated Environment Using Micro-Aerial Vehicles
In Proceedings of the IEEE International Conference on Robotics and Automation in Industry 2023
@INPROCEEDINGS{syed2023uav4canal, author={Ullah, Syed Izzat and Muhammad, Abubakr}, booktitle={2023 International Conference on Robotics and Automation in Industry (ICRAI)}, title={Autonomous Navigation and Mapping of Water Channels in a Simulated Environment Using Micro-Aerial Vehicles}, year={2023}, volume={}, number={}, pages={1-7}, keywords={Meters;Irrigation;Automation;Navigation;Atmospheric modeling;Inspection;Path planning;UAV;navigation;Mapping;Path Planning;ROS;Unreal engine;Airsim;RRT*}, doi={10.1109/ICRAI57502.2023.10089599} }Irrigation canal networks serve as the bedrock of agriculture sectors across the globe as they are the primary channel through which water runs from major sources to agricultural lands. However, the water-carrying capacity of these water channels significantly reduces over time because of erosion, structural deterioration, and silt accumulation. As a result, routine inspections are required to analyze and repair these water channels which necessitates automation because of the vast length of these channels. We present a framework that enables Micro-Aerial Vehicles(MAVs) not only to navigate in an unknown cluttered canal environment but also to provide a complete 3-Dimensional map for the inspection. The framework consists of three main components (mapping, path planning, and mission planner) that gradually explore the environment while solving for start to local goal queries. We use Octomap; an octree-based representation of the environment for mapping, and we extended the Informed Rapidly-exploring Random Tree (Informed-RRT*) for optimal path planning and replan paths with respect to the static nearby and dynamic obstacles perceived during the execution of the mission. A simulated 2,378 meters length of canal environment is implemented and demonstrated by using the Airsim simulation in the Unreal engine, running on Robot Operation System (ROS) and Linux OS. Results obtained show that the framework enables the MAV to navigate over a simulated canal environment and allows the MAV to map the 3D structure of the canal.
2021
Motion Planning for a Snake Robot using Double Deep Q-Learning
In Proceedings of the IEEE International Conference on Artificial Intelligence (ICAI) 2021.
@INPROCEEDINGS{khan2023ddqn, author={Khan, Semab Neimat and Mahmood, Tallat and Ullah, Syed Izzat and Ali, Khawar and Ullah, Anayat}, booktitle={2021 International Conference on Artificial Intelligence (ICAI)}, title={Motion Planning for a Snake Robot using Double Deep Q-Learning}, year={2021}, volume={}, number={}, pages={264-270}, keywords={Training;Adaptation models;Simulation;Snake robots;Object detection;Concrete;Planning;Double deep Q learning;Experience replay memory;Model free;Motion planning;Off-policy;Snake robot}, doi={10.1109/ICAI52203.2021.9445200} }Motion planning for a snake robot in an unknown complex environment is a long-standing research problem because of the complex control of the modular mechanism. We propose deep reinforcement learning-based novel framework for motion planning. In this model-free framework, we propose a double deep Q-leaming-based technique to learn the optimal policy for reaching the goal point from a random start point; in a minimum number of steps in various unknown environments. In this approach, the agent learns to minimize the distance between the current and goal positions by aligning its yaw angle to the goal points through controlling multiple locomotive gaits. For experimental evaluation, we trained and tested the model in obstacle-free terrains. For training, we selected the model on the mud-terrain and tested for 50 episodes on five different terrains concrete, default, metallic, mud, and wooden. From simulation results, we observe the learned-optimal policy shows promising results for all unknown environments with a performance efficiency of 100% for all terrains except the wooden-terrain where it fails for only one episode and achieves 98% efficiency.
2019
Vision-Based Autonomous Mapping & Obstacle Avoidance for a Micro-Aerial Vehicle (MAV) Navigating Canal
Lahore University of Management Sciences (LUMS) 2019.
@INPROCEEDINGS{syed2023uav4canal, author={Ullah, Syed Izzat and Muhammad, Abubakr}, booktitle={2023 International Conference on Robotics and Automation in Industry (ICRAI)}, title={Autonomous Navigation and Mapping of Water Channels in a Simulated Environment Using Micro-Aerial Vehicles}, year={2023}, volume={}, number={}, pages={1-7}, keywords={Meters;Irrigation;Automation;Navigation;Atmospheric modeling;Inspection;Path planning;UAV;navigation;Mapping;Path Planning;ROS;Unreal engine;Airsim;RRT*}, doi={10.1109/ICRAI57502.2023.10089599} }For optimal agriculture yields, the land designated for agriculture requires proper irrigation. The major part of Pakistan’s irrigation system is served through canals that are outflows of dams, barrages, and rivers. The total length of the irrigating canal network in the Indus Basin is about 57,000 kilometers[47]. The manual labor for inspecting and analyzing the condition of water canals and then repairing during the closure period each year cannot be possible by laborious manual inspection. Hence, there is a need for automation in inspection tasks related to the water channels. The capability of moving robots in accomplishing desired tasks with the help of a couple of sensors mounted on them, not only decreases computation time but guarantees task completion as well. To be able to use robots for canal-like environments, one requirement is obstacle avoidance. We have developed a near to real canal environment in the Unreal engine, with real textures of trees, brushes, and bridges. Various conceivable obstacles that an aerial vehicle can confront in a canal-like environment is added in the simulation. Robot Operating System (ROS)[10] plays an important role in the simulation as it offers methods and groups of useful libraries (libraries for frame transformations, point cloud processing, visualization, and data monitoring to name a few). The Airsim plugin[55] in the Unreal engine[11] is used to simulate the MAV in the canal. We used Stereo image P roc[9] for disparity and point cloud construction, Octomap server[61] for Octomap Mapping, Informed RRT *[31] for path planning, and Airsim Simple FlightController[55] for vehicle control. The vehicle is capable of profiling the canal channels quickly and effectively that will assist the human operator in surveying the canal during an annual canal closure. The developed system autonomously flies over the canal, not only build a threedimensional map of the canal environment but detects obstacles in the path of the vehicle and eventually avoid these obstacles.