Academic homepage

Panagiota Grosdouli

I study how autonomous systems can perceive, predict, and act safely around people under uncertainty.

Diploma candidate in Electrical & Computer Engineering, Democritus University of Thrace — preparing for doctoral study in robotics and autonomous systems.

Trajectory Prediction
Safe Autonomy
Human-Aware Robotics
Visual-Inertial SLAM
Planning Under Uncertainty

My work treats autonomy as a coupled problem of perception, prediction, planning, and safety, rather than a single engineering pipeline. I am interested in how a system can represent what it does not yet know and how that representation should shape what it does next.

Research agenda Selected projects GitHub

Research agenda

Questions I want to spend the next years answering

How should an autonomous system act before a person's future trajectory is fully observable?
Most prediction models are evaluated on accuracy alone. I am interested in how uncertainty in a forecast should propagate into a decision, particularly in mixed-traffic settings involving pedestrians and cyclists.
Can perception systems know when they are wrong?
Visual and visual-inertial estimation can degrade quietly under blur, low light, or sensor dropout. I want to study how a system can detect its own degradation and recover rather than fail silently.
What does it mean for an intersection, rather than a single vehicle, to reason about safety?
Infrastructure-level sensing offers a vantage point no single agent has. I am interested in how shared perception and prediction could reduce risk for vulnerable road users.
How should a robot plan when its map of the world is incomplete?
Navigation in unknown, dynamic environments requires continual decisions about whether the current plan is still trustworthy or should be revised.

Current research

Trajectory prediction and uncertainty-aware autonomy for vulnerable road users

My diploma thesis is the primary line of research I am pursuing at present. It addresses the forecasting of pedestrian and vulnerable-road-user motion in traffic scenes where an autonomous system must commit to a decision before the future trajectory is fully observable.

This direction connects to Self-Healing VIO, which studies how perception degrades and recovers under adverse sensing conditions, and DynNav, which extends the same robustness perspective to planning and navigation in unknown environments.

Research themes

Areas my work is organized around

— Trajectory Prediction
— Safe Autonomous Systems
— Intelligent Intersections
— Human Motion Forecasting
— Visual-Inertial SLAM
— Adaptive Multi-Modal Perception
— Navigation Under Uncertainty

Selected research projects

Five research directions

Diploma Thesis

In progress

Research question

How can autonomous systems forecast vulnerable-road-user motion before making safety-critical decisions at intersections?

Scientific motivation

Pedestrian and cyclist behavior is variable and only partially observable at decision time. Prediction should therefore be evaluated not only by accuracy, but by how uncertainty affects downstream decisions.

Technical contribution

Trajectory-prediction models and an uncertainty-aware evaluation protocol for smart-intersection scenarios involving vulnerable road users.

SafeCrossAI

Active

Research question

How can intelligent intersections model interactions between vulnerable road users and connected or autonomous vehicles?

Scientific motivation

Infrastructure-level sensors observe interactions that no single vehicle can, offering an opportunity to reason about risk collectively rather than agent by agent.

Technical contribution

An infrastructure-level perception and prediction framework for safety-aware reasoning around vulnerable road users at intersections.

Repository →

Self-Healing VIO

Exploratory

Research question

How can visual-inertial SLAM remain reliable when tracking quality degrades under blur, low light, sensor noise, or frame dropout?

Scientific motivation

Visual-inertial estimators tend to fail quietly under adverse conditions, without a clear internal signal that tracking has degraded.

Technical contribution

A self-healing VIO approach that predicts tracking degradation and adapts its sensing or recovery strategy for robust UAV navigation.

Repository →

Adaptive Multi-Modal SLAM

Exploratory

Research question

How can high-speed UAV navigation adaptively fuse visual foundation models, event cameras, and uncertainty-aware sensing?

Scientific motivation

No single sensing modality is reliable across all conditions; combining modalities with explicit uncertainty may allow graceful degradation.

Technical contribution

A multi-modal SLAM framework with uncertainty-aware sensor fusion for high-speed navigation in visually challenging conditions.

Repository →

DynNav

Exploratory

Research question

How should a robot re-plan when it moves through unknown environments containing dynamic obstacles?

Scientific motivation

Navigation in unmapped, dynamic settings requires deciding whether a current plan is still trustworthy or should be revised.

Technical contribution

A dynamic navigation and rerouting approach for unknown environments, combining obstacle avoidance with planning under uncertainty.

Repository →

Research vision

Where I want this work to be in five years

Scientific challenges

Building prediction models whose confidence is calibrated enough to support safety-critical decisions; developing perception systems that can recognize their own degradation; and planning under incomplete maps and unobserved agents.

Longer-term direction

I hope to contribute to research connecting human-motion forecasting, robust multi-modal perception, and safe decision-making, narrowing the gap between benchmark performance and systems that can share space with people.

Publications

Research manuscripts, technical reports and publications will be shared here as they become available, with links to the corresponding papers and code.

News

2026Started diploma thesis research on trajectory prediction and uncertainty-aware autonomy for vulnerable road users.
2025Began development of SafeCrossAI, an intelligent-intersection research project.
2025Opened research repositories for Self-Healing VIO, Adaptive Multi-Modal SLAM, and DynNav.

Contact

panagros1@ee.duth.gr github.com/panagiotagrosdouli