Research Project · Democritus University of Thrace · ECE

Navigation that knows
what it doesn't know.

DynNav is a 26-contribution research framework for uncertainty-aware, risk-sensitive autonomous navigation in unknown environments — built on ROS 2 Humble, validated on TurtleBot3, spanning belief-space planning, formal safety shields, and Byzantine-fault-tolerant swarm consensus.

26
Research modules
79.4K
Lines of Python
ROS 2
Humble / TurtleBot3
72/72
Unit tests passing
01

Abstract & Motivation

Most navigation stacks assume the map, the sensors, and other agents behave well. Real deployments — disaster response, warehouse fleets, planetary rovers, assistive robots — break that assumption constantly. DynNav treats the failure modes of autonomy as first-class research problems rather than edge cases, decomposing navigation under uncertainty into 26 independent, testable contributions: each pairs a focused research question with implementation, experiments, and reproducible results, integrated into a common ROS 2 stack.

uncertainty quantificationrisk-sensitive planningCVaR optimization formal safety (STL/CBF)belief-space planningmulti-robot consensus reinforcement learningROS 2
Open research questions
  • State uncertaintyWhat should a robot do when it is uncertain about its own state, not just the world's?
  • Formal guaranteesHow do you give verifiable guarantees around components that are fundamentally statistical?
  • Adversarial robustnessWhat happens to multi-robot systems when an agent lies or fails, and how much can decentralization buy back?
  • ExplainabilityCan a robot explain its own failures in terms a human can act on, rather than a stack trace?
02

System Architecture

Four layers sit above a ROS 2 / Gazebo / TurtleBot3 foundation. Foundation models and the learning layer feed structured planning; the safety layer filters every command before it reaches the robot.

Foundation Models 11 · VLM Scene Agent 19 · LLM Mission Planner 20 · Failure Explainer Learning Layer 01 · Learned A* Heuristic 21 · PPO Nav Agent 22 · Curriculum RL · 16 · Federated Safety Layer 18 · STL + CBF Shield 05 · Safe-Mode FSM 04 · Returnability Coordination Layer 26 · Swarm BFT Consensus 16 · Federated Nav Learning 09 · Multi-Robot Coordination Planning Core A* / D* · Belief-Space & Risk Planning (03) · CVaR-weighted cost · Next-Best-View Exploration (07) Perception Layer LiDAR SLAM · EKF/UKF (02) · Gaussian Splatting (23) · NeRF Uncertainty (24) · Neuromorphic DVS+SNN (15) Security Layer Intrusion Detection (08) · Adversarial Attack Simulation (25) · Causal Risk Attribution (14) ROS 2 Humble · TurtleBot3 · Gazebo · Nav2 · slam_toolbox

Replanning pipeline

01

Sense

LiDAR / odometry stream into EKF–UKF belief-state estimation.

02

Estimate risk

Belief-space occupancy is converted into a CVaR-weighted risk map.

03

Plan

Risk-weighted A* / D* searches over the belief-space cost field.

04

Shield

STL monitors and CBF filters check the candidate command against safety constraints.

05

Execute / replan

Safe-mode FSM triggers conservative behaviour or replanning if risk crosses threshold.

03

26 Research Contributions

Each contribution is an independent module: a research question, an implementation, an experiment script, and a results folder. Filter by category, or browse all.

04

Selected Results

Figures reported below are drawn from each module's own experiment logs (contributions/*/results/*.csv). They reflect each module's standalone evaluation protocol, not a single unified benchmark.

Formal Safety Shields
Contribution 18 · STL monitor + CBF filter
MetricNo shieldWith shield
Constraint violations / episode4.2 avg0.3 avg
Path length overhead< 8%
Avg. command correction0.026 m/s
Byzantine Swarm Consensus
Contribution 26 · weighted-median BFT, 6 robots / 1 Byzantine
MetricNaive majorityBFT weighted median
Byzantine detection rate60%91%
Correct plan selected71%96%
Byzantine fault tolerancef < N/2f < N/3
Federated Navigation Learning
Contribution 16 · FedAvg + DP across 6 robots
RoundCentralised (MSE)Federated (MSE)
10.410.37
100.180.21
200.120.14
Curriculum Reinforcement Learning
Contribution 22 · adaptive 5-stage difficulty curriculum
TrainingEpisodes to "hard" stageFinal success rate
Flat (no curriculum)N/A23%
Adaptive curriculum~20061%
Reproducibility note — these tables are reported per-module results, not yet reproduced by a single top-level orchestration script. A unified benchmark harness producing all four tables from one command is on the project roadmap (see Future Work, §06).
05

Figures & Demonstrations

Placeholders for simulation captures and demo recordings — replace with exported assets from data/plots/ and Gazebo screen recordings.

RISK MAP
occupancy + CVaR overlay
PLANNER COMPARISON
A* vs learned-heuristic A*
ROS 2 NODE GRAPH
rqt_graph export
STL/CBF SHIELD
trajectory correction plot
SWARM CONSENSUS
Byzantine detection over time
DEMO VIDEO
TurtleBot3 / Gazebo run
06

Future Work

·Unified orchestration script reproducing all reported result tables from one command.
·End-to-end integration of the sense → risk → plan → shield → execute loop as one running ROS 2 demo, not 26 separate evaluations.
·Hardware validation of the safety-shield and safe-mode contributions on the physical TurtleBot3 Burger.
·Ablation studies isolating the contribution of risk-weighting versus learned heuristics in planning quality.
·Extending Byzantine-fault-tolerant consensus to heterogeneous robot teams with asymmetric sensing.
·Formal verification coverage report for the STL/CBF safety layer across all contributions that route through it.
07

Paper & FAQ

A full technical writeup is in TECHNICAL_REPORT.md in the repository. The questions below cover the project's scope honestly.

Is DynNav one integrated navigation system?

Not yet. It is 26 independent research modules sharing a common ROS 2 target and simulation environment. Several modules are integrated into the core planning/safety loop; others are standalone studies. See Future Work for the integration roadmap.

Are the result tables independently reproducible today?

Each table is produced by that module's own evaluation script. A single command reproducing every table at once does not yet exist — this is the top item in Future Work.

What hardware was this tested on?

TurtleBot3 Burger (physical) and TurtleBot3 Waffle in Gazebo, on ROS 2 Humble / Ubuntu 22.04, with WSL2 supported for development.

Can I build on this for my own thesis or project?

Yes — the repository is Apache-2.0 licensed. See the Citation section below for how to credit the work.

Author

Panagiota Grosdouli

Electrical & Computer Engineering
Democritus University of Thrace

Research interests: uncertainty-aware planning, risk-sensitive control, formal safety for learning-enabled robotic systems, multi-robot coordination.

Open to research collaboration and graduate study opportunities in robotics & autonomous systems.

GitHub Profile Contact
08

Citation

If you use DynNav in your research, please cite the software release.

@software{dynnav2026, author = {Grosdouli, Panagiota}, title = {{DynNav}: Dynamic Navigation Rerouting in Unknown Environments}, year = {2026}, publisher = {GitHub}, url = {https://github.com/panagiotagrosdouli/DynNav-Dynamic-Navigation-Rerouting-in-Unknown-Environments}, license = {Apache-2.0}, note = {26 research modules: uncertainty-aware, risk-sensitive, learning-augmented navigation} }