Panagiotis
Sarikas

Task: Navigate a quadrotor through increasingly complex indoor environments (open → cluttered → maze-like) visiting multiple goal checkpoints, while avoiding both known static obstacles and unknown dynamic ones that appear mid-flight.

Solution: A three-layer hierarchical architecture combining global path planning with local reactive control. A roadmap (PRM or Sukharev grid) is built once from the static map, and A*/Dijkstra computes shortest paths for each goal. An MPPI (Model Predictive Path Integral) controller follows this path in real-time, sampling thousands of candidate trajectories to track the reference while dodging unknown obstacles. A low-level PD attitude controller handles drone stabilisation. The system is evaluated in 2D and 3D simulation using gym-pybullet-drones.

Task: Develop a knowledge-based decision-making system for a mobile robot capable of autonomously tidying a simulated apartment, reasoning about object properties and locations, planning actions, and executing household tasks using symbolic planning and knowledge representation techniques.

The assignment included three tasks:
1. Simple Tidy Up: pick up all objects and place them in the closest drop location within the same room. 2. Tidy Up Following House Rules: place each object in the correct drop location based on semantic rules — tableware, toys, trash, or bedroom items. 3. Find and Bring an Object: locate and deliver a requested book using contextual clues and handling blocked doorways when needed.
Solution: Implements a hybrid autonomous robot architecture that combines symbolic planning (PDDL, PlanSys2) with a semantic knowledge base (TypeDB) for decision-making in dynamic, partially observable environments. The system enables adaptive task execution through real-time world modelling, semantic inference, and modular multi-stage planning. Evaluated on RoboCup-inspired tasks including object manipulation, semantic tidying, and contextual object retrieval in ROS2/Gazebo simulation.

Task: A mobile robot navigates simulated rooms collecting yellow “ducks.” A human operator manually controls the robot, generating demonstration data (RGB images + actions). The goal is to train a classifier that imitates the human’s behaviour — mapping raw visual observations to discrete navigation actions (move forward, turn right, turn left).

Solution: A robot learns a policy that maps RGB camera observations to discrete control actions by imitating human demonstrations. The task is formulated as a supervised multi-class classification problem, where each observation is classified into one of three actions: turn left, turn right, or move forward. A classifier (Random Forest, outperforming Decision Trees and CNN on this task) is trained on human demonstrations. The learned policies are evaluated both offline using classification metrics and online in a simulator to assess real-time control performance and generalisation to unseen environments.

Task: Detect pedestrians in 3D space (camera reference frame) at distances of 5–40m, for pedestrian heights of 1.2–1.9m.

Solution: A detect-then-localize pipeline that combines YOLO-based 2D pedestrian detection with LiDAR and radar data to estimate accurate 3D pedestrian positions in the camera frame. The system applies geometric constraints, sensor fusion, and ground-plane backprojection to robustly localise pedestrians under varying distances and scene conditions.

Task: A mobile manipulator robot (Mirte) must autonomously navigate a greenhouse environment, detect and interact with plants, and maintain a live digital twin of the greenhouse state. The system must integrate perception, navigation, manipulation, and knowledge representation to support autonomous plant monitoring and data collection.

Solution: A multidisciplinary ROS 2 system combining autonomous navigation (Nav2), plant detection via computer vision, manipulation planning, and real-time knowledge graph updates. The digital twin is maintained through continuous sensor fusion and environment mapping, providing a queryable representation of the greenhouse state for higher-level task planning.

Task: A simulated mobile robot (Mirte) must autonomously navigate a cone-defined track, avoiding 3D obstacles detected from a depth camera point cloud and stopping permanently when a pedestrian is detected nearby in the RGB camera image.

Solution: A modular ROS 2 system with three nodes communicating via topics. A PCL-based obstacle detector processes depth camera point clouds to extract 3D bounding boxes of cones/barrels. An OpenCV-based pedestrian detector (pre-built) produces 2D bounding boxes of people. A control node fuses both detection streams, transforms obstacle positions from camera frame to robot frame using TF2, and publishes velocity commands to steer around obstacles and halt for pedestrians.

Task: Design a constrained controller for a vehicle lane-keeping assist system that regulates lateral position and heading errors while respecting steering limits, lane boundaries, and model validity constraints — something PID and LQR cannot guarantee.

Solution: A constrained MPC framework built on a linearized dynamic bicycle model. The vehicle state tracks lateral position error, lateral velocity error, heading error, and yaw rate error, with steering angle as the single control input. Terminal ingredients from the discrete algebraic Riccati equation (DARE) ensure closed-loop stability. Two formulations are developed: a regulation MPC for driving errors to zero, and an offset-free output-feedback MPC that uses a Luenberger observer to reject persistent road-curvature disturbances and measurement noise. Performance is benchmarked against LQR.

Task: Subtask 1 uses a 4-joint redundant arm to track an endpoint trajectory while keeping an intermediate joint near the origin as a secondary task. Subtask 2 derives the full physics model of a 2-joint torque-controlled arm and controls it with an impedance controller along the same trajectory.

Solution: In subtask 1, two PID controllers run in parallel. The primary one tracks the endpoint using a damped pseudo-inverse Jacobian, while the secondary one targets the mid-chain joint and its velocity is projected into the null space so it never disturbs the endpoint. In subtask 2, the Lagrangian method produces the mass, Coriolis, and gravity terms. These drive a forward dynamics simulation each timestep. An impedance controller generates a spring-damper force from position and velocity error, converted to joint torques via the Jacobian transpose.

Modeling and simulation of an electronic differential for a lightweight electric vehicle with Ackermann steering geometry. Compares two in-wheel motor architectures — DC motor and PMSM — using FOC and SVPWM to evaluate vehicle stability, traction performance, and cornering behavior.