Subject Training

Abstract: Modern aerial robots must plan, perceive, and control robustly in dynamic environments in a safe manner. This research combines classical aspects from control and planning to Simultaneous Localization and Mapping (SLAM), then adds a novel reinforcement learning (RL) layer to push autonomy beyond hand tuned pipelines, all built and tested in ROS + Gazebo and high fidelity simulators (CARLA, AirSim). Scholarship Duration: 6 months Period: February 2026 - July 2026

Description: The project envisages a SLAM-first autonomy stack in which perception, simulation engineering, planning, and control are composed with safety as a primary requirement. The work centers on the development of a visual/visual-inertial SLAM module that provides pose, mapping, and basic uncertainty estimates and serves as the backbone for downstream decision and control. Simulation is a core activity: a ROS/ Gazebo pipeline will be engineered for rapid iteration and systematically bridged with AirSim to enable high-fidelity sensing/physics, domain randomization, and scenario stress-testing (dynamic obstacles, wind, sensor artifacts). On top of SLAM, motion planning will be implemented using graph-based (A*/D* Lite) and sampling-based (RRT/RRT*) methods with time-parameterized, kinodynamically feasible trajectories that account for safety margins and, where relevant, SLAM uncertainty. Control strategies will be investigated from PID to optimal control (LQR/MPC) for point-to-point motion and constrained trajectory following. Safe control learning is a principal objective: safety filters (e.g., Control Barrier Function-based QPs), runtime assurance, and robust constraint handling will be integrated into both the classical controllers and the learning layer. As a novelty component, reinforcement-learning approaches (residual or safety-shielded PPO/SAC) will be developed to augment tracking and local avoidance, with risk-aware rewards and ablations against classical baselines. Deliverables include a reproducible codebase (ROS/Gazebo ? AirSim), quantitative benchmarks (classical vs. classical+RL), a concise research report, a demo video, and preparation of a manuscript for submission to reputed venues. Objectives include:
⬢ Engineer a SLAM backbone: Implement and benchmark a visual/VIO SLAM pipeline (pose, mapping, basic uncertainty), establishing it as the primary state-estimation module.
⬢ Build the simulation stack: Configure a reproducible ROS/Gazebo environment (worlds, sensors, plugins) with scripted scenarios for rapid iteration.
⬢ Bridge high-fidelity simulation: Integrate AirSim ? ROS/Gazebo (time sync, topics, sensors) to enable physics- and sensor-realistic stress tests and domain randomization.
⬢ Path planning & trajectories: Implement A*/D* Lite and RRT/RRT*; generate time-parameterized, kinodynamically feasible trajectories (e.g., jerk-limited splines).
⬢ Baseline control: Develop basic PID based cntrol and preferebly extend to optimal control (LQR) for point-to-point motion and constrained trajectory tracking.
⬢ Safety layer for control: Design CBF-QP safety filters and/or runtime assurance (Simplex) to guarantee constraint satisfaction and minimize safety violations.
⬢ RL novelty: Develop residual or safety-shielded RL (e.g., PPO/SAC) that augments the nominal controller or local avoidance while respecting safety filters..