Autonomous drone navigation with Soft Actor-Critic
Master thesis (UiA) · high-fidelity simulation before real-world deployment.
Goal
Train a policy so a drone can navigate a 3D environment safely and efficiently, using reinforcement learning where exploration is cheap in simulation compared to the real world.
Approach
- Algorithm: Soft Actor-Critic (SAC) for continuous control.
- World: Unreal Engine with Microsoft AirSim for photorealistic scenes, sensors, and physics-oriented feedback.
- Loop: agent observes state → chooses continuous actions → receives reward / penalty shaped around progress, obstacles, and stability.
Illustration placeholder
Add a screenshot from Unreal/AirSim, a training reward curve, or a short trajectory GIF when you are ready — keeps this page lightweight for now.
Outcome
Demonstrates end-to-end RL experimentation in a realistic sim stack: environment integration, reward design, training stability, and evaluation — relevant to robotics and autonomy roles that start in simulation.