The project investigates a progressive reinforcement learning–driven autonomy for UAVs, beginning with PPO-based micro-waypoint navigation in a fixed simulation environment and extending toward camera-based (monocular and stereo) vision policies for obstacle avoidance. Building upon learned goal-directed control, the framework is designed to incorporate perception-aware decision making, enabling policies conditioned on visual observations rather than purely geometric state inputs. The research direction advances toward bird’s-eye-view (BEV)–aware multi-agent localization and coordination, where multiple UAVs share spatial context for cooperative navigation.
Unlike high-level planners, the proposed system directly learns micro-waypoint navigation policies that output continuous position setpoints in the NED frame. The learned policy demonstrates stable convergence to goal locations, obstacle avoidance, and generalizable behavior under deterministic inference.
| Component | Description |
|---|---|
| ΔN | Normalized North distance to goal |
| ΔE | Normalized East distance to goal |
| Roll | UAV roll angle |
| Pitch | UAV pitch angle |
The action space includes:
| Action | Description |
|---|---|
| aₙ | Northward micro-step |
| aₑ | Eastward micro-step |
Rewards Function is designed to be:
r = 3 · (dₜ₋₁ − dₜ) − 0.005 · ||a||² − 0.001 + 10.0 if goal reached
Figure1: World Environment, Figure2: Quadcopter cam_down FOV
Refer to: