Alpha Drone is an experimental drone flight control architecture focused on agile, sub-human level acrobatic flight and reactive obstacle avoidance. The system is designed to bridge the gap between high-level path planning and low-level instinctive flight maneuvers.
Currently, the project is deployed as a fully simulated proof-of-concept, serving as a foundation for future physical hardware integration.
The control system is divided into a two-layer hierarchy to manage complex flight dynamics:
- The Brain (High-Level Planning): Utilizes Latent Diffusion Planning to generate sophisticated, predictive flight paths through complex environments.
- The Spine (Low-Level Reflexes): Acts as the reactive layer, handling immediate, short-latency obstacle avoidance and executing the "instinctive" acrobatic maneuvers required to stay airborne in tight spaces.
This project is built using modern robotics middleware and simulation tools. The current development environment runs locally and consists of:
- Middleware: ROS 2 (Jazzy)
- Flight Controller: PX4 Autopilot (SITL)
- Simulation: Gazebo (Harmonic)
- Simulation Only: The project is currently running entirely in simulation to safely train and validate the Latent Diffusion Planning models and the Spine reflex integration.
- Active Development: Focus is currently on optimizing the latency and communication between the Brain and Spine layers to ensure seamless acrobatic responsiveness.
- Hardware: No physical drone hardware is currently implemented. Hardware-in-the-loop (HITL) and real-world deployment are planned for future phases.
- Improve avoidance effectiveness and environment awareness for both Brain and Spine layers
- Finalize integration structure and latency optimization between th two layers.
- Expand simulated environments to include more complex, dynamic obstacles.
- Deploy onto physical edge computing hardware onboard a real drone.
Free environment fly:
Training loop environment:

