Inspiration
Racing has always been about precision, speed, and split-second decisions — exactly what AI does best. We wanted to push the boundaries of drone racing by fusing F1-style strategy with swarm intelligence, letting 50 autonomous drones learn, adapt, and race as a team. The idea: build an AI swarm that races like pros before ever touching real hardware.
What it does
DroneF1 Swarm is a multi-drone AI racing simulator where 50 autonomous drones compete through F1-inspired gates. Each drone uses reinforcement learning to minimize crashes, optimize turns, and coordinate with others in real time. The learned strategies can then be exported and deployed to Haas drones, bridging the gap between simulation and real-world performance.
How we built it
Simulation Engine: Built using Pygame, later ported to the web using Grok for real-time browser demos. AI Core: Reinforcement learning models trained to optimize flight paths, collision avoidance, and gate navigation. Interface: A Gradio dashboard for controlling race parameters, viewing telemetry, and replaying race footage. Visualization: Integrated live race cam and heatmap tracking for swarm coordination insights.
Challenges we ran into
Managing 50 independent drone simulations in real-time without frame drops. Designing RL reward functions that balance speed with safety. Handling collision physics and gate detection under web deployment constraints. Synchronizing data between Grok’s browser simulation and local AI training environments.
Accomplishments that we're proud of
Achieved stable multi-drone racing at 60 FPS in browser. Reduced simulated crash rates by 70% after RL fine-tuning. Developed a plug-and-play export system to Haas drones for real-world testing. Built a visually stunning “Track 3” demo that wowed testers with swarm precision.
What we learned
How to apply multi-agent reinforcement learning in high-speed environments. The importance of balancing compute efficiency with visual realism. That collaboration between drones — not just competition — creates better AI behavior. Browser ports of traditionally local simulations can be surprisingly powerful when optimized.
What's next for DroneF1 Swarm
Integrate real sensor data from physical Haas drones for hybrid training loops. Add dynamic weather and obstacle conditions for realism. Explore team-based swarm strategies (e.g., drafting, blocking, pit strategies). Open-source the simulator to the AI racing community. Host the first-ever DroneF1 Virtual Grand Prix online.
Log in or sign up for Devpost to join the conversation.