New AI cuts quadrotor formation errors by 31%, runs 10x faster
Physics-informed learning enables tight drone flight with just 30 seconds of training data
Quadrotors flying in tight formations face severe aerodynamic turbulence—especially downwash from neighboring drones—that can cause collisions. Traditional model-based controllers struggle to model these complex interactions, while nonlinear model predictive control (NMPC) offers high accuracy but demands heavy computation, limiting real-time use on small drones.
To solve this, the team introduces a physics-informed residual learning framework that models aerodynamic disturbances while keeping the multi-quadrotor system differentially flat. This preserved flatness allows a computationally cheap feedback linearization controller to cancel disturbances via feedforward compensation. In hardware experiments, their approach reduced average tracking errors by 31% compared to standard baselines. Crucially, it matched the tracking performance of state-of-the-art NMPC while requiring an order of magnitude less computation. The system can achieve stable, tight formation flight with under 30 seconds of training data and a 5ms loop rate—making it viable for compute-constrained flight stacks. The work was accepted at IROS '26.
- 31% reduction in average tracking errors vs. nominal baselines in real hardware tests
- Matches NMPC performance but requires 10x less computation
- Achieves stable tight formation with under 30 seconds of training data and a 5ms control loop rate
Why It Matters
Enables tight formation flight on low-cost drones, unlocking swarm applications previously limited by compute constraints.