New self-adaptive learning method for drone tracking achieves no-regret control
A self-adaptive control method learns unknown target dynamics with near-optimal performance even under adversarial motion
A self-adaptive online learning for control method is proposed for tracking unknown target dynamics that may switch between structured, random, and adversarial motion. The method simultaneously learns multiple predictors from scratch via self-supervised, one-shot, and computationally efficient learning, adaptively selecting the best one to match observed behavior. It provides finite-time near-optimality guarantees in expectation, asymptotically matching the optimal non-causal control policy that knows the target dynamics a priori. The method is validated in Crazyflie simulations and hardware experiments across structured, random, and adversarial target trajectories, in comparison with non-stochastic, kernel-based, and neural-network-based methods.
- Simultaneously learns multiple predictors from scratch using self-supervised one-shot learning, avoiding the need for labeled data or pre-training.
- Provides finite-time near-optimality guarantees with regret proportional to learning error and switching frequency, asymptotically matching the optimal non-causal policy.
- Validated on Crazyflie quadrotors in both simulation and hardware experiments, outperforming RFF, kernel, and neural network baselines under structured, random, and adversarial target motion.
Why It Matters
Enables drones and robots to robustly track unknown moving targets without prior knowledge, critical for autonomous navigation and pursuit-evasion in dynamic environments.