In-Situ Learning in Drones: Rethinking Embodied Intelligence Without Sim-to-Real

This article introduces in-situ learning (ISL), a paradigm in which drones adapt directly through real-world interaction rather than relying on simulation-to-reality (sim-to-real) transfer. Whereas sim-to-real separates training from embodiment, ISL integrates the physical interaction features into the learning loop, treating disturbances such as wind gusts, ground effect, and collisions not as nuisances but as informative signals. We formulate the in-situ learning problem for quadrotor UAVs and develop an adaptive reservoir critic with continuous-time weight updates. The framework admits formal guarantees of boundedness, convergence, and input-to-state stability, and introduces interaction-gated adaptation, where natural disturbances provide the persistent excitation required for learning. Experimental studies via both simulations and a real-world quadrotor demonstrate emergent behaviors such as disturbance anticipation, collision-driven adaptation, and context-sensitive thrust regulation. These results support a philosophical shift: robustness and intelligence emerge not from increasingly complex simulators but from embodied interaction itself. We argue that ISL reframes physical interaction as a renewable computational resource and motivates new benchmarks and design principles for embodied intelligence beyond the sim-to-real paradigm.