The Drone Dreams the Gate but Simulates Its Own Body

Agile flight through moving gates requires predicting the environment without losing high-frequency control of the vehicle. DiffDreamer separates those jobs within one reinforcement-learning update. A learned latent world model propagates external motion and history-dependent features; a differentiable simulator propagates the quadrotor's physical state. The actor objective also separates critic-driven task rewards from differentiable motion rewards. On aggressive tracking and dynamic-gate racing, the authors report faster convergence and stronger final performance than model-free, model-based, differentiable-RL and classical planning-and-control comparisons, plus proof-of-concept real flights. The abstract provides no universal margin, and simulation structure, reward design and selected tasks constrain the claim.