FALCON: A Fast and Feasible End-to-End Trajectory Planner via a Cascaded Bernstein Decoder

Abstract

Navigation in cluttered environments with dynamic obstacles remains a major challenge for uncrewed aerial vehicles (UAVs), especially when both fast response and reliable trajectory generation are required. Traditional modular planners suffer from the latency introduced by mapping and iterative optimization, whereas end-to-end methods, despite their efficient inference, often lack explicit feasibility constraints on the generated trajectories. To address these limitations, we propose a self-supervised end-to-end UAV planning framework that combines Bernstein polynomial-based feasible trajectory generation with point-cloud-based differentiable supervision for collision-aware learning. Specifically, the perception network predicts unconstrained latent variables, which are subsequently transformed into executable trajectories by a carefully designed differentiable cascaded decoder. By leveraging the properties of Bernstein polynomials, the decoder generates trajectories that naturally satisfy velocity and acceleration constraints, thus requiring no back-end optimization. Meanwhile, the supervision module uses raw point clouds as privileged information and extracts differentiable distance and gradient cues through local quadratic surface fitting. Because local quadratic surfaces can be fitted efficiently online as the scene changes, this approach enables training in dynamic environments. Extensive simulation and real-world experiments validate the effectiveness, real-time performance, and zero-shot sim-to-real capability of the proposed method. Note to Practitioners—In indoor inspection, warehouse navigation, emergency response, and other cluttered workspace applications, practitioners often need small aerial robots to react quickly to nearby static and moving obstacles while producing motions that can be reliably executed by the vehicle. For practical deployment, a local planner is expected to operate with limited onboard computation, handle changes in the surrounding environment, and output trajectories that are compatible with downstream tracking controllers. This work presents an end-to-end local planning framework that directly generates short-horizon trajectories from onboard depth observations and vehicle state measurements. By training the planner in changing simulated scenes with moving obstacles and point-cloud-based collision supervision, the system learns to generate collision-aware motions under moving-obstacle interference. The proposed system is designed to provide low-latency local obstacle avoidance while keeping the generated trajectories within prescribed velocity and acceleration limits. This property can facilitate integration with downstream tracking controllers and reduce the risk of infeasible motion commands. During deployment, the planner only requires onboard depth sensing and state measurements, making it suitable for resource-constrained aerial platforms. The current implementation is intended for local planning and does not replace global mission planning, state estimation, or low-level control. Its performance may depend on sensing quality and training-scene coverage. Future extensions may include stronger uncertainty handling, richer dynamic-obstacle prediction, and multi-robot coordination. The same idea may also be useful for other automated mobile platforms that require fast and feasible local motion generation.