Real-Time Visual Obstacle-Avoidance Trajectory Generation for UAVs Using Imitation Learning and Privileged Geometry-Guided Residual Reinforcement Learning

Abstract

Real-time flight in unknown cluttered environments requires unmanned aerial vehicle (UAV) planners to generate local trajectories from limited onboard visual observations within short decision cycles, while balancing obstacle avoidance and progress toward a goal. Vision-based imitation-learning planners can acquire useful planning priors from expert demonstrations, but their performance is constrained by demonstration coverage and may deteriorate under closed-loop distribution shifts caused by unseen environments, perception errors, and trajectory-tracking deviations. This degradation can reduce obstacle clearance or impede progress. We introduce a hierarchical trajectory-generation method that combines behavior cloning with privileged geometry-guided residual reinforcement learning. A privileged expert planner generates demonstrations through kinodynamic search, Euclidean signed distance field (ESDF)-constrained cubic B-spline optimization, and temporal adjustment, and uses the optimized free control points as supervision. The behavior-cloning policy predicts nominal control points from depth observations, vehicle states, and relative-goal information. A residual policy then directly modifies these control points. Its action scale is tied to the local planning range, allowing substantial restructuring when necessary while retaining the expert planning prior. The fused control points define a time-parameterized cubic B-spline trajectory. During training, the ESDF provides privileged clearance rewards; at deployment, the policy uses only onboard depth, flight state, and relative-goal information.