Core Concepts
Quadrotor trajectory planning with a focus on perception-awareness to enhance state estimation accuracy.
Abstract
I. Introduction
Various perception-aware planning approaches aim to improve state estimation accuracy.
Feature matchability is crucial but often overlooked in trajectory planning.
APACE framework focuses on generating perception-aware trajectories for quadrotors.
Objective: Reduce visual-based estimator error while ensuring smoothness, safety, agility, and dynamics constraints.
II. Related Work
Studies bridge perception and planning using metrics like Fisher information.
Existing works overlook feature matchability or consider it as a soft constraint.
Approaches differ in optimizing execution trajectories or considering clustered features.
III. System Overview
Position and yaw trajectories are optimized separately using B-spline curves.
Differential flatness property of quadrotors enables trajectory decomposition.
Visibility model maximizes covisible features while maintaining small parallax angles.
IV. Methodology
A. Visibility Model
Differentiable visibility model approximates exact visibility function accurately.
B. Position Trajectory Optimization
Cost function includes vertical covisibility and parallax cost terms.
C. Yaw Trajectory Generation
Primitives Search: Graph search maximizes covisibility along the trajectory.
Trajectory Optimization: Minimizes jerk and follows yaw primitives guidance.
V. Experiments
A. Experimental Setup
Simulation experiments conducted in AirSim with comparisons to existing methods.
B. Photorealistic Simulation Benchmarks
Our method outperforms perception-agnostic and existing perception-aware planners in simulation benchmarks.
C. Real-world Experiments
Real-world experiments demonstrate the feasibility of our method in challenging low-texture environments.
VI. Conclusions
APACE framework significantly improves state estimation accuracy for aggressive quadrotor flights through perception-aware trajectory generation.
Stats
"RMSE reduced by up to an order of magnitude."
"Average goal error of 0.54m and RMSE of 0.40m."