MINER-RRT*: A Hierarchical and Fast Trajectory Planning Framework in 3D Cluttered Environments

TL;DR

MINER-RRT* tackles fast and reliable quadrotor trajectory planning in 3D clutter by a hierarchical approach that couples a neural-heuristic front-end with a differential-flatness-based back-end. The front-end uses 3D-sGAN-RRT* and a 3D conditional sGAN to produce a safe, connected heuristic region that bias-samples the RRT* search, yielding high-quality initial paths quickly. The back-end then applies minimum control effort trajectory generation with closed-form, time-allocated polynomial segments derived from optimality conditions, ensuring collision-free, smooth multi-stage trajectories. Experiments in simulation and real indoor environments show substantial speedups and improved connectivity/safety over SOTA methods, validating the framework’s practicality for fast, high-quality 3D planning in cluttered spaces.

Abstract

Trajectory planning for quadrotors in cluttered environments has been challenging in recent years. While many trajectory planning frameworks have been successful, there still exists potential for improvements, particularly in enhancing the speed of generating efficient trajectories. In this paper, we present a novel hierarchical trajectory planning framework to reduce computational time and memory usage called MINER-RRT*, which consists of two main components. First, we propose a sampling-based path planning method boosted by neural networks, where the predicted heuristic region accelerates the convergence of rapidly-exploring random trees. Second, we utilize the optimal conditions derived from the quadrotor's differential flatness properties to construct polynomial trajectories that minimize control effort in multiple stages. Extensive simulation and real-world experimental results demonstrate that, compared to several state-of-the-art (SOTA) approaches, our method can generate high-quality trajectories with better performance in 3D cluttered environments.

Figures (11)

• Figure 1: Illustration of our hierarchical trajectory planning framework MINER-RRT*. (a) Front end: The predicted heuristic region is shown in orange, the edges of the RRT are shown in light red, and the generated initial path is shown in green. (b) Back end: The resulting minimum snap trajectory is shown in light blue.
• Figure 2: Architecture of our system. The localization module and controller are on both sides of the figure (Sec. \ref{['sec:overview']}). The front end of our framework is the initial path finding guided by the heuristic map (Sec. \ref{['sec:Frond-end']}). The back end of our framework is the optimal trajectory generation (Sec. \ref{['sec:Back-end']}).
• Figure 3: Architecture of the proposed network.
• Figure 4: The visualization of trajectory planning process: original map, RRT*'s front-end path, heuristic region, our front-end path, our final trajectory (Simulation map1).
• Figure 5: The visualization of trajectory planning process: original map, RRT*'s front-end path, heuristic region, our front-end path, our final trajectory (Simulation map2).

Theorems & Definitions (2)

• proof
• proof