AERO-MPPI: Anchor-Guided Ensemble Trajectory Optimization for Agile Mapless Drone Navigation

Abstract

Agile mapless navigation in cluttered 3D environments poses significant challenges for autonomous drones. Conventional mapping-planning-control pipelines incur high computational cost and propagate estimation errors. We present AERO-MPPI, a fully GPU-accelerated framework that unifies perception and planning through an anchor-guided ensemble of Model Predictive Path Integral (MPPI) optimizers. Specifically, we design a multi-resolution LiDAR point-cloud representation that rapidly extracts spatially distributed "anchors" as look-ahead intermediate endpoints, from which we construct polynomial trajectory guides to explore distinct homotopy path classes. At each planning step, we run multiple MPPI instances in parallel and evaluate them with a two-stage multi-objective cost that balances collision avoidance and goal reaching. Implemented entirely with NVIDIA Warp GPU kernels, AERO-MPPI achieves real-time onboard operation and mitigates the local-minima failures of single-MPPI approaches. Extensive simulations in forests, verticals, and inclines demonstrate sustained reliable flight above 7 m/s, with success rates above 80% and smoother trajectories compared to state-of-the-art baselines. Real-world experiments on a LiDAR-equipped quadrotor with NVIDIA Jetson Orin NX 16G confirm that AERO-MPPI runs in real time onboard and consistently achieves safe, agile, and robust flight in complex cluttered environments. Code is available at https://github.com/XinChen-stars/AERO_MPPI.

Figures (7)

• Figure 1: AERO-MPPI navigating a challenging obstacle environment. The algorithm samples spread anchor points and generates corresponding guiding trajectories (gray lines) on the LiDAR point cloud. These trajectories are then optimized by parallel MPPI instances to produce multiple trajectory proposals (orange–purple curves). The best is the one with the lowest cost.
• Figure 2: AERO-MPPI System Overview. We perform state estimation by fusing IMU data with raw LiDAR point clouds to recover the system state $\left(\mathbf{p}, \mathbf{v}, \mathbf{q}\right)$. To enhance efficiency, we design a multi-resolution partitioning scheme that downsamples the nearest point clouds onto coarsely defined partitions to obtain refined anchor endpoints. For free-space exploration, the drone leverages the refined endpoints together with the current state to compute guiding trajectories, which serve as initial guesses for parallel multi-MPPI optimization. After obtaining the candidate control sequences from multi-MPPI, we evaluate them through explicit rollouts and select the best control input.
• Figure 3: Multi-resolution partition for generating guiding trajectories. Black points denote raw point clouds within the sensing sphere. The left hemisphere illustrates fine-grained cells, where selected cells are highlighted by the corresponding red spherical cones extending to the nearest LiDAR points (yellow). The right hemisphere depicts a coarse-grained partition used to adjust the initial anchor endpoints (gray) to refined endpoints (blue). Facilitated by this multi-resolution partition, guiding trajectories are generated efficiently from the current drone's pose to the refined endpoints.
• Figure 4: Comparison across obstacle numbers (200–1000) and height ranges (1–6 m) in a $40\,\mathrm{m} \times 40\,\mathrm{m}$ map. (a) Randomized heights between 1–6 m: the five subfigures correspond to obstacle counts from 200 to 1000. With 1000 obstacles (each with a width ranging from 0.4 m to 1.1 m), the maximum velocity reaches 8.20 m/s and the mean velocity is 5.42 m/s, demonstrating the agility of the proposed AERO-MPPI. (b) Fixed height of 6 m: the uniform obstacle height prevents the drone from exploiting $z$-axis variation for avoidance. This configuration is more challenging, with narrow gaps that often cause local minima, reducing the maximum velocity to 7.68 m/s.
• Figure 5: Benchmark results with previous works egoplanner2021topoplanner2020liu2024integratedfastplanner2019robust under different scenarios in a $40\,\mathrm{m} \times 40\,\mathrm{m}$ map. (a) Forest: All methods generate smooth trajectories of approximately the same length. (b) Verticals: Mapping-based methods exhibit frequent replanning, leading to oscillatory and non-smooth trajectories, as tall vertical obstacles often cause severe occlusions. In contrast, our method directly leverages local perception for optimization, ensuring continuity and smoothness. (c) Inclines: Similar to the Verticals case, frequent replanning occurs, and some trajectories even become trapped in local geometric pitfalls (e.g., small holes or narrow gaps), causing stagnation at local optima. Our method employs explicit parallel optimization and robust sampling to effectively avoid such traps.