EAROL: UAV Autonomy with Tilted LiDAR

• EAROL is a UAV autonomy framework that uses a downward-mounted 20° tilted LiDAR and tightly-coupled LiDAR-Inertial Odometry to overcome weak overhead structures in open-top environments.
• It employs degeneracy-aware mapping with adaptive voxel filtering and hierarchical trajectory-yaw optimization to balance environmental exploration, target tracking, and energy efficiency.
• Experimental evaluations demonstrate an 81% reduction in tracking error, near-zero vertical drift, and a 22% improvement in perceptual coverage compared to conventional approaches.

EAROL is a unified UAV autonomy framework for Environmental Augmented perception-aware planning and Robust Odometry via Downward-Mounted tilted LiDAR, proposed for open-top scenarios such as collapsed buildings and roofless mazes, where conventional LiDAR-SLAM pipelines can suffer from weak overhead structure, localization drift, and poor coupling between perception and planning. The framework combines a downward-mounted tilted LiDAR configuration with a tightly-coupled LiDAR-Inertial Odometry (LIO) system and a hierarchical trajectory-yaw optimization algorithm, with the stated objective of improving 6-DoF localization, forward environmental awareness, target tracking, perceptual coverage, and energy efficiency in post-disaster search-and-rescue settings (Liang et al., 20 Aug 2025).

1. Problem setting and system definition

EAROL is designed for UAV operation in feature-sparse and degenerate environments in which a large fraction of the observable structure lies below or ahead of the vehicle rather than above it. The motivating failure mode is that standard horizontal or upward-facing LiDAR configurations often encounter unobservable vertical states in open environments, which leads to localization drift, especially along the vertical axis. A second motivating issue is perception-planning coupling: when sensing is directional, trajectory generation and yaw control cannot be treated as independent subproblems without degrading coverage or target observability (Liang et al., 20 Aug 2025).

The framework is presented as a hardware-algorithm co-design paradigm. Its hardware innovation is a downward-mounted tilted LiDAR with 20° inclination. Its algorithmic stack contains two main components. The first is a tightly-coupled LIO system based on an Iterative Error-State Kalman Filter (IESKF) with dynamic motion compensation, intended to maintain robust localization in open-top scenes. The second is a hierarchical trajectory-yaw optimization algorithm that explicitly balances environmental exploration, target tracking precision, and energy efficiency (Liang et al., 20 Aug 2025).

A plausible implication is that EAROL treats observability, mapping, and control as a single coupled design problem rather than as separate modules connected only through a pose estimate.

2. Downward-mounted tilted LiDAR and constraint enhancement

The defining hardware choice in EAROL is a LiDAR mounted downward and tilted forward by 20°. According to the reported design rationale, this geometry allows the field of view to cover both the ground below and the space ahead. The ground-facing component yields dense ground point cloud acquisition, which is used for constraint enhancement, especially for vertical observability. The forward-facing component provides forward environmental awareness for dynamic obstacle detection and target-oriented navigation (Liang et al., 20 Aug 2025).

This sensor pose is intended to overcome a specific limitation of conventional UAV LiDAR layouts in open-top environments: the lack of overhead geometric features. By emphasizing the ground plane and forward scene, EAROL increases the availability of geometric constraints that remain observable even when ceilings or other upper-hemisphere structures are absent. The reported outcome is near-zero vertical drift, which the framework attributes to the joint effect of the tilted LiDAR geometry and degeneracy-aware state estimation (Liang et al., 20 Aug 2025).

The system also includes adaptive tilt compensation to correct installation error at startup using IMU data. With IMU-measured accelerations $a_x$ and $a_z$, the pitch is computed as

$$pitch = \arctan\left(\frac{a_x^{filtered}}{a_z^{filtered}}\right)$$

with low-pass filtering applied to smooth the readings. This startup calibration is described as a countermeasure against mechanical inaccuracies in sensor installation (Liang et al., 20 Aug 2025).

3. Tightly-coupled LiDAR-Inertial Odometry

EAROL’s odometry subsystem tightly integrates LiDAR and IMU measurements, with the reported sensor rates of 10 Hz for LiDAR and 200 Hz for the IMU. The estimator is an IESKF operating on a state vector

$$\xi = [\mathbf{t}_W^I, \phi_W^I, \mathbf{v}_W^I, \beta_a, \beta_g, \gamma_W]^\top$$

where $\mathbf{t}_W^I$, $\phi_W^I$, and $\mathbf{v}_W^I$ are IMU position, orientation, and velocity in the world frame; $\beta_a$ and $\beta_g$ are accelerometer and gyroscope biases; and $\gamma_W$ is a gravity correction term (Liang et al., 20 Aug 2025).

A central component is dynamic motion compensation for point-cloud undistortion. The method uses BÉzier interpolation of angular velocity and acceleration; for angular velocity,

$a_z$0

Each point $a_z$1 is then transformed to an undistorted position as

$a_z$2

The update step uses the linearized observation model

$a_z$3

with iterative corrections mapped back to the rotation manifold. Residuals are computed as weighted Mahalanobis point-to-plane distances,

$a_z$4

These details indicate that the LIO module is not merely loosely fused sensor processing but a manifold-based state estimator with explicit scan undistortion and geometric residual modeling (Liang et al., 20 Aug 2025).

EAROL also includes dual-layer voxel hash tables with coarse and fine layers for local map management, and a relocalization mechanism based on spherical harmonic encoding of local map regions combined with pose graph optimization (Liang et al., 20 Aug 2025).

4. Degeneracy-aware mapping and hierarchical planning

A distinguishing feature of the odometry pipeline is explicit degeneracy-aware processing. The system quantifies local geometric constraint quality using entropy computed from PCA eigenvalues $a_z$5 of point-cloud covariance:

$a_z$6

Low $a_z$7 is interpreted as indicating degenerate or underconstrained directions; these directions are then damped during the state update via projection matrices. EAROL additionally uses adaptive voxel filtering, with voxel size updated as

$a_z$8

to maintain map density under changing scene structure (Liang et al., 20 Aug 2025).

On the planning side, EAROL adopts a hierarchical trajectory-yaw optimization algorithm because directional sensing makes tangent-aligned yaw insufficient. The trajectory $a_z$9 is represented as a sequence of polynomial segments with objective

$$pitch = \arctan\left(\frac{a_x^{filtered}}{a_z^{filtered}}\right)$$0

where the cost terms include $$pitch = \arctan\left(\frac{a_x^{filtered}}{a_z^{filtered}}\right)$$1 for smoothness, $$pitch = \arctan\left(\frac{a_x^{filtered}}{a_z^{filtered}}\right)$$2 for dynamic feasibility, $$pitch = \arctan\left(\frac{a_x^{filtered}}{a_z^{filtered}}\right)$$3 for obstacle avoidance, and $$pitch = \arctan\left(\frac{a_x^{filtered}}{a_z^{filtered}}\right)$$4 for total flight time. The smoothness term is exemplified by

$$pitch = \arctan\left(\frac{a_x^{filtered}}{a_z^{filtered}}\right)$$5

Yaw optimization is then performed on a discretized horizon using graph search (Dijkstra) over sampled candidate yaw angles. The node cost is

$$pitch = \arctan\left(\frac{a_x^{filtered}}{a_z^{filtered}}\right)$$6

where $$pitch = \arctan\left(\frac{a_x^{filtered}}{a_z^{filtered}}\right)$$7 is local map entropy, $$pitch = \arctan\left(\frac{a_x^{filtered}}{a_z^{filtered}}\right)$$8 is the normalized angular deviation of the tracked target from the field-of-view center,

$$pitch = \arctan\left(\frac{a_x^{filtered}}{a_z^{filtered}}\right)$$9

and $$\xi = [\mathbf{t}_W^I, \phi_W^I, \mathbf{v}_W^I, \beta_a, \beta_g, \gamma_W]^\top$$0 penalizes yaw changes,

$$\xi = [\mathbf{t}_W^I, \phi_W^I, \mathbf{v}_W^I, \beta_a, \beta_g, \gamma_W]^\top$$1

The entropy term for grid cell $$\xi = [\mathbf{t}_W^I, \phi_W^I, \mathbf{v}_W^I, \beta_a, \beta_g, \gamma_W]^\top$$2 is

$$\xi = [\mathbf{t}_W^I, \phi_W^I, \mathbf{v}_W^I, \beta_a, \beta_g, \gamma_W]^\top$$3

Graph pruning restricts yaw transitions to those consistent with feasible maximum angular velocities. The stated result is synchronized, collision-free trajectory and yaw generation that seeks to maximize perceptual coverage, keep moving targets within the LiDAR field of view, and reduce unnecessary rotation energy (Liang et al., 20 Aug 2025).

5. Experimental evaluation

The reported evaluation includes physical experiments in both indoor and outdoor environments. The indoor setting is an open-top maze with no ceiling, sized 17×12 m with a 12 m corridor. The outdoor setting is a 60×60 m playground containing dense vegetation, containers, mazes, and dynamic obstacles. A further experiment addresses dynamic tracking in a maze, in which the UAV tracks a moving target described as a ring on UGV (Liang et al., 20 Aug 2025).

In indoor mapping, EAROL is reported to produce accurate, crisp outline maps with no discernible drift, whereas FAST-LIO2, used as a standard LIO baseline with a conventional LiDAR mount, exhibits significant drift and boundary errors. In outdoor mapping, the framework is reported to reconstruct accurate, high-fidelity maps over large areas while negotiating obstacles. In the dynamic tracking experiment, the yaw planner keeps the target within the field of view with $$\xi = [\mathbf{t}_W^I, \phi_W^I, \mathbf{v}_W^I, \beta_a, \beta_g, \gamma_W]^\top$$4 error, while a standard (differential) yaw control reaches up to $$\xi = [\mathbf{t}_W^I, \phi_W^I, \mathbf{v}_W^I, \beta_a, \beta_g, \gamma_W]^\top$$5 deviation (Liang et al., 20 Aug 2025).

The quantitative results reported for EAROL are:

• 81% reduction in average cumulative tracking error
• 73% reduction in instantaneous maximum error compared to baseline
• 22% improvement in global map entropy, used as a measure of explored/unexplored area
• Near-zero vertical drift

The abstract summarizes these results as 81% tracking error reduction, 22% improvement in perceptual coverage, and near-zero vertical drift across indoor maze and 60-meter-scale outdoor scenarios (Liang et al., 20 Aug 2025).

6. Significance, applications, and nomenclature

EAROL is positioned as a solution for post-disaster search and rescue missions, especially where UAVs must operate in structurally open, partially collapsed, or otherwise geometrically degenerate environments. The framework’s significance lies in the claim that robust autonomy in such settings requires simultaneous attention to sensor placement, odometry robustness, and perception-aware planning, rather than software-only modification of a conventional sensor stack (Liang et al., 20 Aug 2025).

The paper explicitly states that the authors will release the software and hardware as an open-source package. This suggests an intention to make the system reproducible at both the algorithmic and platform levels, though the operational impact of that release necessarily depends on downstream adoption (Liang et al., 20 Aug 2025).

Because acronym collisions exist in recent literature, EAROL should be distinguished from unrelated systems using similar names, including EARL for egocentric interaction reasoning and pixel grounding (Su et al., 14 May 2026) and EARL for energy-aware adaptive antenna control in O-RAN cell-free massive MIMO networks (Ge et al., 13 Feb 2026). Within UAV autonomy, however, EAROL specifically denotes the downward-mounted tilted-LiDAR framework for environmental augmented perception-aware planning and robust odometry (Liang et al., 20 Aug 2025).