AWARE: Adaptive Whole-body Active Rotating Control
for Enhanced LiDAR-Inertial Odometry under Human-in-the-Loop Interaction

Yizhe Zhang1, Jianping Li2*, Liangliang Yin1, Zhen Dong1, Bisheng Yang1
1State Key Laboratory of Information Engineering in Surveying, Mapping and Remote Sensing, Wuhan University, Wuhan 430079, China
2School of Electrical and Electronic Engineering, Nanyang Technological University, 639798, Singapore
*Corresponding author
arXiv Video Code

Project Video

Video

AWARE is a bio-inspired active sensing framework designed for human-in-the-loop UAV operation in challenging real-world environments. By combining whole-body active yawing with hybrid adaptive RL-MPC control, it expands LiDAR perception, improves localization robustness, and supports safe real-time flight.

Research Overview

Abstract

Human-in-the-loop (HITL) UAV operation in complex and safety-critical environments demands reliable onboard localization, yet lightweight platforms equipped with narrow-field-of-view LiDAR often suffer from limited sensing coverage. As a result, LiDAR-inertial odometry can become fragile in geometrically degenerate and feature-sparse scenes.

AWARE addresses this challenge with a bio-inspired whole-body active yawing strategy that expands the effective sensing horizon without additional actuation. The framework couples differentiable Model Predictive Control (MPC) with Reinforcement Learning (RL): MPC selects observability-aware yaw actions, RL adapts control weights online to balance information gain and flight stability, and a Safe Flight Corridor preserves operator intent for safe human-autonomy cooperation. Across both simulated and real-world experiments, AWARE consistently surpasses passive scanning and static optimization baselines while maintaining real-time performance on resource-constrained UAV hardware. These properties make AWARE particularly valuable for UAV-based surveying and mapping, where robust localization is essential for reliable geospatial data acquisition in complex field environments.

Key Features & Contributions

Bio-inspired Active Yawing

Inspired by biological active sensing, AWARE uses whole-body yaw rotation to expand the effective LiDAR sensing horizon and improve perception quality without extra mechanical actuation.

Hybrid RL + MPC

A differentiable MPC searches observability-aware yaw actions, while a lightweight RL policy adapts control weights online to balance information gain and flight stability.

Human-autonomy Cooperation

A Safe Flight Corridor preserves operator navigation intent while decoupling autonomous yaw optimization, enabling safe and effective human-in-the-loop interaction.

Real-time Performance

Extensive simulation and real-world experiments show robust localization gains over passive scanning baselines while maintaining onboard real-time performance.

Real-world Experiments

AWARE is designed for challenging real-world and human-in-the-loop UAV scenarios where sensing coverage, localization robustness, and safe cooperative control must all be maintained.

Field Testing Results

Abandoned Athletic Infrastructure

Underground Tunnel

Abandoned Bunker

Dense Forest

Learning & Practice in Simulation Environments

We build a point cloud-based simulation platform and dataset for scalable active sensing training. The simulator covers eight sequences across structured, degraded, and unstructured environments, and pairs dense point cloud maps with expert-demonstrated trajectories and randomized initialization for robust training and evaluation.

Simulator

Point-cloud active sensing platform

Supports PX4-style trajectory rollout, expert reference guidance, and policy learning under diverse geometry conditions.

Dataset

Eight curated training sequences

Scenarios span structured, degraded, and unstructured environments with different observability and occlusion characteristics.

Training Setup

Train and evaluation split

The first 50% of each reference trajectory is used for training, while the remaining 50% is reserved for evaluation.

Simulation Dataset & Training Setup

Overview of the simulation dataset and training configuration for AWARE.
Overview of the simulator, dataset organization, and trajectory-based training setup used for AWARE active sensing policy learning.

Eight Simulation Sequences

Seq01 Structured

Spine

952 m

Linear corridor with repetitive vertical columns.

Seq02 Structured

Wuhan Subway

344 m

Large-scale indoor scene with repetitive patterns.

Seq03 Structured

Building

690 m

Multi-level complex environment with partition walls.

Seq04 Degraded

Istanbul Tunnel

2314 m

Featureless tunnel with strong geometric degeneracy.

Seq05 Degraded

Wuhan Tunnel

378 m

Unexposed space dominated by linear structure.

Seq06 Unstructured

Lava Tube

292 m

Irregular curvature with severe occlusion.

Seq07 Unstructured

Forest

1153 m

High clutter and unstructured vegetation.

Seq08 Unstructured

Karst Cave

607 m

Complex topology with uneven terrain.

BibTeX

@misc{zhang2026awareadaptivewholebodyactive,
  title={AWARE: Adaptive Whole-body Active Rotating Control for Enhanced LiDAR-Inertial Odometry under Human-in-the-Loop Interaction},
  author={Yizhe Zhang and Jianping Li and Liangliang Yin and Zhen Dong and Bisheng Yang},
  year={2026},
  eprint={2604.10598},
  archivePrefix={arXiv},
  primaryClass={cs.RO},
  url={https://arxiv.org/abs/2604.10598},
}