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

1. Problem Statement

The paper addresses a critical challenge in Human-in-the-Loop (HITL) UAV operations within complex, safety-critical environments (e.g., underground mines, disaster zones, dense forests).

• Hardware Constraints: Resource-constrained UAVs often rely on lightweight, solid-state LiDARs with Narrow Fields of View (FoV) due to payload and power limitations.
• The Degeneracy Issue: In geometrically degenerate or feature-sparse scenes (e.g., long tunnels, corridors), the limited FoV severely weakens the observability of LiDAR-Inertial Odometry (LIO). This leads to drift accumulation, unstable state estimation, and degraded geometric accuracy.
• The Control Gap: Existing solutions either add heavy mechanical gimbals (reducing flight endurance) or use passive rotation strategies that lack adaptability. Furthermore, current active perception methods often treat the UAV as fully autonomous, failing to safely integrate human operator intent with autonomous yaw optimization.
• Core Challenge: How to maintain high-accuracy, robust LIO in degenerate environments using only the UAV's existing agility (yaw rotation) without adding mechanical complexity, while ensuring safety and respecting human pilot commands.

2. Methodology: The AWARE Framework

The authors propose AWARE, a bio-inspired framework that treats the UAV's whole-body yaw motion as a "gimbal-like" sensing action. The system operates on a Hybrid RL-MPC architecture.

A. System Architecture & Perception

• Unified Panoramic Representation: The system aggregates recent LiDAR scans into a ScanBlock and projects them onto a 2D Panoramic Depth Map (spherical coordinate system). This creates a unified view of the environment relative to the UAV, preserving geometric features while reducing dimensionality.
• Predictive Observability Analysis: Using the panoramic map, the system simulates virtual scans for candidate yaw angles. It computes the Fisher Information Matrix (FIM) to derive an observability metric ($\mathcal{J}_{obs}$). This metric predicts which yaw direction minimizes pose uncertainty (maximizes information gain) based on the Fisher Information Matrix trace.

B. Hybrid Control Strategy (RL + MPC)

To balance observability (sensing quality) with flight stability and human intent, AWARE uses a two-layer control loop:

1. Model Predictive Control (MPC) Layer:

   • Formulates a constrained Quadratic Programming (QP) problem.
   • Task Decoupling: The human pilot controls translational motion ($x, y, z$), while the MPC autonomously optimizes the yaw angle ($\psi$) to maximize observability.
   • Safe Flight Corridor (SFC): Constraints are applied to ensure the trajectory remains within a collision-free convex region, decoupling safety from the active sensing optimization.
   • Differentiable Approximation: Since the observability metric is non-differentiable, the system fits a local quadratic function to the observability landscape to integrate it into the MPC cost function.

2. Reinforcement Learning (RL) Meta-Controller:

   • Role: A lightweight RL agent (Actor-Critic network trained with PPO) acts as a meta-controller.
   • Input: It takes the UAV's kinematic state and a downsampled version of the panoramic depth map as input.
   • Output: It dynamically adjusts the cost weights of the MPC (specifically the weight $\lambda_{obs}$ for observability vs. tracking/stability weights) in real-time.
   • Benefit: This allows the system to adapt to changing environmental contexts (e.g., increasing yaw rotation in tunnels, reducing it in forests) without manual parameter tuning.

C. Training Pipeline

• Sim-to-Real: The system is trained in a high-fidelity, point-cloud-based simulation environment using real-world maps (tunnels, forests, urban structures).
• Reward Function: The RL agent is rewarded for SLAM accuracy (minimizing Relative Trajectory Error), exploration completeness, and control smoothness.
• Human-in-the-Loop: Training uses expert pilot trajectories as reference paths, ensuring the agent learns to optimize sensing while following human intent.

3. Key Contributions

1. Adaptive Whole-Body Active Rotation: A novel framework that exploits the UAV's inherent yaw agility to extend the effective sensor horizon, eliminating the need for heavy mechanical gimbals.
2. Hybrid RL-MPC Architecture: A unique integration where an RL agent dynamically tunes MPC cost weights based on scene geometry. This resolves the trade-off between sensing gain and flight stability, outperforming static optimization baselines.
3. HITL-Safe Control: The introduction of a Safe Flight Corridor (SFC) mechanism that decouples the operator's navigational intent from autonomous yaw optimization, ensuring safety while allowing the system to actively improve localization.
4. High-Fidelity Simulation Platform: Development of a closed-loop, point-cloud-based simulator that supports real-to-sim-to-real transfer, validated across diverse structured, degraded, and unstructured environments.

4. Experimental Results

The framework was validated in extensive simulations and real-world flights on a custom quadrotor with an edge-grade processor (Intel N305).

• Simulation Performance:

  • AWARE achieved an average Absolute Pose Error (APE) of 3.41 m across diverse scenes.
  • This represents a 27.8% improvement over the best static MPC baseline and an 84.9% improvement over passive fixed-rate scanning.
  • In highly degenerate environments (e.g., Istanbul Tunnel), passive methods failed catastrophically (APE > 100m), while AWARE maintained bounded errors (~17.86m).

• Real-World Deployment:

  • Tested in four challenging sites: Underground Tunnel, Abandoned Bunker, Abandoned Athletic Infrastructure, and Dense Forest.
  • Generalization: The policy trained entirely in simulation transferred successfully to real-world environments without fine-tuning.
  • Drift Rates: AWARE maintained low drift rates (e.g., 0.05% in open areas, 0.66% in the highly degenerate tunnel), demonstrating robustness against noise, dust, and occlusion.
  • Real-Time Performance: The control loop runs at an average latency of 46.90 ms (95th percentile < 81 ms) on the onboard processor, satisfying the 10 Hz LiDAR update rate.

• Interpretability: Analysis showed that the RL agent correctly increased the observability weight ($\lambda_{obs}$) and commanded higher yaw rates in feature-poor tunnels, while relaxing these actions in feature-rich forests, confirming the system learns a scene-aware strategy.

5. Significance

• Operational Safety: By decoupling human intent from autonomous sensing optimization via SFC, AWARE enables safer HITL operations in critical scenarios where localization failure could be fatal.
• Hardware Efficiency: It proves that software-defined active perception can replace hardware-heavy solutions, making high-precision surveying feasible on lightweight, low-cost UAVs.
• Robustness: The ability to maintain accurate state estimation in geometrically degenerate environments (tunnels, caves) significantly expands the operational envelope of autonomous and semi-autonomous aerial systems.
• Scalability: The real-time performance and sim-to-real transfer capability suggest this approach is ready for deployment in real-world inspection, search and rescue, and mapping missions.