Technical Summary: HUNT (High-Speed UAV Navigation and Tracking in Unstructured Environments via Instantaneous Relative Frames)

1. Problem Statement

Search-and-Rescue (SAR) operations require Unmanned Aerial Vehicles (UAVs) to perform two distinct but critical tasks: traversing unknown, unstructured environments at high speed to locate victims, and reliably tracking detected targets once identified. Current autonomy solutions face significant limitations in degraded sensing conditions (e.g., dense forests, urban canyons, collapsed buildings) where GPS is unavailable and visual features are sparse or dynamic.

• Global Navigation Limitations: Traditional approaches relying on GPS, Visual-Inertial Odometry (VIO), or SLAM degrade rapidly in these environments. Their dependence on persistent landmarks and loop closures leads to drift and unsafe behavior at high speeds.
• Relative Navigation Limitations: Recent object-centric methods successfully track targets by anchoring control to a detected object. However, these methods assume a target is visible from the start and remains continuously observable. They lack a mechanism for safe, high-speed traversal in unknown terrain when no target is in the field of view.

The core challenge addressed is achieving a unified framework that enables safe high-speed traversal without global localization and seamlessly transitions to target tracking without reverting to fragile global pose estimation.

2. Methodology

The proposed framework, HUNT, unifies traversal, acquisition, and tracking within a single instantaneous relative formulation. It operates entirely within a relative reference frame $R_t$ reconstructed at every control cycle from directly observable onboard quantities (attitude, altitude, instantaneous velocity, and relative target position). This eliminates dependence on unobservable states like horizontal translation or absolute yaw.

A. Operating Regimes

HUNT operates in two modes, managed by a unified perception-control pipeline:

1. Loitering Mode (Search):

   • Context: No target is detected with high confidence.
   • Reference Frame: $R_t$ is anchored to gravity and the initial heading, remaining fixed over time.
   • State Estimation: An Unscented Kalman Filter (UKF) estimates relative position, velocity, and attitude. While horizontal position drifts due to lack of direct observation, navigation remains robust because reference generation defines goals relative to the current state estimate.
   • Reference Generation: The UAV regulates altitude, forward speed, and heading. The heading maintains the initial yaw but can be weakly biased by low-confidence detections to steer toward potential targets without triggering a mode switch.
   • Navigation: The UAV traverses cluttered terrain by defining reference trajectories based on the current estimate, ensuring control objectives remain well-posed despite drift.

2. Tracking Mode (Pursuit):

   • Context: A target is detected with high confidence.
   • Reference Frame: $R_t$ is re-anchored at every cycle to the detected target's position. This keeps the target stationary in $R_t$ regardless of global drift.
   • State Estimation: The UKF updates the state using target-relative displacement derived from visual detection and stereo disparity. The linear acceleration is not used in the process model to avoid assumptions about target motion.
   • Reference Generation: The UAV maintains a safety standoff distance ($d_{xy}$) and vertical offset ($d_z$) while orienting the body axis toward the target.

B. Control Architecture

• Unified Controller: Both modes are executed by a single Nonlinear Model Predictive Controller (NMPC).
• Safety Constraints: High-order Control Barrier Functions (CBFs) are embedded directly into the NMPC. Obstacles are identified from stereo depth images, and a time-to-collision heuristic selects the most imminent collision points. The CBFs enforce a quadratic separation constraint, guaranteeing dynamically feasible, collision-free trajectories in real time.
• Mode Switching: A hysteresis policy based on the estimator's covariance trace ($tr(P_t)$) governs transitions.
  • Loiter $\to$ Track: Occurs when uncertainty drops below a threshold ($\tau_\uparrow$).
  • Track $\to$ Loiter: Occurs when uncertainty exceeds a higher threshold ($\tau_\downarrow$).
  • During loitering, low-confidence detections bias the heading toward the target, facilitating a smooth transition once confidence is sufficient.

3. Key Contributions

• Unified Instantaneous Relative Formulation: HUNT is the first framework to integrate high-speed traversal, target acquisition, and tracking in cluttered, GPS-denied environments under a single relative formulation, eliminating reliance on global pose.
• Drift-Resilient Navigation: By anchoring objectives to instantaneous observables rather than global maps, the system maintains safe high-speed flight even when horizontal position estimates drift significantly.
• Seamless Mode Transition: The framework enables reactive switching between searching (loitering) and tracking without requiring global state re-initialization or reverting to fragile SLAM/VIO methods.
• Real-Time Safety: The integration of CBFs within the NMPC ensures collision avoidance in dynamic, unstructured environments without requiring prior maps.

4. Experimental Results

The framework was validated in extensive outdoor experiments across diverse environments: urban container compounds, semi-structured layouts, and dense forests.

• Loitering Performance:
  • In urban environments, the UAV maintained steady forward progress at ~4.1 m/s with heading errors bounded within 2° relative to GPS ground truth.
  • In dense forests (under canopy), the UAV successfully traversed clutter at 2.0 m/s, maintaining clearance margins >1.0 m. GPS logs in these conditions showed horizontal errors of 2–6 m, which would render map-based navigation unreliable, yet HUNT sustained drift-free flight.
• Full SAR Missions:
  • Urban Search: The UAV located a stationary vehicle wreckage, switching from loitering to tracking with displacement errors bounded within ±0.5 m.
  • Urban Pursuit: The UAV tracked a moving target (mannequin on an ATV) at speeds up to 11.4 m/s, keeping lateral error within a 1.2 m corridor. Low-confidence detections successfully biased the loitering heading toward the target before the switch.
  • Forest Search: The UAV traversed dense trees at 2.0 m/s, switched to tracking a mannequin, and sustained pursuit despite severe reflections and heading corrections up to 25°.
• Quantitative Metrics: Across all missions, HUNT achieved consistent high-speed performance (mean velocities up to 8.85 m/s) and agile navigation (peak accelerations up to 31.11 m/s²) without parameter retuning.

5. Significance and Claims

The paper claims that HUNT represents a fundamental shift in autonomy for SAR operations by demonstrating that both traversal and tracking can be expressed without global pose or mapping.

• Robustness in Degraded Sensing: The system achieves reliable autonomy in conditions where traditional GPS, VIO, and SLAM methods fail due to drift, feature sparsity, or occlusion.
• Unified Capability: It bridges the gap between "searching" (traversing unknown terrain) and "tracking" (following a target), a capability previously missing in relative navigation frameworks.
• Safety and Agility: By embedding CBFs into the control loop, the framework guarantees safety and dynamic feasibility at high speeds in unstructured environments.

The authors note that while the current implementation uses a fixed initial heading for loitering, future work could incorporate semantic reasoning (e.g., recognizing doors/windows) or lightweight short-term memory to improve exploration efficiency and prevent cyclic loitering, all while maintaining the instantaneous relative formulation.