Robust Active-Perception Control for Global-State-Free Aerial-Ground Cooperation

Aerial-ground cooperation typically requires both robots to estimate their positions in a shared global coordinate frame, which is computationally expensive and prone to localization errors. This paper addresses the challenge of performing cooperative target tracking and relative motion control without relying on global-state measurements, specifically focusing on how to maintain target visibility when the UAV's necessary flight maneuvers conflict with the camera's field-of-view (FOV) constraints.

The authors propose COPA (Cooperative Non-Inertial Perception-Aware Framework), which shifts the perception and control problem into a non-inertial, UGV-attached reference frame. To solve the visibility conflict, the system mounts a single-axis gimbal on the UAV to decouple the camera's optical axis from the UAV's body pitch. The framework integrates three key components:

• Active-Perception Model: A mathematical model that relates UAV motion, gimbal angle, and UGV motion to the target's position on the camera's image plane.
• Temporal Convolutional Network (TCN): A predictive module that uses recent UGV motion history to forecast short-horizon acceleration and angular velocity, replacing the common (but inaccurate) constant-parameter assumption during dynamic maneuvers.
• Unified MPC: A model predictive control formulation that jointly optimizes UAV flight commands and gimbal angles to minimize tracking error while keeping the target centered in the camera's FOV.

COPA demonstrates significant improvements in tracking robustness compared to baseline methods. In simulation, it achieved a 100% success rate across various randomized UGV trajectories, whereas baselines frequently failed due to FOV loss or tracking instability. The TCN-based prediction module was shown to reduce peak tracking errors during rapid motion transitions, such as sharp turns or sudden accelerations. Real-world experiments confirmed that the gimbal effectively compensates for aggressive UAV pitch maneuvers, allowing the system to maintain continuous target visibility even when the UAV performs complex relative orbits.