Aerial-Balance-Bench: A Controlled and Reproducible Drone-Ball Balancing Benchmark for Indirect Dynamic Aerial Manipulation

Aerial manipulation studies and benchmarks have mostly emphasized direct interaction and static or quasi-static tasks, leaving indirect dynamic manipulation insufficiently studied. We present Aerial-Balance-Bench, a controlled and reproducible drone-ball balancing benchmark in which a vertically constrained tethered drone tilts a beam to drive a rolling ball to a target position or along a reference trajectory. The benchmark unifies: 1) target-position balancing and trajectory-tracking tasks; 2) thrust-, velocity-, and position-command interfaces with corresponding dynamic models; 3) a Gym-style Isaac Lab environment with standardized observations, actions, and task-dependent rewards; and 4) evaluation metrics and robustness tests. To demonstrate the benchmark and provide reference results, we develop a delay-aware hierarchical control framework for the velocity-command interface, combining delay identification with model-based state prediction to compensate for low-level tracking latency. Within this framework, we compare three high-level baselines: cascaded PID, nonlinear model predictive control (MPC), and model-free reinforcement learning (RL). Simulation and real-world experiments show the benchmark's feasibility. These results establish Aerial-Balance-Bench as a practical benchmark and baseline suite for studying the core balancing dynamics in indirect dynamic aerial manipulation.