AM-Bench: A Modular Simulation Suite and Benchmark for Aerial Manipulation Policy Learning

Yutong Wang^*, Dongjae Lee^*, Xiaofeng Guo^*, Yuanzhu Zhan, Yufei Jiang, Bavin Saravanan, Muqing Cao, Jia Xie, Chenyang Mao, Sebastian Scherer, Junyi Geng, and Guanya Shi

Carnegie Mellon University · Pennsylvania State University · Kyung Hee University

^* Equal contribution

Documentation

Paper coming soon

Code coming soon

Overview Video

Overview video coming soon

A concise tour of the tasks, aerial platforms, policy interfaces, and dynamics-critical evaluation.

Overview

AM-Bench is designed to systematically evaluate aerial manipulation across diverse dimensions: (Left) a range of platforms, including underactuated (UA-Quad, UA-Hexa) to fully/over actuated (FA-Hexa, Omni-Hexa) systems and different low-level controllers; (Center) a diverse suite of manipulation tasks with domain randomization in spatial poses and textures; (Right) various visuomotor policy architectures and realistic physical constraints, including aerodynamic disturbances (ground/wall effects) and actuator saturation. u, φ(·, ·), s, a, π(·), and o refer to control input, controller, state, high-level action, policy, and observation, respectively.

Headline Results

The current paper evaluates each high-level policy over 30 rollouts on each of 12 tasks using FA-Hexa and the EE-target IK–PID interface. These figures are a fixed paper snapshot, not a live leaderboard.

52.22%

Macro task success

Best reported result, achieved by PI0.5 after multi-task and task-specific fine-tuning.

67.31%

Macro subtask completion

Best reported intermediate-progress result under the same evaluation setting.

+46.4 pp

AM-domain adaptation gain

PI0.5 macro success improvement from zero-shot evaluation to multi-task fine-tuning.

48%

Near-ground RMSE reduction

Ground-effect modeling reduced simulation-to-real EE-height RMSE from 1.91 cm to 0.99 cm in the tested interval.

Read the experimental findings and evaluation scope →

Tasks

Twelve tasks span instantaneous interaction, object transport, articulated objects, and constrained contact.

Task 1: Press Button

Task 2: Peg in Hole

Task 3: Frame Assembly

Task 4: Cabinet Pick and Place

Task 5: Lemon Harvesting

Task 6: Rotate Valve

Task 7: Push Slider

Task 8: Pull Lever

Task 9: Open Door

Task 10: Wipe Window

Task 11: Toss Ball

Task 12: NDT

Domain Randomization

Configurable geometry, placement, textures, and object appearance expose policies to controlled visual and spatial variation.

Wall Texture

Wall Tilt

Randomized object placement in the lemon harvesting task

Randomized object color in the pull lever task

Disturbances

Aerodynamic and actuation models test policy behavior under wind, ground effect, near-wall interaction, and rotor saturation.

Wind Disturbance

Ground Effect Disturbance

Embodiments

Four multirotor embodiments cover underactuated, fully actuated, and overactuated aerial manipulation platforms under a shared benchmark interface.

UA-Quad

UA-Hexa

FA-Hexa

Omni-Hexa

System Diagram

The framework follows a modular pipeline composed of diverse task environments, robot models, disturbance, actuator saturation, low-level controller and high-level visuomotor policy, under a high-fidelity simulation environment.

Community benchmark

Leaderboard coming soon

The public leaderboard will compare submissions under matched task, embodiment, controller, disturbance, rollout, and seed settings. The current paper results remain available as a fixed experimental snapshot.

Leaderboard coming soon

Citation

Please use the provisional citation below. The archival paper link and publication record will replace it when they become available.

@article{wang2026ambench,
  title  = {{AM-Bench}: A Modular Simulation Suite and Benchmark for Aerial Manipulation Policy Learning},
  author = {Wang, Yutong and Lee, Dongjae and Guo, Xiaofeng and Zhan, Yuanzhu and
            Jiang, Yufei and Saravanan, Bavin and Cao, Muqing and Xie, Jia and
            Mao, Chenyang and Scherer, Sebastian and Geng, Junyi and Shi, Guanya},
  year   = {2026},
  note   = {Preprint}
}