AM-Bench Shows Aerial Manipulation Depends on More Than the Learned Policy

A flying arm is a coupled system

A manipulation policy that works on a table cannot simply be attached to a drone and expected to work in the air. A ground robot can push against its base while reaching, turning a valve or sliding an object. An aerial manipulator must generate those interaction forces while keeping itself aloft. Motion of its arm can disturb the aircraft, contact can push the entire vehicle away, and a controller may request more thrust than the rotors can produce.

AM-Bench, a preprint submitted on September 1, turns those interactions into a modular research environment. Its authors built a simulation suite with 12 tasks, four multirotor embodiments, multiple low-level controllers, two high-level command interfaces, aerodynamic disturbances and actuator limits. Researchers can vary those parts independently instead of reducing every result to a single policy success rate.

That structure is useful because a failed aerial task can have several causes. The learned policy may select the wrong movement. The command interface may expose too many aircraft variables. The low-level controller may track poorly near a surface. The airframe may lack force in the required direction, or its motors may saturate. AM-Bench is designed to separate these mechanisms before researchers commit to extensive hardware trials.

Twelve tasks extend beyond airborne picking

The simulated suite divides its 12 tasks into three groups. Brief interactions include pressing a button and inserting a peg. Transport tasks include frame assembly, cabinet pick-and-place, lemon harvesting and throwing a ball. Constrained or articulated interactions include rotating a valve, pushing a slider, pulling a lever, opening a door, wiping a window and performing a non-destructive-testing motion.

These tasks demand more than reaching a coordinate. Opening a door constrains the end effector to an arc. Sliding or wiping requires sustained contact without destabilizing the vehicle. Carrying an object changes the system’s dynamics, while throwing requires a rapid coordinated motion. Built-in randomization can alter such properties as surface texture, friction and wall orientation.

The suite represents four aerial platforms: underactuated quadrotor and hexarotor designs, a fully actuated hexarotor and an overactuated hexarotor. Underactuated aircraft must tilt to accelerate laterally. Fully actuated designs can produce lateral force without the same attitude change. The benchmark therefore exposes how identical manipulation objectives can generate different base motion, tracking error and rotor demand on different aircraft.

The interface between policy and controller matters

AM-Bench separates the high-level policy from the low-level machinery that keeps the aircraft stable. One interface asks the policy for a desired end-effector target, such as a gripper pose. Inverse kinematics or whole-body model-predictive control then determines how the aircraft and arm should move together. The other interface lets the policy command base and arm targets more directly.

This distinction changes what the learned model must master. An end-effector policy can concentrate on the manipulation objective while an engineering controller handles redundancy, aircraft attitude and constraints. A policy that issues base and joint targets must learn more of that coordination from data.

In representative simulation ablations using Diffusion Policy, the end-effector interface matched or exceeded the corresponding direct base-and-arm interface in the displayed task comparisons. For pushing a slider with proportional-integral-derivative control, the end-effector configuration succeeded in 63.3 percent of 30 trials, while the direct interface recorded no successes. Whole-body model-predictive control reached 83.3 percent on that task and 100 percent on pressing a button.

Those figures are task-specific simulation results, not general controller rankings. Tracking-error statistics were also calculated only over successful runs, so a configuration that fails frequently cannot be judged from its successful-trial precision alone. The broader result is that a high-level policy score depends partly on the control layer beneath it.

Zero-shot VLA transfer remained weak

The researchers compared specialist imitation-learning methods with two pretrained vision-language-action models, known as pi-zero and pi-zero-point-five. For each task, they collected 80 successful scripted demonstrations in simulation. Every reported method was then evaluated over 30 rollouts for each of the 12 tasks, with macro averages giving every task equal weight.

Without aerial-manipulation fine-tuning, the two generalist models achieved macro-average task success rates of 8.61 and 2.78 percent. Multi-task fine-tuning raised those rates to 36.67 and 49.17 percent. Adding task-specific fine-tuning produced 45.00 percent for pi-zero and 52.22 percent for pi-zero-point-five.

The improvement is evidence of adaptation within this benchmark, not proof of autonomous deployment. The policies were trained and evaluated against AM-Bench tasks, observations and controllers. Their rankings could change with different cameras, aircraft, demonstrations, disturbances or definitions of success. Specialist Diffusion Policy also remained competitive, reaching 41.39 percent macro-average success without being a general-purpose VLA.

Subtask completion was higher than complete-task success for every reported method. That gap indicates that policies often performed useful intermediate behavior but failed before finishing the full sequence. Frame assembly, valve rotation and ball throwing remained difficult. A model that reaches or grasps correctly can still lose control during contact, transport or release.

Simulation includes disturbances, but not complete aerodynamics

AM-Bench models wind and drag as well as ground and near-wall aerodynamic effects. It also imposes rotor-thrust saturation, meaning a commanded force can exceed what the simulated aircraft is allowed to deliver. An optional reduced-order actuator model adds a first-order motor response and a limit on how quickly normalized rotor speed can change.

These additions make the simulator more informative than an idealized environment in which every command is executed immediately. They also remain approximations. The paper does not model complete motor, propeller and electronic-speed-controller dynamics or the full range of disturbances an aircraft could encounter outside the laboratory.

Embodiment comparisons demonstrate why those limits matter. For selected pushing and door-opening trials, underactuated platforms used more tilt and often approached rotor limits more heavily than the fully actuated platform. The overactuated design had a larger available wrench space but could adopt tilted configurations because its solver had more options. More actuation therefore changes the trade space rather than automatically producing the best value in every metric.

Hardware tests validate selected pieces, not the full benchmark

The paper includes physical experiments, but they have a narrower purpose than the 12-task simulation results. In one test, a real fully actuated hexarotor followed a vertical end-effector path near the ground. Adding the benchmark’s ground-effect model reduced simulation-to-real vertical-position RMSE from 1.91 to 0.99 centimeters in the designated near-ground interval, a 48 percent reduction reported by the authors.

At higher altitude, the two simulated configurations were comparable: 0.75 centimeters without the ground-effect model and 0.84 centimeters with it. This supports the model for the tested trajectory and altitude intervals. It does not establish aerodynamic fidelity around arbitrary walls, during forceful contact or in outdoor wind.

The researchers also instantiated their demonstration collection, Diffusion Policy training and end-effector command pipeline on a physical fully actuated hexarotor for lemon harvesting. The paper documents that hardware implementation but does not present a physical counterpart to the complete policy-ranking table. It would therefore be incorrect to describe the simulation success rates as demonstrated real-world performance across the 12 tasks.

AM-Bench also excludes cooperative aircraft, cable-suspended loads, continuum manipulators and soft grippers. It focuses on single multirotor platforms with rigid arms. Comprehensive transfer studies across tasks and embodiments remain future work, and the preprint has no independent replication reported.

A diagnostic benchmark can accelerate practical progress

Aerial manipulation could eventually reduce the need for people to reach elevated or constrained equipment for inspection, light maintenance or sampling. That benefit depends on reliable contact, not merely stable flight near an object. A benchmark that exposes whether failures come from perception, policy, controller, aircraft geometry or actuator limits can help researchers direct hardware time toward the right problem.

The most valuable result here is not that one VLA or aircraft won. It is that system choices visibly changed the outcome. Fine-tuning closed part of the policy domain gap. End-effector commands often simplified high-level learning. Fully actuated flight reduced some coupling between lateral movement and tool orientation. Reduced-order aerodynamic modeling brought one simulated trajectory closer to its physical counterpart.

Those findings remain bounded by the authors’ simulator and selected hardware experiments. Even so, AM-Bench provides a practical framework for testing combinations that would be expensive and risky to explore only through flight. Used with physical validation rather than as a substitute for it, that structure can make aerial-robot development more measurable and less dependent on isolated demonstrations.