Downwash-Aware Viewpoint Selection for Flying Camera Observers

A small drone can act as a movable camera for a robot arm. When the arm's own links block the view of the gripper, the drone flies to a spot where it can still see. The trouble is that a drone is not a passive tripod. Its rotors push a jet of air downward, and near a table that jet spreads sideways. To see the gripper well, the drone wants to come close and look down, but that is also where its air blast pushes light objects, sways the arm, and turns a still tabletop into a moving one.

We propose to measure this effect, which we call the drone's downwash footprint on the workspace, and to put it directly into the viewpoint planner. The camera drone would then choose views that trade line of sight against how much it disturbs the task. Put in two sentences: current flying-camera systems pick viewpoints as if the drone's airflow did not exist. We learn a small model of where that airflow lands on the table and arm, and we show that planning with it gives better manipulation outcomes than either ignoring the airflow or banning a fixed cone beneath the drone.

Core Claim

The closest system is the automated drone viewpoint method of Senft et al.. It flies a Ryze Tello around a Franka arm and splits planning into two layers. A 100 Hz local tracker keeps the end-effector centered and in sight. A 5 Hz global search proposes alternative "ghost" viewpoints so the tracker does not get stuck in a poor view. The system handles the drone's flight drift by inflating obstacles, but it treats the drone as a geometric hull with no airflow. Its environment model is also a single static planar scan. It reports 100% visibility on its primary e-stop and valve tasks, zero collisions, and 1–5 s occlusions in complex scenes. It reports nothing about whether the drone moved objects or disturbed the arm. That is the gap this proposal targets.

Our claim is that the workspace disturbance a hovering observer causes depends strongly on the drone's pose. It is not well described by a simple exclusion cone. It also depends on what lies under the drone, because the arm itself deflects the jet. A compact learned model of this disturbance, used as a cost in the slow layer of the existing two-layer planner, should give a better trade-off between visibility and disturbance than any tuned geometric rule. The payoff should be higher task success on manipulation of light, unanchored objects.

Grounding in Prior Work

Two lines of work motivate the gap.

First, Wang et al.'s indoor ego-airflow study shows that a quadrotor's wake rebounding from walls and floors raises the drone's own acceleration variance by more than 20× compared with open air. They model this as a voxel map of disturbance variance that covers 82–97% of the observed variance. They use that map to set state-dependent safety margins for the drone's own trajectory. We borrow the idea of a spatial disturbance map, but we turn it around. Wang et al. ask how the wall disturbs the drone; we ask how the drone disturbs the table, the objects and the arm.

Second, the ground-effect work of Yang et al. shows that near a flat surface the compressed airflow changes lift and adds leveling torques. This confirms that the near-surface flow is strong and structured rather than diffuse.

Three lines of work supply the method. Neural-Swarm learns residual downwash forces between drones with a neural network. It keeps the network's Lipschitz constant bounded through spectral normalization, a weight rescaling that limits how sharply the output can change with the input. It cuts worst-case height error from about 9.4 cm to about 2.4 cm and runs in under 4 ms on a microcontroller. Shankar et al.'s docking work and its SO(2)-equivariant companion build in rotational symmetry about the vertical axis, so a downwash model can be trained from about five minutes of flight. We reuse both ideas: bounded smoothness, so the planner's cost landscape is well behaved, and vertical-axis symmetry, which holds exactly over a bare table.

The new element is that the arm breaks that symmetry. Our model therefore takes the arm's occupancy as an input, and we test whether that input matters.

The downwash-aware swarm planners, such as Downwash-aware trajectory planning for large quadrotor teams and Modeling Aggregate Downwash Forces for Dense Multirotor Flight, constrain the proposal. They show that an exclusion-cone rule is the accepted cheap answer, so we must beat a tuned cone rather than an untuned straw man.

On the vision side, AV-ALOHA shows that moving the camera matters for precision tasks. With a dedicated 7-DoF camera arm, needle threading improves from 30% to 52% and hidden picking from 40% to 60%. It works on a rigid tabletop with no airflow, and it requires an extra arm bolted beside the workspace. AIR-VLA finds that vision-language-action policies for aerial manipulation collapse when the fixed third-person "God view" is removed. That is evidence that an external, repositionable view carries real value. It does not model rotor wash against the work surface.

Research Program

The work has three parts.

The first part is to measure the footprint. We hover the observer drone over a grid of poses: height above the table, lateral offset from a reference point, and camera tilt. For each pose we record the airflow's effect at points on the table with a low-cost pressure or force-sensing array. We also record the displacement of a standard set of light objects, such as paper cards, foam blocks and small plastic parts. A third measurement is the force on the arm's end-effector, read from the Franka's joint-torque estimates. We repeat these measurements with the arm in a range of configurations, because a link beneath the drone redirects the jet sideways.

The second part is to learn a disturbance model. The model is a small network $d_\theta(v, q)$ that takes a drone viewpoint $v$ (position and camera orientation relative to the table) and the arm configuration $q$. It predicts two quantities: the peak horizontal force on a reference object at the target location, and the force on the end-effector. The table-only part of the model is built to respect vertical-axis symmetry. The arm-dependent part takes a coarse occupancy grid of the arm projected under the drone. Spectral normalization bounds the network's smoothness, as in Neural-Swarm. Before planning with the model, we test whether it predicts object displacement on held-out arm configurations and objects.

The third part is to plan with the model. In the slow, 5 Hz layer of the Senft-style planner, each candidate viewpoint is scored by

$$J(v) = C_{\text{vis}}(v, q) + \lambda \, d_\theta(v, q),$$

where $C_{\text{vis}}$ is the existing cost for occlusion and image resolution of the end-effector, $d_\theta$ is the predicted workspace disturbance, and $\lambda$ is a single trade-off weight. The fast, 100 Hz tracker is unchanged, except that it treats high-disturbance regions as soft keep-out zones using a precomputed lookup. Separating the two rates this way keeps the learned model out of the inner control loop.

We also treat the environment as non-static in one targeted way. When the arm moves, the footprint changes, so the slow layer re-evaluates the current viewpoint whenever the arm configuration changes by more than a threshold. It does not rely on the initial scan alone.

Baselines

The naive reference point is the aero-blind planner of Senft et al., reimplemented on our hardware with the same visibility cost and $\lambda = 0$. This shows what happens when a flying camera ignores its own airflow. The paper's own reported results are 100% visibility on primary tasks and zero collisions. On workspace disturbance or object displacement, no reported result was found on Lacuna for this system. We therefore measure its displacement ourselves.

The strongest existing competitor is a geometric downwash exclusion cone of the kind used in downwash-aware swarm planning (large-team planning), here applied to the workspace. The rule forbids viewpoints whose downward cone intersects the table within radius $r$ of the target or the arm. This is the cheap, standard answer, which is why it is the main threat. To keep the comparison fair, the cone gets the same tuning budget as our method. We sweep its radius and half-angle over the same number of trial episodes we spend sweeping $\lambda$, and we report the best setting found. Both methods use the same drone, optimizer, candidate generator and visibility cost. Our method beats the cone if its curve of visibility against disturbance strictly dominates the cone's best curve, and if it gives higher task success on light-object tasks at matched visibility.

We also compare against a static multi-camera array of three fixed cameras, which costs roughly the same in hardware, and against wrist-camera-only operation. Both are evaluated on the same manipulation policy.

Two ablations isolate the key ingredient. One drops the arm occupancy input and keeps only the symmetric table model, which tests whether arm deflection matters. The other replaces $d_\theta$ with an isotropic penalty on distance to the target, which tests whether learning the shape of the footprint matters at all.

Why not just…?

Why not just enforce a fixed exclusion cone under the drone? This is the alternative most likely to work about as well, and the proposal is designed so that it can win. Two observations suggest a cone wastes good viewpoints. Wang et al. find that wake rebounding off surfaces spreads the disturbance well beyond the vertical jet. Yang et al. show that ground effect near a surface produces leveling torques, not just more lift. If the footprint is lopsided or depends on the arm's configuration, a cone must be made large enough to cover the worst case. It then excludes useful low, oblique views that disturb little. The tuned-cone comparison settles the question directly. If the tuned cone matches our learned footprint on the visibility-versus-disturbance curve, we will report that the simple rule suffices and that the measured footprint dataset is the main contribution.

Why not just use fixed cameras or a dedicated camera arm? On a lab tabletop, a camera arm like AV-ALOHA avoids airflow entirely, and it is a fair choice there. The case for a drone is in settings where nothing can be mounted around the work site, such as field panels, valves and switchgear. Those are the tasks Senft et al. target, and AIR-VLA shows policies degrade badly when the fixed external view is unavailable. Static arrays also suffer self-occlusion at the contact point. AV-ALOHA reports only 30% needle-threading success with static views against 52% with active viewing. We include the static array as a baseline rather than assuming it loses.

Why not just hover far away with a zoom lens? Occlusion depends on viewing angle, not distance, so a distant drone can in principle choose the same angle. The costs are that a long lens magnifies the drone's hover jitter in the image, and that indoors the distant line of sight is more often blocked by clutter. We test this directly by adding a far-standoff variant, fixed above the height where the disturbance becomes measurable, and comparing its end-effector visibility and image stability against the other planners.

Proposed Experiments

Experiments use a Franka arm and a small observer drone. The tasks follow the e-stop, circuit-breaker and valve setup of Senft et al., extended with pick-and-place of light objects (cards, foam, small parts) where downwash should matter. We add matched heavy objects as a control, where it should not. The manipulation policy is held fixed across camera strategies, either a teleoperator with a fixed protocol or a diffusion policy trained once.

The metrics are:

  • the fraction of time the end-effector is visible;
  • the end-effector's size in the image, in pixels;
  • object displacement before grasp, in millimeters;
  • end-effector deviation from its intended path;
  • task success.

Three checks decide the outcome.

The first is the visibility-versus-disturbance curve against the tuned cone, which is the main claim.

The second is a dose-response control. The gap between the aero-blind planner and ours should be large on light objects and near zero on heavy ones. If the gap is similar for both, our gains come from something other than airflow.

The third is a held-out test of the learned model. Its predictions of displacement must transfer to arm configurations not seen in training. Separately, the ablation without the arm input must lose accuracy specifically when an arm link sits beneath the drone.