Aerial Manipulation — Drones with Arms, Tethered, Inspection, Construction

Scope. This note covers UAVs (Unmanned Aerial Vehicles) that physically interact with their environment — drones with arms, cable-suspended manipulators, tethered platforms, and the broader class of “aerial physical interaction.” It also covers the production drone-inspection ecosystem (Skydio, FlyAbility, DJI Matrice, Percepto) and the delivery-by-air operators (Zipline, Wing, Manna), which are increasingly the commercial substrate that funds aerial-manipulation research. For multirotor aerodynamics + control without contact, see [[Robotics/multirotor-design]]. For swarms of drones (no manipulation), see [[Robotics/swarm-robotics]].

1. Definition

Aerial manipulation is the union of UAV flight and physical manipulation: grasping, pushing, pulling, applying force, attaching tools, or otherwise interacting with the environment while airborne. It sits at the intersection of two historically separate disciplines — aerial robotics (where the goal was usually “don’t touch anything”) and manipulation (where the base was assumed fixed or wheel-mobile). The combination introduces tightly-coupled dynamics that neither discipline handled alone.

The distinguishing problem: the manipulator and the flight platform share a single floating base, so any arm motion induces a base reaction (conservation of momentum), and any contact force feeds back into the attitude loop. Designing the control + mechanical system as a coupled whole is the entire technical content of the field.

2. Platforms

2.1 Multirotor

The dominant aerial-manipulation base. Quadrotors (DJI Matrice 350 RTK, Inspired Flight IF1200A, Freefly Alta X) up through octocopters (Acecore Zoe, Aurelia X8) and 16-rotor heavy-lift. Payload ranges from 0.5 kg (small inspection drones) to 25 kg (Alta X, Aurelia X8 Mark II).

Trade-offs:

• + Hover capability, low-speed precision.
• + Mechanical simplicity (only rotor control; no swash plate).
• − Energy-inefficient hover (vs. wing lift); 15–45 min flight time on Li-ion.
• − Strong downwash interferes with low-altitude tasks.

2.2 Tiltrotor and convertible

Rotors that tilt between vertical (hover) and horizontal (forward flight). Cyphy Works LVL 1 (defunct 2020), Joby S4 (passenger eVTOL but the kinematic concept transfers), and the experimental hexa-tiltrotor designs from ETH (Brescianini-D’Andrea, 2018) that achieve fully actuated 6-DOF flight — any orientation independently of position.

The crucial 2018 result (Brescianini-D’Andrea): with non-coplanar rotors, a multirotor becomes omnidirectional, eliminating the underactuation that makes manipulation hard. The cost is mechanical complexity + reduced efficiency.

2.3 Coaxial

Two rotors stacked on a shared axis, counter-rotating. Compact footprint useful in confined inspection (FlyAbility Elios uses coaxial inside a cage; Sky-Hero Loki for tactical). Lower efficiency than non-coaxial pairs of the same total area.

2.4 Tethered

Power + signal via a tether to a ground station. Endurance becomes essentially unlimited (limited by tether length, typically 50–200 m). Elistair Orion 2 (~$30k) is the canonical commercial tethered drone, used by police, military, border control. CyPhy Persistent Aerial Reconnaissance and Communication (PARC) — defunct CyPhy Works, technology absorbed elsewhere. Hoverfly Tether is a current vendor.

For manipulation: tether removes battery-mass constraint, enables continuous force application, but constrains workspace + introduces tether dynamics that perturb the platform.

2.5 Fixed-wing

Endurance + speed, but no hover → fixed-wing manipulation is essentially limited to “perch-and-grasp” research demos (e.g., UPenn Cory Lab perching falcon-style); not practical for general manipulation.

2.6 Hybrid VTOL fixed-wing

Vertical takeoff + landing transitioning to wing-borne forward flight. Production examples: Skydio X10D (US-built inspection drone, 2024), Quantum-Systems Trinity Pro/F90+ (mapping; German), Wingtra WingtraOne (Swiss). Range > 100 km, hover capability for the inspection moment. No arm currently.

3. Arms on drones

Arm class	DOF	Examples	Notes
Fixed gripper / mount	0	DJI Matrice payload mount (camera/sensor)	Most “manipulation” in production is sensor placement
Single-DOF jaw	1	Pelican drop mechanism, parachute drop	Drop / pickup binary
2-3 DOF parallel	2-3	Delta-style mini-arms	Low coupling to flight base; limited workspace
Serial 4-6 DOF	4-6	ETH AeRoArms (6-DOF arm on hexacopter); IIT modified KUKA youBot	Full pose control of end-effector relative to base
Soft / compliant	varies	Festo BionicHopper-style, Floating Robotics origami grippers	Contact tolerance; tendon-driven
Cable-suspended	6 (workspace position via cable lengths)	Mellinger-Kumar UPenn 2010 cooperative transport	Workspace is the convex hull of UAVs holding cables

Yale OpenHand (Dollar lab) and 3D-printed underactuated grippers are the de-facto choice for research-grade aerial grippers — passively conform to objects with low actuation cost. Soft Robotics Inc suction-style end-effectors have been integrated in some industrial trials.

The AEROARMS EU H2020 project (2015–2019, coordinated by Anibal Ollero, Univ. of Seville) produced the most ambitious manipulator-equipped flying platforms — dual-arm hexacopters for industrial inspection + maintenance, including bolt-tightening on pipes.

4. Key challenges

4.1 Underactuation + coupling

A standard quadrotor has 4 control inputs (rotor thrusts) for 6-DOF body motion → underactuated. Arm motion at the wrist induces a reaction torque + linear momentum change on the base, which the attitude loop must reject in real time. The cleanest fix is fully-actuated platforms (omnidirectional hexacopters, tilt-rotor designs) — at the cost of weight + efficiency.

4.2 Center of mass migration

Picking up a 1 kg object on a 5 kg drone shifts the CoM by 20% of the arm length — a substantial parameter change that the attitude controller must handle as either an unknown disturbance (robust control) or an online estimate (adaptive control). Trajectory plans must respect the resulting feasible-thrust envelope.

4.3 Environmental disturbances

• Wind — multirotors lose ~50% of effective payload at 10 m/s sustained wind. Industrial inspection thresholds: DJI Matrice 350 rated to 12 m/s, FlyAbility Elios 3 to 7 m/s.
• Downwash — own rotor wash recirculates near surfaces, induces ground effect + interaction with the manipulated object.
• Turbulence near structures (wind turbines, tall buildings) — Reynolds-stress effects locally unpredictable.

4.4 Limited payload + endurance

Practical envelope: payload ≈ 1/4 to 1/3 of total takeoff mass, endurance scales inversely with payload fraction. A 25 kg-payload Alta X has 18 min hover; a 5 kg-payload Matrice 350 has 55 min nominal.

4.5 Contact stability

Touching a surface introduces non-smooth dynamics (the constraint set changes) and eigenfrequencies from surface compliance. Naive position control rejects the contact like a disturbance; needs explicit impedance or admittance control so the drone behaves like a compliant body at contact.

5. Control approaches

5.1 Cascaded (the practical default)

high-level mission → trajectory generator → outer position loop (~50 Hz)
   → attitude loop (~250 Hz) → motor mixer → ESC current control (~kHz)
                                arm joint controllers (parallel ~500 Hz)

This is what runs on production inspection drones (Skydio Autonomy, DJI N3/M30). Manipulation actions are scheduled when the position loop is stable.

5.2 Whole-body MPC (the research frontier)

Brescianini-D’Andrea (ETH, 2018) and the IIT/Naples groups (Lippiello, Ruggiero) formulate the combined drone-plus-arm dynamics as one model + solve MPC over the whole system at 100–500 Hz. Permits aggressive maneuvers (knock down a door, throw an object) where the cascade would lose stability. Computationally expensive; deployed only on heavyweight platforms with onboard Jetson AGX Orin or similar.

5.3 Impedance + admittance control

Hogan 1985 formulated impedance control: instead of commanding position, command a dynamic relationship between motion and force ($M\ddot{x}+B\dot{x}+Kx=F$). For aerial manipulation, the platform is commanded to behave as a virtual mass-spring-damper at the end-effector. Admittance control is the dual: measure force, integrate to motion. Both are essential for safe surface contact + tool-use. Ott et al. (DLR, 2008) generalized to floating-base systems.

5.4 Aerial physical interaction (API)

Umbrella for the research subfield established by Brescianini, the Vijay Kumar group (UPenn GRASP), Tognon-Franchi (LAAS-CNRS), and the Naples group around Lippiello. Focuses on tasks requiring sustained force: pushing a button, drilling, polishing, applying a sticker, tightening a bolt.

6. Notable designs and research platforms

• SAM — “Six-DoF Aerial Manipulator,” ETH Zurich Autonomous Systems Lab. Hexacopter with non-coplanar rotors + 6-DOF arm. Demonstrated valve-turning + force application.
• AEROARMS — EU H2020 project (Seville-led consortium). Multi-arm aerial robots for industrial pipe inspection + maintenance.
• PUMA — Tognon-Franchi 2020; tethered aerial manipulator with controlled tether tension as additional actuator.
• AeroX — Naples (Lippiello-Ruggiero), various platforms.
• HiPeRLab Berkeley (Mark Mueller) — research platform for aggressive multirotor maneuvers; some manipulation work.
• Skydio X10D (2024) — inspection drone with high-quality camera + flashlight, no arm; canonical example of contact-light aerial inspection. Aimed at military + utility customers; US-manufactured.
• DJI Matrice 350 RTK + payload H30T — thermal + zoom + LiDAR + visible spectrum; the workhorse industrial-inspection platform. ~$10kbase;payloads$5–15k each.

7. Inspection — the production-grade application

This is where 95% of aerial-manipulation-adjacent revenue actually lives, even though most of it doesn’t require an arm.

7.1 Wind turbine blades

• SkySpecs (Ann Arbor): autonomous drone blade inspection; ~$2k/turbine(compare$5–10k rope-access; 3× faster). Acquired by ONYX Insight 2024.
• Pegasus Aero, Avitas Systems (subsidiary of Baker Hughes GE, acquired 2017), Sky Futures — multiple commercial operators.
• Typical workflow: drone autonomously flies a programmed pattern, captures multi-angle high-res imagery of all three blades, defect detection via ML on the ground.

7.2 Bridges + civil infrastructure

• FlyAbility Elios 3 (Switzerland, 2022): caged collision-tolerant drone with LiDAR + 4K + thermal, for confined / GNSS-denied inspection (bridge interiors, sewers, ship ballast tanks). ~$45k.
• DJI M30T + payloads — bridge underside, bridge bearings.
• Niricson — software stack for bridge defect mapping from drone imagery.

7.3 Power lines + utility

• Sharper Shape (Finland): autonomous transmission-line inspection with LiDAR + photogrammetry.
• Percepto AIM (Israel): autonomous drone-in-a-box; fully unattended patrols, charging dock. Used by ENEL, Verizon, Florida Power & Light.
• Aeroseed, VHive, Cyberhawk.
• EPRI + EDF lead the utility R&D side.

7.4 Solar farms

• Above Surveying, Heliolytics (Toronto): IR + RGB inspection at scale; one drone covers 25 MW/day; identifies failed cells from thermal anomalies.
• Pix4D + DJI workflows are the most common DIY stack.
• Sitemark — analytics layer.

7.5 Oil and gas

• Sky-Futures (acquired by MISTRAS Group 2017), Avitas Systems / BHGE GE, Cyberhawk (UK): offshore platform + refinery inspection — flare stacks, tank tops, piping isometrics. BP + Shell + Equinor + ADNOC are the major end-customers.
• Confined-space + atmosphere-monitoring use cases push toward FlyAbility Elios + tethered platforms.

7.6 Tank + pressure-vessel interiors

When the asset is internal + isolated:

• Square Robot (Boston): submersible robots inside above-ground storage tanks (no need to empty + degas).
• Vertikal Robotics — climbing crawlers.
• FlyAbility Elios 3 — caged drones inside vessels.

8. Construction + cooperative aerial 3D printing

• ETH Aerial-AM (Kovac group, Imperial → ETH partnership, Nature 2022): multiple drones cooperatively 3D-printing structures with cementitious + polymer materials. Demonstrated dome + complex geometry construction outdoors.
• Imperial Aerial Robotics Laboratory (ARL) (Mirko Kovac) — built-in environment manufacturing; Aerial-BuiLT (2014) brick-stacking demonstration.
• Tongji University + Skanska — drone-printed bridge concepts (early-stage research).
• Spider-bots / Spider-cam-like suspended platforms (not drones strictly) are the production state of the art for stadium-scale “aerial work.”

9. Light shows + drone choreography (mass-precision swarms)

• Verity Studios (Switzerland, Raffaello D’Andrea spin-out 2014): indoor + outdoor synchronized drone shows; Madison Square Garden residencies, Olympic ceremonies. The reference name for high-precision indoor swarms.
• Intel Shooting Star: 1,218 drones at Pyeongchang Winter Olympics 2018 (the world record at the time; broken since by Chinese operators); Tokyo Olympics 2020.
• Pixis Drones (UK), Skymagic (UK/Singapore), EHang Falcon, High Great (Shenzhen — 5,200+ drones at Xi’an 2021).
• These are swarm deployments rather than manipulation, but the precision-control + multi-vehicle-coordination tech is shared.

10. Delivery by drone

Operator	Founded	Volume	Notes
Zipline	2014 (US, ops Rwanda)	> 1,000,000 commercial deliveries (Q1 2025)	Fixed-wing P2 + new P2 quadrotor “Sparrow” 2024; medical supplies Rwanda + Ghana; US Walmart (Pea Ridge AR), Salt Lake City
Wing	2012 (Alphabet)	hundreds of thousands	US suburbs (Frisco TX, Christiansburg VA, Coppell TX, DFW); Australia
Manna	2018 (Ireland)	tens of thousands	Dublin + Texas pilot 2023
Matternet	2011 (Swiss-US)	thousands	Hospital networks Switzerland, UNC Health NC
Flytrex	2013 (Israel)	tens of thousands	NC + TX pilots 2022–2024
DroneUp	2016	thousands	Walmart partnership (defunct as of 2024)
JD.com + Meituan	China	hundreds of thousands	Shenzhen + Beijing routes

Zipline’s volume + reliability is the existence proof that drone delivery is a real market, at least for high-value low-weight payloads. Note that Zipline-style cable-winch delivery is a degenerate form of aerial manipulation — the drone hovers + the package descends on a winch, no arm needed. This evolutionary path (avoid the manipulation problem by lowering things on strings) is winning commercially.

11. Regulatory framework

• FAA Part 107 (US, commercial small UAV ≤ 25 kg, daylight, VLOS) — the regulatory baseline since 2016.
• Part 137 — agricultural operations (spraying); Hylio + XAG operate under this.
• Section 44807 exemption — FAA authority to grant operational waivers; the main BVLOS path 2020–2024.
• Part 108 (proposed BVLOS rule, NPRM expected 2025) — would create a streamlined BVLOS-by-default operational class.
• UTM (Unmanned Traffic Management) — NASA-FAA-industry initiative; provides airspace deconfliction for many simultaneous drones below 400 ft AGL.
• Remote ID rule — FAA Part 89, in effect Sep 2023; all US drones must broadcast ID + position.
• EU: Categories A (open) / B (specific) / C (certified) under EU regulation 2019/947, with C0–C6 product classes. EASA + national CAAs.
• ASTM F38 — standards for UAS operations and design.
• Operator certificates: LUC (Light UAS Operator Certificate) in EU; Part 91/121/135 adjacency in US for commercial passenger operations (relevant for eVTOL).

For aerial manipulation specifically — i.e., not just flight, but contact — no specific certification framework yet exists. Operations happen under research or specific-category waivers.

12. Cable-suspended cooperative manipulation

A separate branch where multiple UAVs cooperatively carry a single payload via cables. The seminal demonstrations:

• Mellinger-Kumar UPenn 2010: three quadrotors carry a beam via cables; coordinated trajectory in real time.
• Bullo / Bicchi / Franchi — formal control theory of cooperative cable transport.
• The geometry is rich because cable tensions can only be positive — this is a complementarity constraint that becomes its own optimization problem.

Practical applications: heavy-lift in disaster relief, agriculture (pesticide tank transport across terrain inaccessible to vehicles).

13. Open research problems

1. Long-endurance manipulation: battery + payload trade-off remains binding; tethered + cable-suspended hybrids most plausible.
2. Contact-rich tasks without explicit force sensing: estimating contact forces from rotor disturbances (similar to legged proprioception).
3. Outdoor full-pose maneuvers in wind: gust-rejection during contact remains brittle.
4. Certification for human-collaborative operations: no regulatory path yet.
5. Multi-arm coordination + dual-handed grasping on a single floating base.
6. Learning-based aerial manipulation: sim-to-real for contact remains harder than for free flight (contact dynamics are stiff + discontinuous).

Adjacent

• [[Robotics/multirotor-design]] — quadrotor + multirotor aerodynamics + control, without contact.
• [[Robotics/swarm-robotics]] — multi-UAV coordination, choreography, formations.
• [[Robotics/impedance-control]] — Hogan impedance + admittance, the core of safe contact.
• [[Robotics/end-effectors]] — grippers + tool-mounts (most apply to aerial too).
• [[Engineering/aerospace-systems]] — broader aerospace engineering reference.
• [[Compute/perception-stack]] — VIO + SLAM for GNSS-denied flight.
• [[Math/optimal-control]] — MPC, contact-implicit optimization.