Omnidirectional Aerial Manipulator (OAM)

Updated 9 July 2026

OAM is a flying robotic system that integrates full 6-DoF force and torque control, enabling independent translational and rotational motion for contact-based tasks.
Key designs range from tilt-rotor hexarotors and dual-arm platforms to hybrid models, illustrating trade-offs between full wrench generation and flight efficiency.
Advanced control and sensing strategies facilitate precise disturbance rejection, real-time wrench allocation, and robust contact detection in diverse operational environments.

An omnidirectional aerial manipulator (OAM) is a flying robotic system that combines aerial locomotion with manipulation while retaining independent control of translational and rotational motion in three dimensions. In the literature, the term covers multiple embodiments: fully actuated tilt-rotor aerial systems with rigidly mounted end-effectors for contact inspection, omnidirectional multirotors carrying articulated rigid or soft manipulators, dual-arm platforms with omnidirectional arm pivots, and hybrid vehicles that trade full wrench generation for improved range and cruise efficiency. The defining technical distinction from conventional underactuated multirotors is the ability to command body force and body torque with substantial decoupling, enabling stable contact, disturbance rejection, arbitrary-pose manipulation in $\mathsf{SE}(3)$ , or, in nearly omnidirectional variants, a reduced but still manipulation-oriented wrench space (Bodie et al., 2019, Lee et al., 27 Aug 2025, Pantic et al., 2023).

1. Conceptual scope and nomenclature

In the fully actuated formulation, an OAM can independently generate forces and torques in all six degrees of freedom, so that translational and rotational dynamics are effectively decoupled within actuator limits. This property underlies the contact-based inspection platform that uses a fully actuated tilt-rotor hexarotor with a rigidly mounted end-effector and the arbitrary-pose manipulation platform that treats the aerial base as a floating body in $\mathbb{R}^3 \times SO(3)$ and jointly plans base and arm motion in the full configuration space (Bodie et al., 2019, Lee et al., 27 Aug 2025).

The term also encompasses platforms in which the manipulator is not a distal serial arm in the conventional sense. One line of work realizes omnidirectional aerial manipulation with a rigid manipulator arm fixed to the body and a tool frame aligned for contact tasks such as whiteboard interaction and non-destructive testing; another uses a fork-like end-effector mounted on a tiltable quadrotor for valve turning under teleoperation; yet another attaches a carbon-fiber hook to an omnidirectional micro aerial vehicle for door opening (Bodie et al., 2019, Li et al., 17 Jun 2025, Cuniato et al., 2023).

A narrower, but important, interpretation appears in hybrid platforms such as Soliro. Soliro is described as a hybrid rotary-/fixed-wing aerial manipulator that achieves nearly omnidirectional $(5\ \mathrm{DOF})$ force and torque generation with a minimal set of actuators. Its omission of one independent wrench axis is explicit: it cannot independently generate lateral force in body- $Y$ without coupling to yaw or tilt, and therefore differs from fully actuated $6$-DoF OAMs even though it remains manipulation-oriented (Pantic et al., 2023).

2. Morphologies and actuation architectures

Representative OAM morphologies differ primarily in how they realize wrench generation, how much of the manipulation burden is placed on the airframe rather than on an attached arm, and whether efficiency in forward flight is prioritized alongside omnidirectionality. These differences are structural rather than cosmetic: the allocation map, actuator count, workspace geometry, and contact strategy all follow from the morphology (Bodie et al., 2019, Brummelhuis et al., 11 Feb 2026, Szász et al., 2021, Bao et al., 23 May 2025).

System	Morphology	Reported capability
Contact-inspection OAM	Hexarotor with six double-propeller groups, each on a servo-actuated tilt axis, plus rigid end-effector	Arbitrary $6$-DoF wrench; contact-based inspection
ETH-Zurich OMAV door-opening platform	Six rigid arms, each with two counter-rotating propellers, plus hook end-effector	Arbitrary $6$-DoF wrench for articulated-object manipulation
Arbitrary-pose $\mathsf{SE}(3)$ OAM	Fully actuated floating base with six tiltable rotors and a $4$-DoF serial arm	Stationary base at arbitrary $6$D pose with whole-body planning
Dual-arm OAM	Standard quadrotor with two $\mathbb{R}^3 \times SO(3)$ 0 serial arms on omnidirectional $\mathbb{R}^3 \times SO(3)$ 1 pivots	Omnidirectional $\mathbb{R}^3 \times SO(3)$ 2D workspace and proprioceptive probing
Soft-arm OAM	Fully actuated OMAV with a soft continuum arm	Unified floating-base and soft-body modeling
Soliro	Hybrid rotary-/fixed-wing split tilt-wing aerial manipulator	Nearly omnidirectional $\mathbb{R}^3 \times SO(3)$ 3-DoF wrench and efficient cruise
MorphEUS	Morphable co-axial quadrotor with paired servos on each arm	Full controllability almost everywhere

The contact-inspection platform uses six double-propeller groups equally spaced around the body $\mathbb{R}^3 \times SO(3)$ 4-axis, each mounted on a servo-actuated tilt axis. By commanding rotor speeds $\mathbb{R}^3 \times SO(3)$ 5 and tilt angles $\mathbb{R}^3 \times SO(3)$ 6, it instantaneously reshapes its global thrust wrench. The same general fully actuated logic appears in the ETH-Zurich OMAV used for door opening, where six tiltable arms and twelve propellers permit arbitrary body-frame wrench production regardless of attitude (Bodie et al., 2019, Cuniato et al., 2023).

The arbitrary-pose manipulation platform employs a fully actuated floating base with six tiltable rotors and a rigid $\mathbb{R}^3 \times SO(3)$ 7-DoF arm. MorphEUS instead uses four arms with coaxial counter-rotating propellers and a paired servo mechanism per arm, so that each thrust vector can be pointed in an arbitrary direction. The MorphEUS summary states that this produces higher and more uniform force/torque reachability with a smaller footprint and minimum thrust cancellations, and that full actuation holds almost everywhere when $\mathbb{R}^3 \times SO(3)$ 8 for generic geometry and generic servo angles (Lee et al., 27 Aug 2025, Bao et al., 23 May 2025).

The dual-arm OAM shifts emphasis from base omnidirectionality to arm omnidirectionality. It uses a standard quadrotor frame with two $\mathbb{R}^3 \times SO(3)$ 9 serial manipulators, each attached through a custom planetary-gear pivot joint that rotates freely through $(5\ \mathrm{DOF})$ 0 about the body $(5\ \mathrm{DOF})$ 1-axis. Experimentally sampled configurations show symmetric reachability in all planar directions and substantial out-of-plane reach up to $(5\ \mathrm{DOF})$ 2 (Brummelhuis et al., 11 Feb 2026).

The soft-arm OAM replaces a rigid serial arm with a soft continuum arm modeled through Piecewise Constant Curvature (PCC) and an Augmented Rigid Body Model (ARBM), unified with a floating-base formulation. Soliro occupies a different point in the design space: two main arms extended into full wings, reversible tail rotor, wing tilt servos $(5\ \mathrm{DOF})$ 3, and rotor-tilt servos $(5\ \mathrm{DOF})$ 4, with no classical control surfaces or flaps (Szász et al., 2021, Pantic et al., 2023).

3. Modeling, wrench generation, and control

A canonical rigid-body OAM model writes the body wrench balance as

$(5\ \mathrm{DOF})$ 5

with $(5\ \mathrm{DOF})$ 6 the actuator-generated wrench and $(5\ \mathrm{DOF})$ 7 an external disturbance. In the tilt-rotor inspection platform, the wrench is

$(5\ \mathrm{DOF})$ 8

so the mapping from $(5\ \mathrm{DOF})$ 9 to $Y$ 0 spans all six dimensions and the system can produce arbitrary wrenches within its thrust limits. The arbitrary-pose $Y$ 1 OAM uses the equivalent Newton–Euler decomposition

$Y$ 2

with control inputs given directly as body-frame force $Y$ 3 and torque $Y$ 4 and allocated through a $Y$ 5 map for six tiltable rotors (Bodie et al., 2019, Lee et al., 27 Aug 2025).

Several control paradigms recur across the literature. For contact tasks, the inspection platform formulates impedance control with selective apparent inertia,

$Y$ 6

and derives a continuous-time control law that combines desired impedance dynamics with estimated external wrench compensation. This allows low apparent mass along the tool normal and high apparent mass in tangential directions and rotational axes, avoiding discontinuous switching between free flight and contact (Bodie et al., 2019).

For arbitrary-pose flight and manipulation, the fully actuated floating-base OAM uses a geometric robust controller on $Y$ 7 built from translational and rotational error states such as

$Y$ 8

Its nominal terms resemble geometric nonlinear PID, while robust terms add integrals of $Y$ 9 for disturbance rejection. The summary reports Lyapunov functions $6$0 and $6$1 yielding ultimate bounds that can be made arbitrarily small by sufficiently large integral-tanh gains (Lee et al., 27 Aug 2025).

The soft-arm OAM generalizes the state to include continuum-arm coordinates and a floating base. The reduced dynamics in PCC coordinates are expressed as

$6$2

where $6$3 are squared rotor speeds and $6$4 are soft-arm chamber pressures. Control is hierarchical and dynamically consistent: end-effector orientation, end-effector position, and OMAV body orientation are treated as prioritized tasks using nullspace projectors based on a dynamically consistent inverse (Szász et al., 2021).

Soliro highlights how minimal-actuator near-omnidirectional control changes allocation. Its high-level command specifies attitude quaternion $6$5, vertical acceleration $6$6, and overall wing tilt $6$7. An inner-loop attitude controller and rate PID produce torque setpoints, while the unified allocation block at $6$8 inverts closed-form expressions for main propeller thrusts, differential wing deflection $6$9, and tail-rotor thrust. Roll is entirely via differential wing tilt, yaw via thrust difference, and pitch via tail-rotor thrust, with collective thrust and body-force vector orientation governed by overall wing tilt $6$0 (Pantic et al., 2023).

MorphEUS formulates the actuation problem in terms of a virtual thrust vector $6$1 and a control-allocation matrix $6$2,

$6$3

Its reported thrust-energy optimum is the pseudoinverse solution

$6$4

which globally minimizes $6$5 subject to force and torque constraints (Bao et al., 23 May 2025).

4. Perception, contact inference, planning, and human interaction

OAM research couples force-capable aerial platforms with sensing and planning strategies that are explicitly contact-aware. In the inspection platform, external wrench estimation is performed with a generalized-momentum observer,

$6$6

which yields the first-order filter $6$7. On-board state estimation combines a visual-inertial sensor running Rovio with a forward-facing time-of-flight camera aligned to the tool frame. The ToF point cloud is filtered near the tool $6$8-axis, a least-squares plane fit estimates a local surface normal $6$9, and the resulting $6$0 setpoint supports autonomous approach, braking, contact, and sliding without mode switching (Bodie et al., 2019).

The dual-arm OAM replaces exteroceptive surface sensing with blind probing. Its momentum-based external-torque observer estimates body torque from changes in angular momentum,

$6$1

and detects contact when $6$2 exceeds a threshold during designated probing intervals. Contact localization then uses virtual torque matching:

$6$3

choosing the arm whose virtual torque direction best aligns with the observed residual. Surface inclination follows from the estimated normal using

$6$4

This pipeline is explicitly designed for slanted-roof landing (Brummelhuis et al., 11 Feb 2026).

Whole-body planning appears most clearly in the arbitrary-pose $6$5 OAM. Its planner is split into an offline end-effector trajectory optimizer over $6$6 and an online kinematic NMPC that jointly optimizes base pose and arm configuration under joint limits, self-collision constraints, obstacle constraints, and manipulability objectives. The offline problems are solved with CasADi+IPOPT using $6$7th-order Runge–Kutta discretization, while the onboard NMPC replans every $6$8 (Lee et al., 27 Aug 2025).

Learning-based interaction replaces explicit online optimization in the door-opening OAM. The PPO policy observes a $6$9-dimensional state consisting of body linear and angular velocities, hook-to-handle vector, flattened body-to-door rotation, and door hinge angle. Its $\mathsf{SE}(3)$ 0-dimensional action specifies a pose correction in the door frame through a translation offset and two orientation vectors that are orthogonalized by Gram–Schmidt before being passed to the inner pose controller. The reward decomposes into hook proximity, attitude alignment, door opening, velocity penalty, and control effort (Cuniato et al., 2023).

Teleoperation provides a complementary interface when autonomy is insufficient or when direct human dexterity is required. The hand-based OAM system maps shoulder-anchored hand motion and finger gestures to $\mathsf{SE}(3)$ 1 setpoints for a tiltable-quadrotor through four modes: Operation Mode, Locking Mode, Spherical Mode, and Cartesian Mode. Hand and shoulder markers are streamed at $\mathsf{SE}(3)$ 2 from OptiTrack, glove data classify gestures, and an actuator-level NMPC tracks the commanded pose (Li et al., 17 Jun 2025).

5. Experimental capabilities and validated performance

The contact-inspection OAM demonstrates the classical fully actuated contact regime. In rope-pull disturbance tests, low apparent mass produces deflections of $\mathsf{SE}(3)$ 3– $\mathsf{SE}(3)$ 4 under $\mathsf{SE}(3)$ 5– $\mathsf{SE}(3)$ 6 pulls, while high apparent mass reduces deflection below $\mathsf{SE}(3)$ 7 for $\mathsf{SE}(3)$ 8 lateral forces and $\mathsf{SE}(3)$ 9 torques. In push-and-slide on a whiteboard, the platform maintains constant contact force of approximately $4$0, rejects frictional tangential drag, and holds attitude error under $4$1 while sliding at speeds up to $4$2. For concrete-vault interaction using only on-board sensing, the local normal is updated at $4$3, the platform translates $4$4 along the surface, pose tracking stays within approximately $4$5 of flight-capture ground truth, and attitude aligns within $4$6. In contact-based non-destructive testing with a copper-sulfate electrode sensor, nine potential measurements at $4$7 intervals are taken while the controller holds approximately $4$8 contact force and position error below $4$9 normal to the wall; the measured half-cell potentials correctly identify the corroded points. The generalized-momentum observer tracks normal forces up to approximately $6$0 with RMS error of $6$1–$6$2 (Bodie et al., 2019).

The dual-arm OAM validates proprioceptive landing on slanted roofs. Over nine trials on slopes of $6$3, $6$4, and $6$5, all nine landings succeed with the vehicle body remaining level upon touchdown. The mean absolute inclination errors are $6$6, $6$7, and $6$8 for the three slope sets, with overall average $6$9. The report attributes robustness partly to probing-interval gating that suppresses false positives during body roll and notes that ground-effect perturbations caused small translational offsets without preventing landing (Brummelhuis et al., 11 Feb 2026).

The arbitrary-pose $\mathbb{R}^3 \times SO(3)$ 00 OAM demonstrates free-flight and contact-adjacent manipulation at extreme attitudes. In controller comparison with the arm oscillating between $\mathbb{R}^3 \times SO(3)$ 01, the proposed gRITE controller attains approximately $\mathbb{R}^3 \times SO(3)$ 02 RMS position error and approximately $\mathbb{R}^3 \times SO(3)$ 03 RMS attitude-geodesic error at $\mathbb{R}^3 \times SO(3)$ 04 pitch, and $\mathbb{R}^3 \times SO(3)$ 05 and $\mathbb{R}^3 \times SO(3)$ 06 at $\mathbb{R}^3 \times SO(3)$ 07 pitch. In grasp-and-pull tasks, the base hovers at $\mathbb{R}^3 \times SO(3)$ 08– $\mathbb{R}^3 \times SO(3)$ 09 pitch while the arm reaches under a bar, grasps, and retracts; reported tracking errors are $\mathbb{R}^3 \times SO(3)$ 10– $\mathbb{R}^3 \times SO(3)$ 11 RMS in position and $\mathbb{R}^3 \times SO(3)$ 12– $\mathbb{R}^3 \times SO(3)$ 13 RMS in orientation in ground scenarios, and $\mathbb{R}^3 \times SO(3)$ 14– $\mathbb{R}^3 \times SO(3)$ 15 and $\mathbb{R}^3 \times SO(3)$ 16– $\mathbb{R}^3 \times SO(3)$ 17 near a table, with NMPC solve times of $\mathbb{R}^3 \times SO(3)$ 18– $\mathbb{R}^3 \times SO(3)$ 19 and $\mathbb{R}^3 \times SO(3)$ 20– $\mathbb{R}^3 \times SO(3)$ 21 respectively (Lee et al., 27 Aug 2025).

The reinforcement-learning door-opening OAM emphasizes robustness to mismatch and large initial offsets. Against a state-of-the-art MPPI baseline running at $\mathbb{R}^3 \times SO(3)$ 22, the PPO policy achieves $\mathbb{R}^3 \times SO(3)$ 23 success for hook distance up to $\mathbb{R}^3 \times SO(3)$ 24, while MPPI drops below $\mathbb{R}^3 \times SO(3)$ 25 beyond approximately $\mathbb{R}^3 \times SO(3)$ 26. For lateral and vertical handle-position offsets up to $\mathbb{R}^3 \times SO(3)$ 27, RL remains above $\mathbb{R}^3 \times SO(3)$ 28 success. Completion time is always below $\mathbb{R}^3 \times SO(3)$ 29 for RL, whereas MPPI often requires $\mathbb{R}^3 \times SO(3)$ 30– $\mathbb{R}^3 \times SO(3)$ 31. Zero-shot transfer initially opens the door to $\mathbb{R}^3 \times SO(3)$ 32; saturating the maximum action norm and retraining produces a smoother policy that reliably reaches $\mathbb{R}^3 \times SO(3)$ 33 (Cuniato et al., 2023).

Teleoperated aerial manipulation is validated on a valve-turning task in a $\mathbb{R}^3 \times SO(3)$ 34 arena. The procedure uses Spherical Mode to pass an industrial ladder, Operation Mode to align and turn the valve, Locking Mode to resolve occlusion, and Cartesian Mode to exit a corridor. Reported performance includes positional RMSE of $\mathbb{R}^3 \times SO(3)$ 35 during valve turning, orientational RMSE of $\mathbb{R}^3 \times SO(3)$ 36, end-to-end latency of $\mathbb{R}^3 \times SO(3)$ 37– $\mathbb{R}^3 \times SO(3)$ 38, task completion time of $\mathbb{R}^3 \times SO(3)$ 39 averaged over three trials, corridor tracking error within $\mathbb{R}^3 \times SO(3)$ 40 of the centerline, and gesture recognition above $\mathbb{R}^3 \times SO(3)$ 41 (Li et al., 17 Jun 2025).

Hybrid and morphable OAM-related platforms show that omnidirectional manipulation can be extended toward efficient cruise or morphable close-proximity inspection. Soliro’s wind-tunnel study yields a $\mathbb{R}^3 \times SO(3)$ 42-parameter aerodynamic model, $\mathbb{R}^3 \times SO(3)$ 43 and $\mathbb{R}^3 \times SO(3)$ 44 fits accurate to within $\mathbb{R}^3 \times SO(3)$ 45, differential wing-torque sensitivity of $\mathbb{R}^3 \times SO(3)$ 46 per $\mathbb{R}^3 \times SO(3)$ 47 of $\mathbb{R}^3 \times SO(3)$ 48, and minimum stable $\mathbb{R}^3 \times SO(3)$ 49 increment of $\mathbb{R}^3 \times SO(3)$ 50, corresponding to $\mathbb{R}^3 \times SO(3)$ 51 resolution. Flight tests achieve stable hover-to-cruise-to-hover transitions, cruise speed up to $\mathbb{R}^3 \times SO(3)$ 52, and approximately $\mathbb{R}^3 \times SO(3)$ 53 power reduction at $\mathbb{R}^3 \times SO(3)$ 54 relative to hover, with single-battery range extension of approximately $\mathbb{R}^3 \times SO(3)$ 55 and up to $\mathbb{R}^3 \times SO(3)$ 56 under drag optimizations (Pantic et al., 2023).

MorphEUS is currently validated in high-fidelity simulation. In continuous contact inspection of a water tower, translational tracking RMS is reported as $\mathbb{R}^3 \times SO(3)$ 57 with peak $\mathbb{R}^3 \times SO(3)$ 58, and orientation error $\mathbb{R}^3 \times SO(3)$ 59 never exceeds $\mathbb{R}^3 \times SO(3)$ 60. A constrained-pipe traversal succeeds with minimum clearance greater than $\mathbb{R}^3 \times SO(3)$ 61, and a corkscrew-view inspection task achieves angular tracking error $\mathbb{R}^3 \times SO(3)$ 62 while matching the theoretical energy optimum (Bao et al., 23 May 2025).

The soft-arm OAM is also presently simulation-based. In continuous nullspace motion, the end-effector holds $\mathbb{R}^3 \times SO(3)$ 63 with $\mathbb{R}^3 \times SO(3)$ 64 about its $\mathbb{R}^3 \times SO(3)$ 65-axis while the OMAV rotates $\mathbb{R}^3 \times SO(3)$ 66 in alternating axes, keeping position error below $\mathbb{R}^3 \times SO(3)$ 67 and orientation error below $\mathbb{R}^3 \times SO(3)$ 68 for $\mathbb{R}^3 \times SO(3)$ 69. Under disturbance rejection tests, end-effector position returns within $\mathbb{R}^3 \times SO(3)$ 70 and orientation within $\mathbb{R}^3 \times SO(3)$ 71 in under $\mathbb{R}^3 \times SO(3)$ 72, and dynamic trajectory tracking on a horizontal circle yields mean position error of approximately $\mathbb{R}^3 \times SO(3)$ 73 and orientation error below $\mathbb{R}^3 \times SO(3)$ 74 (Szász et al., 2021).

6. Trade-offs, misconceptions, and research directions

A central point of comparison is the distinction between full $\mathbb{R}^3 \times SO(3)$ 75-DoF wrench generation and reduced-order near-omnidirectionality. Fully actuated OAMs can command arbitrary wrenches within thrust limits, which is crucial for contact-rich tasks requiring decoupled translation and rotation. Soliro explicitly trades this away: its $\mathbb{R}^3 \times SO(3)$ 76-DoF design reduces mechanical complexity and weight, avoids extra motors or linkages, and blends over-actuated hover with efficient fixed-wing cruise, but cannot independently generate lateral body- $\mathbb{R}^3 \times SO(3)$ 77 force and therefore cannot provide arbitrary wrench generation in all six axes when full wrench arbitrage is needed (Bodie et al., 2019, Pantic et al., 2023).

Another recurrent misconception is that omnidirectionality is solely a controller property. In the reported systems, controllability is inseparable from morphology and allocation rank: servo-actuated tilt axes, paired-servo thrust vectoring, endless-rotation pivots, coaxial drag-torque cancellation, or split tilt-wings determine whether the wrench map spans six dimensions, how uniformly the reachable set is distributed, and how efficiently forces and torques can be produced. MorphEUS makes this explicit through the condition $\mathbb{R}^3 \times SO(3)$ 78 almost everywhere, while the door-opening and inspection platforms rely on overactuation to resolve wrench allocation in real time (Bao et al., 23 May 2025, Cuniato et al., 2023).

The literature also identifies practical limitations. The soft-arm OAM reports simulation only, unmodeled aerodynamic effects, relatively high computational load with a $\mathbb{R}^3 \times SO(3)$ 79 system, and fixed gains without learning or adaptation. The teleoperation framework lacks force feedback; its report states that Locking Mode was essential to avoid occlusion during valve turning. The dual-arm landing platform notes that high-friction coverings improved contact stability and proposes compliant, high-friction end-effectors to generalize to low-friction roofs (Szász et al., 2021, Li et al., 17 Jun 2025, Brummelhuis et al., 11 Feb 2026).

A plausible implication is that the field is separating into several mature subproblems rather than converging to a single canonical OAM architecture. One branch prioritizes precise contact and disturbance rejection with full wrench control; another targets arbitrary-pose whole-body planning near obstacles; another incorporates proprioceptive probing for surface inference; another studies human-in-the-loop manipulation; and hybrid or morphable vehicles seek better range, efficiency, or wrench uniformity. The reported future directions are correspondingly diverse: real-world soft-arm experiments with on-board state estimation, gain-scheduling or MPC, hybrid force/motion control, tactile or vision feedback, bilateral haptic feedback, application-specific end-effectors, drag optimization, and compliant contact interfaces (Szász et al., 2021, Li et al., 17 Jun 2025, Pantic et al., 2023).