Omnidirectional Dual-Arm Aerial Manipulator with Proprioceptive Contact Localization for Landing on Slanted Roofs

arXiv:2602.10703 · cs.RO · Submitted 2026-02-11 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: Hi, I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: We're looking at the paper: "Omnidirectional Dual-Arm Aerial Manipulator with Proprioceptive Contact Localization for Landing on Slanted Roofs".

Dev: Here's the gist of the paper. The gist: A novel dual-arm aerial manipulator design featuring an omnidirectional 3D workspace and extended reach, combined with a proprioceptive contact detection and contact localization strategy based on a momentum-based torque observer, enables the UAM to infer the inclination of slanted surfaces blindly through physical interaction prior to touchdown.

Rosa: How it works The core innovation involves developing a dual-arm aerial manipulator with an omnidirectional 3D workspace and low-inertia limbs that minimize the reaction torques on the body <ref:2602.10703#pg6>.

Dev: This design is intended to overcome limitations of existing platforms by enabling the manipulator to operate within a workspace that revolves planarly 360◦ around the vehicle’s body <ref:2602.10703#pg2>.

Rosa: The system's properties include a body mass of 1287 g and a total mass of 4145 g, with rotor arm lengths specified as 0.160 m <ref:2602.10703#pg4>.

Title and authors: Rosa: So, looking at this paper, "Omnidirectional Dual-Arm Aerial Manipulator with Proprioceptive Contact Localization for Landing on Slanted Roofs," the title tells us immediately that they’re focusing on a dual-arm setup and using contact to figure out where the drone is landing.

Dev: That's right. The core idea here is moving past just having a physical arm to actually using that interaction with the surface to get geometry information, which is what they call proprioceptive contact localization <ref:2602.10703#pg5>.

Taro: It’s interesting how they framed the problem by saying conventional sensing methods like cameras or laser rangefinders can be unreliable in poor weather or low light, and this paper wants to avoid relying on external perception hardware <ref:2602.10703#pg2>.

Rosa: Exactly. They are suggesting that you can infer the landing location geometry through physical interaction instead of needing vision or acoustics to figure it out.

Dev: And they introduce this strategy by developing a proprioceptive contact detection and contact localization method based on a momentum-based torque observer <ref:2602.10703#pg5>.

Taro: That sounds like they are using the system's own physics, specifically the change in angular momentum, to detect when contact happens <ref:2602.10703#pg5>.

Rosa: They define the total angular momentum of the system with an equation that includes terms for both the body and the arms <ref:2602.10703#pg5>.

Dev: And then they estimate that external torque by looking at the residual of how much that angular momentum changes over time, which is what triggers a contact registration <ref:2602.10703#pg5>.

The paper's summary: Rosa: So, summarizing what they’ve done with this "Omnidirectional Dual-Arm Aerial Manipulator with Proprioceptive Contact Localization for Landing on Slanted Roofs," the main contribution is moving from guessing the surface geometry to actually sensing it through physical interaction.

Dev: They use a momentum-based external torque observer to detect contact and figure out which arm makes contact, and this allows them to get a localized view of the landing geometry <ref:2602.10703#pg5>.

Taro: It’s interesting how they defined the total angular momentum using an equation like L(t) = I b bω b b(t) + X i=zero I i iω i i(t) + miv b i(t), as shown in that equation, and then estimate the external torque as the residual of the change in angular momentum <ref:2602.10703#pg5>.

Rosa: Exactly. They register a contact point when that estimated torque goes above a certain threshold, and they have a timeout to make sure they don't register multiple contacts just because it bounces around <ref:2602.10703#pg5>.

Dev: Once they have that contact point pc in world frame, they figure out the surface slope by defining a plane using three points on the surface and calculating the surface normal ns from the cross product of two vectors, r1 and r2 on that plane <ref:2602.10703#pg5>.

Taro: After finding that landing plane, they then use equations thirteen and fourteen to plan the required heading psi s and surface incline alpha, which tells them exactly what angle to aim for <ref:2602.10703#pg5>.

Rosa: And the final step is moving the manipulator into position so that the slope between the end-effector and the rear landing gear matches that calculated landing surface. That's how they plan a successful landing on an incline of up to thirty point five degrees, which was validated in flight experiments where they achieved an average estimation error of two point eight seven degrees across nine total flight tests <ref:2602.10703#pg1,of up to 30.5>.

Dev: A two point eight seven degree average error is what they found, and the paper shows that this level of accuracy is sufficient for repeatably planning and executing landings on surfaces up to thirty point five degrees incline, which is pretty good for real-world application <ref:2602.10703#pg1>.

Taro: The fact that they show successful landings across three different surface inclines—eleven point three degrees, twenty point six degrees, and thirty point five degrees—shows the versatility of this method when dealing with different real-world roof geometries <ref:2602.10703#pg6>.

Rosa: It’s about proving that physical interaction can provide a robust way to estimate surface inclination without needing complex external sensing hardware to do it all at once.

The paper's improvements: Dev: Thinking about the improvements suggested by "Omnidirectional Dual-Arm Aerial Manipulator with Proprioceptive Contact Localization for Landing on Slanted Roofs," the main improvement is moving away from relying entirely on exteroceptive sensing hardware like cameras or lasers.

Rosa: They are suggesting that contact can be used to infer landing location geometry, which bypasses the need for external perception to figure out where you landed, which is a major advantage when you're dealing with bad weather or low light.

Taro: And they're addressing the mechanical constraints too. They acknowledge that even if you successfully estimate the surface properties through contact sensing, there can still be mechanical constraints posed by uneven ground that prevent a successful landing, which they note <ref:2602.10703#pg2>.

Dev: So, their design addresses both the perception issue and the physical constraint issue. The dual-arm omnidirectional morphology helps them reach a workspace where they can interact with those surfaces effectively <ref:2602.10703#pg3>.

Rosa: They also propose this system can handle landing on surfaces with inclinations up to thirty point five degrees, which is a significant improvement over what simpler mechanisms might be able to do reliably on their own <ref:2602.10703#pg1>.

Dev: And that error rate of two point eight seven degrees over those nine experiments means they have a solid foundation for repeatably planning and executing these landings, which is what you need for any safety case in real-time control loops <ref:2602.10703#pg1>.

Rosa: This "Omnidirectional Dual-Arm Aerial Manipulator with Proprioceptive Contact Localization for Landing on Slanted Roofs" paper gives us a concrete way to handle the mechanical reality of landing on complex surfaces.

Taro: Yeah, the fact that they used a momentum-based torque estimator is what makes this method interesting because it connects the physical interaction directly to the system's dynamics <ref:2602.10703#pg5>.

Dev: It’s about using those dynamics to find a solution, which is how you build robust control systems that can handle unexpected physical situations during flight <ref:2602.10703#pg5>.

Conclusion: Rosa: So we're wrapping up on "Omnidirectional Dual-Arm Aerial Manipulator with Proprioceptive Contact Localization for Landing on Slanted Roofs." Basically, they figured out how an aerial robot can sense a slanted surface just by physically touching it and using momentum to figure out the angle before it lands.

Dev: Exactly. It's not about having a perfect camera setup; it's about using the robot's own movement and forces to get that information, which is what we care about for real-time control loops.

Taro: I just want to stress that they validated this up to a thirty point five degree slope, which is pretty substantial for something learned through contact rather than pre-programmed geometry <ref:2602.10703#pg1>.

Rosa: And the average error they reported was two point eight seven degrees across nine flight tests, which means it's accurate enough for reliably planning and executing landings on those surfaces <ref:2602.10703#pg1>.

Dev: That accuracy number is what matters for me; if the latency in that torque observer estimation is too high, the safety case falls apart because we won't know what we're landing on in time <ref:2602.10703#pg5>.

Taro: I mean, when things go wrong, like unexpected bounces or rapid changes in contact points during a touchdown attempt, they have a timeout mechanism to prevent registering multiple contacts and getting confused <ref:2602.10703#pg5>.

Rosa: Right. So the whole system is built with that low-inertia dual-arm design specifically to minimize reaction torques on the body so it doesn't become unstable when it’s interacting with that slanted surface <ref:2602.10703#pg3>.

Dev: That low inertia is a key part of why this works in the first place; it keeps the dynamics manageable while the AI is figuring out those contact points <ref:2602.10703#pg3>.

Taro: It shows how physical interaction can bridge the gap between perception and action, which is a really important concept for autonomous systems that have to handle unpredictable environments.

Rosa: Well, I think what this paper means for us is that we might see more aerial platforms operating in complex urban geometries where vision alone isn't enough to define the landing zone.

Dev: It gives us a path toward developing safer landing protocols even when the environment misbehaves and external sensing gets noisy or blocked <ref:2602.10703#pg5>.

Taro: And I think looking at how they handle those failure modes during interaction is more telling than just the final success rate of the experiment.

Rosa: So, that's our take on this paper about the omnidirectional dual-arm aerial manipulator with proprioceptive contact localization for landing on slanted roofs. Next up, we’ve got a look at how other robots are using vision and language models to navigate indoors.

Martijn B.J. Brummelhuis, Nathan F. Lepora, Salua Hamaza

BioMorphic Intelligence Lab, Department of Control & Operations, Faculty of Aerospace Engineering, TU Delft · School of Engineering Mathematics and Technology, Bristol Robotics Laboratory, University of Bristol

cs.RO

Submitted: 2026-02-11

Updated: 2026-02-11

Comments: Accepted into 2026 International Conference on Robotics and Automation (ICRA) in Vienna

Journal ref: 2026 IEEE International Conference on Robotics and Automation (ICRA)

DOI: 10.1109/ICRA57385.2026.11697086

License: http://creativecommons.org/licenses/by/4.0/

Importance score: 80/100

The gist: The gist: A novel dual-arm aerial manipulator design featuring an omnidirectional 3D workspace and extended reach, combined with a proprioceptive contact detection and contact localization strategy

Key concepts

Dual-Arm Aerial Manipulator: This is a robot with two arms mounted on an aerial vehicle that can move in any direction in three dimensions. The design uses low-inertia limbs to keep the main body stable when interacting with uneven surfaces, allowing it to operate safely where traditional robots might struggle.
Proprioceptive Contact Detection: This is a method of sensing physical contact using internal measurements rather than external sensors like cameras. The system detects when an arm touches something by monitoring changes in the robot's total angular momentum, which indicates a physical interaction has occurred.
Momentum-Based Torque Observer: This is the core technique used to find out where and how hard contact happens. By measuring how the robot's total angular momentum changes over time, the system can estimate the external torque applied by a surface, helping it pinpoint which arm made contact and determine the surface's angle.

Summary

The gist: A novel dual-arm aerial manipulator design featuring an omnidirectional 3D workspace and extended reach, combined with a proprioceptive contact detection and contact localization strategy based on a momentum-based torque observer, enables the UAM to infer the inclination of slanted surfaces blindly through physical interaction prior to touchdown.

How it works

The core innovation involves developing a dual-arm aerial manipulator with an omnidirectional 3D workspace and low-inertia limbs that minimize the reaction torques on the body <ref:2602.10703#pg6>. This design is intended to overcome limitations of existing platforms by enabling the manipulator to operate within a workspace that revolves planarly 360◦ around the vehicle’s body <ref:2602.10703#pg2>. The system's properties include a body mass of 1287 g and a total mass of 4145 g, with rotor arm lengths specified as 0.160 m <ref:2602.10703#pg4>.

Proprioceptive Contact Detection and Localization

The strategy for inferring surface inclination relies on a proprioceptive, sensorless contact detection and contact localization which enables the aerial manipulator to learn the inclination of landing slopes <ref:2602.10703#pg5>. This is achieved using a momentum-based external torque observer to detect when contact occurs and on which arm <ref:2602.10703#pg5>. The total angular momentum of the system at a given time instant is defined by the equation L(t) = I b bω b b(t) + X i=0 I i iω i i(t) + miv b i(t), as shown in equation (5) <ref:2602.10703#pg5>. The external torque is then estimated as the residual of the change of angular momentum at any given time instant t, expressed in equation (6) <ref:2602.10703#pg5>. A contact point is registered if the absolute value of the estimated torque rises above a threshold and a timeout prevents registration of multiple contact points due to bouncing <ref:2602.10703#pg5>.

Surface Slope Determination and Landing Planning

For contact localization, a virtual moment is determined to compare with the estimated torque, by assuming a virtual force f b v in the opposite direction of end-effector movement <ref:2602.10703#pg5>. A contact point pc in the world frame is then obtained by pc = pb + Rbp b e,j <ref:2602.10703#pg5>. The surface slope determination involves defining a plane by three points and calculating a surface normal ns using the cross product of two vectors (r1, r2) on the plane <ref:2602.10703#pg5>. Once the landing plane is identified, a landing is planned by determining the required heading ψs and surface incline α using equations (13) and (14) <ref:2602.10703#pg5>. The manipulator is then moved in position such that the slope between the end-effector and the rear landing gear matches the landing surface <ref:2602.10703#pg5>.

Experimental Validation Results

The approach was validated through a flight testing campaign comprising 9 total flight tests on 3 different surface inclines, with surface incline angles of 11.3◦, 20.6◦, and 30.5◦ <ref:2602.10703#pg6>. The average error of the surface inclination estimation algorithm was found to be 2.87◦ across all experiments <ref:2602.10703#pg7>. This average error was shown to be sufficient for repeatably planning and executing landings <ref:2602.10703#pg7>. Over the 3 sets of experiments, all landings were successful, meaning the UAM estimated the slope angle, set the manipulator arms in the correct configuration, and landed on the incline allowing the motors to be shut off <ref:2602.10703#pg7>. The total average incline estimation error was 2.87◦ <ref:2602.10703#pg7>.

Conclusion

A novel omnidirectional dual-arm aerial manipulator design is introduced and demonstrated through the application of landing on slanted surfaces <ref:2602.10703#pg7>. The system successfully performs contact detection and determines on which of the end-effectors the contact occurs, utilizing a momentum-based torque estimator <ref:2602.10703#pg5>. System validation confirms effectiveness up to 30.5◦ slope angle <ref:2602.10703#pg7>. The average error of the surface inclination estimation algorithm is 2.87◦, which was shown to be sufficient for repeatably planning and executing landings <ref:2602.10703#pg7>. Future work will address the assumption that the plane is aligned with the UAM’s heading through adding contact points, constraining the plane in 3D space <ref:2602.10703#pg7>. The versatility of this novel morphology will be explored in advanced application scenarios requiring coordinated dual-arm manipulation and navigation <ref:2602.10703#pg7>.

ACKNOWLEDGEMENT

The authors would like to thank Anton Bredenbeck and Dimitris Chaikalis for their help with the hardware and software.

REFERENCES

[1] A. Ollero, M. Tognon, A. Suarez, D. Lee, and A. Franchi, “Past, present, and future of aerial robotic manipulators,” IEEE Transactions on Robotics, vol. 38, no. 1, pp. 626–645, 2021.

[2] M. Bianconi, M. Brummelhuis, and S. Hamaza, “Dynamically reconfigurable force-balanced aerial manipulator for varying payload tasks,” in 2025 International Conference on Advanced Intelligent Mechatronics.

[3] R. Peng, Y. Wang, M. Lu, and P. Lu, “A dexterous and compliant aerial continuum manipulator for cluttered and constrained environments,” Nature Communications, vol. 16, no. 1, p. 889.

[4] A. Suarez, A. E. Jimenez-Cano, V. M. Vega, G. Heredia, A RodriguezCastano, and A Ollero, “Design of a lightweight dual arm system for aerial manipulation,” Mechatronics, vol. 50, pp. 30–44.

[5] H. Paul, R. Miyazaki, R. Ladig, and K Shimonomura, “Tams: development of a multipurpose three-arm aerial manipulator system,” Advanced Robotics, vol. 35, no. 1, pp. 31–47.

[6] A Caballero, A Suarez, F Real, V M Vega, M Bejar, A RodriguezCastano, and A Ollero, “First experimental results on motion planning for transportation in aerial long-reach manipulators with two arms,” in 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE.

[7] M. Ryll, G Muscio, F Pierri, E Cataldi, G Antonelli, F Caccavale, D Bicego, and A Franchi, “6d interaction control with aerial robots: The flying end-effector paradigm,” The International Journal of Robotics Research, vol. 38, no. 9, pp. 1045–1062.

[8] N. Staub, D Bicego, Q Sable, V Arellano, S Mishra, and A Franchi, “Towards a flying assistant paradigm: The othex,” in 2018 IEEE International Conference on Robotics and Automation (ICRA). IEEE.

[9] S. Park, J Lee, J Ahn, M Kim, J Her, G.-H Yang, and D Lee, “Odar: Aerial manipulation platform enabling omnidirectional wrench generation,” IEEE/ASME Transactions on mechatronics, vol. 23, no. 4, pp. 1907–1918.

Improvements for AI systems

Contact detection and localization for landing on slanted surfaces can be performed blindly through physical interaction using a momentum-based torque observer, enabling inference of surface inclination prior to touchdown. This allows unmanned aerial manipulators (UAMs) to operate safely on complex roof geometries without relying on unreliable external sensing methods like vision or acoustics.
The system can accurately estimate the surface inclination angle with an average error of 2.87 degrees over nine experiments, which is sufficient for robust landing planning and execution on surfaces up to 30.5 degrees incline. This capability directly improves the reliability of autonomous navigation in urban environments where roof geometries vary significantly.
The dual-arm omnidirectional aerial manipulator design features low-inertia limbs that minimize reaction torques on the body, which reduces dynamic instability during interaction with uneven ground. This ensures that the system can successfully execute landings on surfaces of unknown inclination, addressing mechanical constraints posed by uneven ground without requiring complex motion planning algorithms at high computational cost.

Abstract

Operating drones in urban environments often means they need to land on rooftops, which can have different geometries and surface irregularities. Accurately detecting roof inclination using conventional sensing methods, such as vision-based or acoustic techniques, can be unreliable, as measurement quality is strongly influenced by external factors including weather conditions and surface materials. To overcome these challenges, we propose a novel unmanned aerial manipulator morphology featuring a dual-arm aerial manipulator with an omnidirectional 3D workspace and extended reach. Building on this design, we develop a proprioceptive contact detection and contact localization strategy based on a momentum-based torque observer. This enables the UAM to infer the inclination of slanted surfaces blindly - through physical interaction - prior to touchdown. We validate the approach in flight experiments, demonstrating robust landings on surfaces with inclinations of up to 30.5 degrees and achieving an average surface inclination estimation error of 2.87 degrees over 9 experiments at different incline angles.