Introduction

MAVs have become increasingly essential for applications in exploration, surveillance, and environmental monitoring. Their agility and ability to operate in confined or cluttered environments make them particularly suited for urban missions1, time-critical search-and-rescue deployments2, and ecological data collection3–5. Nevertheless, their performance is fundamentally limited by poor flight endurance: even under ideal conditions, most MAVs can travel only a few kilometres, or sustain flight for tens of minutes6. This constraint severely restricts their ability to perform long-duration tasks, such as environmental monitoring over extended time horizons.

Perching refers to the behavior of biological systems—typically birds or arboreal animals—to rest on a surface by maintaining contact at one or more attachment points. In aerial robotics, perching is often exploited in the context of drones to extend mission duration while maintaining an energy-free resting state. A wide spectrum of perching technologies has been explored in aerial robotics. Design-based solutions typically employ robotic grippers mounted underneath the vehicle, such as avian-inspired claws78, bistable elements3, or passive latching systems for fixed-wing drones9. The main drawback of design-based solutions is, however, their reliance on absolute knowledge of the target location and accurate position tracking during the approach phase. These strong assumptions enable the flying gripper to align with the target; however, with little to no tolerance for uncertainty in the target pose estimate.

Besides these fundamental limitations, there are also alternative perching methods such as adhesive1011, suction-based1213, and magnetic1415 mechanisms, which allow surface attachment but impose strict constraints on the surface cleanliness, smoothness, or material composition for perching. Other design-based strategies leverage friction or geometric interlocking: frictional perching for canopy structures16, hook-based attachment for rough surfaces17, custom grippers for well-defined geometries18–20, and lightweight tensile perching designs21, 22.

For problems of the broader class of aerial interaction, prior works have closed the loop with onboard vision: compliant grippers paired with learned target detectors have been demonstrated for aerial grasping23–25, and stereo and RGB-D pipelines for aerial perching26. Across both tasks, these vision-based pipelines share well-known failure modes: they are trained on a specific set of objects, require an unoccluded view, and most crucially lose the target whenever it leaves the camera’s field of view or is occluded by the gripper itself during the approach. The grasping works do, however, demonstrate two enduring advantages of compliant grippers for aerial physical interaction: shape-adaptive grasping across a range of target geometries, and absorption of impact forces transmitted to the platform base. However, in all proposed designs, these grippers require sustained actuation to maintain a closed grasp, making them energy inefficient for perching. Furthermore, they also provide no feedback on contact establishment or grasp quality, leaving the controller blind to the moment of capture and to any post-grasp slip.

Similar lessons about compliance, underactuation, and feedback have been distilled in conventional ground-based manipulation, where the design space of anthropomorphic and underactuated hands is mature and is the subject of systematic survey27. Tendon-driven anthropomorphic hands exploit synergies (low-dimensional joint-space couplings) to produce diverse grasp modalities (e.g., power and precision) from a minimal actuator set, conforming to object geometry without measuring or explicitly modeling the target28. Across this space, underactuated tendon-driven architectures stand out for delivering anthropomorphic dexterity and inherent mechanical compliance from a minimal motor count, as exemplified by postural-synergy designs derived from human grasping data29, adaptive-synergy designs with antagonistic tendons and tactile fingertips30, and design-time parameter optimization of underactuated hands against analytical grasp-quality metrics31. Commercial platforms such as the 980 g BarrettHand32 showcase the industrial usage of underactuated grasping. However, while achieving impressive dexterity with minimal actuation, their mass envelope still disqualifies them from use in aerial robotics and, as in the previous examples of aerial grippers, their dexterity requires actuation to maintain a closed grasp.

Task-space feedback for manipulation in the form of distributed tactile measurements has likewise been extensively studied33. On one side, tactile and joint-angle signals alone have been shown sufficient for blind grasp-stability classification, without any visual or geometric prior34; on the other, distributed tactile feedback can serve as an active control signal to generate whole-hand envelope grasps on unknown targets35. Many of these demonstrations, however, are realized on heavy, ground-fixed manipulators, with high-fidelity tactile sensors while closing the tactile loop to the actuated fingers rather than using minimal tactile sensors that inform the control actions of the base.

Despite substantial progress in perching mechanisms for drones, deploying these methods in unstructured environments remains challenging. All grippers still require extremely precise maneuvers to guarantee alignment, which are impractical in cluttered settings or in the real world. Adhesive, suction, magnetic, and geometry-specific methods further constrain the range of surfaces suitable for perching. Moreover, nearly all existing approaches rely on accurate prior knowledge of the target’s position and orientation. In the studies mentioned, aerial perching is typically conducted in laboratory settings with Motion-Capture systems providing the pose of the perching target beforehand, and the actual maneuver is then executed in an open-loop fashion with a precise estimate of the target pose, tracking a predefined waypoint trajectory. These challenges highlight the need for additional perception modalities to support aerial perching maneuvers, beyond visual guidance. In this context, continuous tactile feedback can more effectively guide the maneuver, enable real-time alignments during grasping to adapt to diverse target geometries, and validate the robustness of the grasp during interaction. Biological systems embody this principle seamlessly: perching animals rely on tactile feedback and bodily compliance when engaging with their surroundings, using touch to adjust posture and grasp to the environment. Inspired by these behaviors, tactile sensing in aerial robotics has emerged as a novel tool to provide continuous feedback in aerial physical interaction, directly in the task-space36–40.

Building on this insight, we propose an aerial tactile perching framework that equips a compliant anthropomorphic hand with soft binary tactile sensors on an aerial platform. Starting from only a rough estimation of the target location in space, the system continuously leverages tactile feedback to infer the precise location and orientation of the perching structure, to guide real-time flight adjustments during the approach, and to evaluate grasp stability upon perching. By closing the feedback loop directly in the task space, our novel approach enables robust, vision-free aerial perching that accommodates substantial misalignments to the perching target and unpredictable surface geometries. The main contributions of this work are:

• A compliant anthropomorphic hand equipped with soft tactile pads that act as sensorized phalanges, providing both environmental perception and passive adaptation to diverse structures.

• An aerial tactile perching framework that continuously refines the MAV pose during the perching maneuver using embodied tactile feedback.

• A tactile-based grasp validation strategy that ensures secure attachment before finalizing the perch, enabling reliable, energy-free operation.