Organizations: School of Automation Science and Electrical Engineering, Beihang University, Beijing, China · School of Electrical and Electronic Engineering, Nanyang Technological University, Singapore · School of Mechanical and Aerospace Engineering, Nanyang Technological University, Nanyang, Singapore · Department of Electrical and Computer Engineering, National University of Singapore, Singapore 119077, Singapore · School of Aeronautic Science and Engineering, Beihang University, Beijing, China · Beijing Key Laboratory of Bio-inspired Intelligent Technology of Unmanned System, China
As a representative form of embodied robotic interaction, aerial grasping requires close coordination between the aerial platform and manipulator. Most existing systems perform aerial grasping with static or slowly moving targets under fixed scenarios. Dynamic and precise grasping across different conditions introduces additional challenges, including target motion, localization drift, dynamic coupling, and external disturbances. In this work, we present CAPTURER, a dexterous aerial manipulator for centimeter-level grasping across diverse environments. To achieve the challenging goal, we first develop an onboard perception system for state estimation and target detection. Drift correction and global pose transformation are introduced to improve target localization during flight. A cooperative motion planner is proposed to coordinate the mobility of the aerial platform with the dexterity of the manipulator according to the relative target state. To maintain grasping accuracy under disturbances, we design a disturbance-rejection controller to compensate for dynamic coupling, model uncertainty, and external disturbances. The capability of CAPTURER is validated through indoor and outdoor experiments with both static and moving targets. The results demonstrate that CAPTURER achieves centimeter-level grasping accuracy for a marker pen measuring only 1.5 cm in diameter at flight speeds exceeding 1.5 m/s.
Figures & tables
Figure 1: Overview of CAPTURER workflow. Body : CAPTURER integrates onboard perception, cooperative motion planning, and anti-disturbance control within a unified framework. Perception : The perception module combines visual–inertial state estimation, target detection, drift correction, global pose transformation, motion prediction, and active target search to provide consistent robot and target states. Planning : Based on the relative target state, the cooperative planner generates coordinated motion references for the aerial platform and manipulator while accounting for the predicted platform attitude and their different motion characteristics. Control : The dynamic controller reconstructs and compensates the strong coupling effects induced by manipulator motion and employs residual observers to handle model uncertainty and external disturbances. The resulting framework supports indoor and outdoor grasping of stationary and moving targets using onboard sensing and online replanning. More details, including CAD models and source code, are available on the Project Site .
Ref.
Platform
Target
Perception
Plan
Control
Motion
Type
Robot
Target
Replan capability *
Disturbance rejection †
A. Hovering grasping of stationary targets
Seo et al. (2017)
Hexarotor + serial arm
Static
Cylinder
Mocap
Visual
✓
✗
Bucki et al. (2023)
Quadrotor + folding arms
Static
Box
Mocap
Known
✗
✗
Wang et al. (2024b)
Quadrotor + serial arm
Static
Marker pen
Mocap
Mocap
✓
✗
Cao et al. (2025a)
Quadrotor + Delta arm
Static
Block
Mocap
Mocap
✗
✓
Table 1: Comparison of aerial manipulation systems.
Figure 2: Schematic of CAPTURER.
Symbols
Physical interpretations
ms∈R
Total mass of the aerial manipulator
J∈R3×3
Inertia matrix of the aerial platform
pb∈R3
The aerial platform position in {I}
pe∈R3
The end-effector position in {I}
ped∈R3
The end-effector position relative to the manipulator base, expressed in {D}
vb∈R3
The UAV linear velocity in {I}
Table 2: Nomenclature.
Figure 3: Schematic of the serial manipulator.
Figure 4: Visualization of the operational capability.
Joint
ϑi
ri (mm)
Mass (g)
1
[001]⊤
[0077]⊤
210
2
[010]⊤
[0077]⊤
146
3
[010]⊤
[240205]⊤
138
4
[010]⊤
[1480205]⊤
336
Table 3: The manipulator system parameters.
Figure 5: (a) CoM shift along Z -axis. (b) CoM shift along Y -axis. (c) The aerial platform is oscillating when the manipulator moves.
Figure 6: The evaluation of CoM position during manipulation motion.
Figure 7: Performance comparison between state estimation via onboard sensors and ground truth in three trajectories. (a) Circle trajectory. (b) Lemniscate trajectory. (c) Four-leaf trajectory.
Figure 8: (a) CAPTURER detects a fixed target during flight. (b) CAPTURER detects a target that is moving randomly. (c) The position estimation results of the moving target. (d) The predicted velocity output of the moving target. (e) Detection error characteristics. The horizontal axis represents the moving velocity of the target and the vertical axis denotes the distance between the detected target and CAPTURER. (f) The process of searching for a target outside the camera’s FoV.
Figure 9: Estimation of P¨bmB using different differentiation methods.
Figure 10: Estimation results of the coupling torque.
Test I
Test II
Test III
T=12
T=8
T=6.28
T=12,ω =3
T=8,ω =3
T=12,ω =3, wind disturbance
Methods
ϵ
μ
ϱ
ϵ
μ
ϱ
ϵ
μ
ϱ
ϵ
μ
ϱ
ϵ
μ
ϱ
ϵ
μ
ϱ
Baseline I
8.41
4.10
1.72
15.01
7.79
2.56
24.79
11.89
4.37
11.97
4.91
2.32
15.58
8.44
2.84
13.07
5.26
2.40
Baseline II
6.91
3.45
1.43
12.37
6.25
2.16
23.52
13.10
4.87
10.76
3.92
1.92
17.18
7.32
3.52
11.48
4.56
2.24
Baseline III
5.04
2.46
0.89
12.22
4.21
1.82
26.31
9.67
4.26
9.70
3.63
1.78
13.96
6.58
2.78
12.58
3.81
1.93
Our scheme
4.85
1.86
0.96
6.38
2.32
1.21
15.59
5.75
2.60
5.79
2.43
0.99
10.66
5.44
1.88
7.31
2.84
1.16
Table 4: Tracking error statistics of different control schemes (Unit: cm).
Figure 11: Baseline I can track the desired trajectory. However, the tracking performance substantially degrades with the increase of the coupling effects. The dynamical performance of Baseline II is enhanced due to the compensation of the first-order disturbance terms. Baseline III with the offline initialization training process achieves better results than Baseline I and Baseline II. By combining the established coupling model and residual observers, our scheme demonstrates the excellent dynamic performance among all experiments.
Figure 12: Hardware system of CAPTURER.
Figure 13: CAPTURER actively detects and grasps a target, then precisely places it into the specified area. The experimental video can be found in the Supplementary Movie S2.
Figure 14: (a) Process of detecting unknown objects. (b) Process of aerial grasp and place. (c) Position tracking error of the aerial platform. (d) The manipulator joints are regulated to grasp the target.
Figure 15: Sequential snapshots of the aerial grasping in an open lawn. The experimental video can be found in the Supplementary Movie S3.
Figure 16: (a) The distribution of state drift across different times. (b) Estimation results of residual disturbance torque. (c) Position tracking error of the aerial platform along three axes. (d) The manipulator state is continuously regulated to compensate for the UAV tracking error.
Figure 17: Key snapshots of the dynamic aerial grasping. The experimental video can be found in the Supplementary Movie S4.
Figure 18: Poses of CAPTURER in the X - Z plane, where the brown box indicates the captured object.
Figure 19: Key snapshots of the dynamic grasping of a moving object. The experimental video can be found in the Supplementary Movie S5.
Figure 20: Poses of CAPTURER in the X - Z plane, where the gray box indicates the target motion process.
Figure 21: The relative motion between CAPTURER and the target. The instant of starting to close the gripper is denoted by a gray circle. The red dashed line denotes the target velocity.
Figure 22: Evolution of the relative distance and motion velocity.
Figure 23: Representative scenarios that evaluate object adaptability, angular tolerance, and vertical reachability. The experimental video can be found in the Supplementary Movie S6.
Figure 24: The aerial platform is required to track a four-leaf trajectory with different lap times while the manipulator is fixed to the underside of the system. (a-c) Tracking error distribution of all methods in Group \raisebox{-0.20ex}{\scriptsize1}⃝, \raisebox{-0.20ex}{\scriptsize2}⃝, and \raisebox{-0.20ex}{\scriptsize3}⃝.
Figure 25: The aerial platform is required to track a four-leaf trajectory with different lap times while the manipulator follows a prescribed trajectory. (a) Key snapshots of the flight setup. (b) Tracking error distribution of all methods in Group \raisebox{-0.20ex}{\scriptsize4}⃝. (c) Tracking error distribution of all methods in Group \raisebox{-0.20ex}{\scriptsize5}⃝. (d)-(e) The evolution of PbmB and P˙bmB . (f)-(g) The estimation of F^m and τ^m . (h) Estimation result of disturbance torque in the rotational loop. (i) The dynamic tracking performance of the rotational loop.
Figure 26: (a) Experimental snapshot. (b) The wind field. (c) Tracking error distribution. (d) The position tracking performance. (e) The attitude tracking performance. (f) Estimation results of disturbance torque in the rotational loop.
Figure 27: The end-effector position tracking errors in Section 7.3 .
Figure 28: The end-effector position tracking errors in Section 7.4 .
Figure 29: The end-effector position tracking errors in Section 7.5 .
Figure 30: The end-effector position tracking errors in Section 7.6 .
Figure 31: Online optimization time at each replanning step during aerial grasping.
Aerial robotics studies the design, perception, control, and coordination of flying robots for navigation, inspection, transport, and physical interaction.
Reliable aerial grasping in cluttered environments remains challenging due to occlusions and collision risks. Existing aerial manipulation pipelines largely rely on centroid-based grasping and lack integration between the grasp pose generation models, active exploration, and language-level task specification, resulting in the absence of a complete end-to-end system. In this work, we present an integrated pipeline for reliable aerial grasping in cluttered environments. Given a scene and a language instruction, the system identifies the target object and actively explores it to gain better views of the object. During exploration, a grasp generation network predicts multiple 6-DoF grasp candidates for each view. Each candidate is evaluated using a collision-aware feasibility framework, and the overall best grasp is selected and executed using standard trajectory generation and control methods. Experiments in cluttered real-world scenarios demonstrate robust and reliable grasp execution, highlighting the effectiveness of combining active perception with feasibility-aware grasp selection for aerial manipulation.
High-speed aerial grasping presents significant challenges due to the high demands on precise, responsive flight control and coordinated gripper manipulation. In this work, we propose Swooper, a deep reinforcement learning (DRL) based approach that achieves both precise flight control and active gripper control using a single lightweight neural network policy. Training such a policy directly via DRL is nontrivial due to the complexity of coordinating flight and grasping. To address this, we adopt a two-stage learning strategy: we first pre-train a flight control policy, and then fine-tune it to acquire grasping skills. With the carefully designed reward functions and training framework, the entire training process completes in under 60 minutes on a standard desktop with an Nvidia RTX 3060 GPU. To validate the trained policy in the real world, we develop a lightweight quadrotor grasping platform equipped with a simple off-the-shelf gripper, and deploy the policy in a zero-shot manner on the onboard Raspberry Pi 4B computer, where each inference takes only about 1.0 ms. In 25 real-world trials, our policy achieves an 84% grasp success rate and grasping speeds of up to 1.5 m/s without any fine-tuning. This matches the robustness and agility of state-of-the-art classical systems with sophisticated grippers, highlighting the capability of DRL for learning a robust control policy that seamlessly integrates high-speed flight and grasping. The supplementary video is available for more results. Video: https://zikenhuang.github.io/Swooper/.
Ziken Huang, Xinze Niu, Bowen Chai +2
School of Automation and Intelligent Sensing, Shanghai Jiao Tong University, Shanghai, 200240, China
Accurate capture of non-cooperative targets is critical. In an attempt to tackle this intractable challenge, an aerial gripper system integrated with a Gradient-based Real-time Inverse-game Predictor and PlannER (GRIPPER) framework is proposed. The interaction is formulated as a general-sum pursuit-evasion game under incomplete information. Specifically, underlying cost parameters of the target are inferred online, and the open-loop Nash equilibrium (OLNE) strategy is iteratively refined within a receding-horizon loop. To ensure high-frequency execution, a computationally friendly gradient-based inverse-game solver is developed. Without explicit computation of the Hessian inverse, the optimized solution is updated (> 50 Hz) based on implicit differentiation and fast Hessian-vector products. Meanwhile, an anti-disturbance controller is developed to overcome disturbances of uncertain payload and gripper actuation, enabling precise tracking of the planned trajectory and accurate grasping of the target. Simulations and real-world experiments illustrate the superior computational efficiency and task performance of GRIPPER. The task of capturing and delivering a non-cooperative target is accomplished, highlighting the robustness, adaptability, and real-time performance of the framework in highly adversarial scenarios.
Zeshuai Chen, Meng Wang, Jindou Jia +2
School of Automation Science and Electrical Engineering, Beihang University, 100191, Beijing, China · School of Mechanical and Aerospace Engineering, Nanyang Technological University, 639798, Singapore