Aerial mobile manipulation requires unmanned aerial vehicles (UAVs) to traverse heterogeneous environments such as indoor and outdoor domains, and connect navigation with subsequent physical interaction within the same trajectories. Existing aerial benchmarks are typically confined to single-environment settings, limiting evaluation of agents across continuous environmental transitions. When interaction is evaluated, agents are often initialized near the target, leaving the transition from navigation to interaction largely unassessed. This work introduces CrossFly, a benchmark for long-horizon interaction-oriented aerial mobility across heterogeneous environments. CrossFly contains 15K trajectories across more than 20 reconstructed scenes under a unified simulation and evaluation framework, covering Indoor, Outdoor, and Indoor↔Outdoor trajectories. It further introduces Interaction Readiness Rate (IRR) to assess whether terminal navigation states provide suitable conditions for downstream interaction without assuming a specific manipulation policy. Experiments with five representative aerial vision-language-action (VLA) agents and the proposed CrossFly-Agent show that Indoor↔Outdoor settings are more challenging than single-environment settings, with the highest success rate reaching 9.7% on unseen scenes. A downstream manipulation case study further validates the relevance of interaction-oriented arrival to subsequent physical interaction. We hope CrossFly helps advance aerial embodied intelligence toward unified mobility and physical interaction in heterogeneous environments.
Figures & tables
Figure 1: Overview of CrossFly benchmark. It focuses on interaction-oriented aerial mobility tasks and consists of a large dataset for unmanned aerial vehicle (UAV) imitation learning, featuring multiple aerial embodiments, heterogeneous environments, photorealistic rendering, and diverse target objects. To enable systematic experimental analysis under interaction-oriented mobility task settings, we additionally provide an evaluation protocol and propose an interaction readiness metric.
Figure 2: Trajectory statistics and scene diversity. (a) Distribution of scene types and task instructions with different numbers of subtasks (§ 3.1 ). (b) Distribution of trajectory lengths and path distances, summarizing the temporal and spatial scales of the dataset. (c) Representative environments in CrossFly , where each scene covers indoor and outdoor environments.
Dataset
Platform
Scene
Ntraj
Nscenes
Environment
Interaction
Dynamics
DoF
R2R ( Anderson et al., 2018 )
Ground Agent
Matterport3D
7,189
90
Indoor
✗
✗
Graph-based
VLN-CE ( Krantz et al., 2020 )
Ground Agent
Habitat
4,475
90
Indoor
✗
✗
2
OpenUAV ( Schmittle et al., 2018 )
UAV
AirSim + UE
12,149
22
Outdoor
✗
✓
6
AVDN ( Fan et al., 2023 )
UAV
xView
6,269
–
Outdoor
✗
✗
3
CityNav ( Lee et al., 2025 )
UAV
SensatUrban
32,637
34
Outdoor
✗
✗
4
Table 1: Comparison with related benchmarks (§ 3.1 ).
Figure 3: Overview of the dataset construction pipeline (§ 3.2 ). CrossFly begins with reconstructed scenes from point clouds and realistic rendering, followed by waypoint annotation and trajectory acquisition through UAV simulation. The collected trajectories are then assembled with RGB observations, UAV states, and task-related annotations to construct the benchmark dataset.
Figure 4: Architecture of CrossFly-Agent (§ 4.1 ) . A high-level VLM planner uses visual context spanning the observed trajectory to predict the active language subtask. Conditioned on this subtask, a low-level VLA policy combines the initial observation with recent visual context to generate flight action chunks and determine when to predict a manipulation-invocation signal.
Indoor
Outdoor
Indoor ↔ Outdoor
Model
NE ↓
SR ↑
OSR ↑
IRR ↑
NE ↓
SR ↑
OSR ↑
IRR ↑
NE ↓
SR ↑
OSR ↑
IRR ↑
π0 ( Black et al., 2024 )
4.07
17.7
23.6
37.2
3.79
20.4
27.8
37.0
7.21
13.2
19.7
29.1
π0 -FAST ( Pertsch et al., 2025 )
3.81
18.5
19.2
39.0
3.42
22.7
30.1
38.9
6.63
15.8
22.1
30.7
OpenFly-Agent ( Gao et al., 2026 )
3.48
15.1
21.4
46.7
3.39
18.1
20.6
43.4
5.93
13.9
17.7
34.2
OpenVLA-OFT ( Kim et al., 2025 )
3.84
17.2
21.7
47.9
3.51
19.5
24.4
38.1
6.87
14.9
21.3
38.8
Table 2: Quantitative comparison on seen scenes (§ 4.2 ).
Indoor
Outdoor
Indoor ↔ Outdoor
Model
NE ↓
SR ↑
OSR ↑
IRR ↑
NE ↓
SR ↑
OSR ↑
IRR ↑
NE ↓
SR ↑
OSR ↑
IRR ↑
π0 ( Black et al., 2024 )
5.72
10.8
13.4
26.7
5.26
13.5
15.8
27.8
9.38
2.4
4.1
21.7
π0 -FAST ( Pertsch et al., 2025 )
5.18
13.2
15.7
28.4
4.83
15.6
17.2
29.1
8.74
3.7
6.3
25.2
OpenFly-Agent ( Gao et al., 2026 )
4.49
10.6
13.9
32.7
4.22
13.4
15.1
26.4
7.61
7.2
7.9
26.2
OpenVLA-OFT ( Kim et al., 2025 )
5.45
11.9
14.6
27.5
4.97
14.1
16.4
28.2
9.02
3.1
4.7
27.9
Table 3: Quantitative comparison on unseen scenes (§ 4.2 ).
Figure 5: Qualitative results of CrossFly-Agent (§ 4.3 ). This example shows the instruction and selected third-person views. Blue text denotes navigation actions, while red text denotes manipulation actions. The top row shows the trajectory from the start to the interaction-ready state, while the bottom row covers the trajectory from the interaction-ready state to task completion.
Task Suite
Traj. Length
# Subtasks
Instr. Length
NE ↓
SR ↑
Indoor
11.2
6.8
18.9
4.15
23.64
Outdoor
11.8
6.7
18.5
4.36
25.28
Indoor ↔ Outdoor
11.5
6.8
18.8
6.43
10.98
Table 4: Performance comparison under matched trajectories.
Figure 6: Manipulation case study (§ 4.4 ). Each row represents a manipulation case. The leftmost image shows the initial observation of UAV. The top case satisfies IRR criteria and achieves a successful grasp, whereas the bottom case violates IRR criteria and fails to grasp.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
Figure 7: Qualitative results of CrossFly-Agent .
Figure 8: Failure case. Sometimes, UAV collides with obstacles when traversing passages across indoor–outdoor boundaries, indicating that navigation in transition regions remains challenging.
Split
# Scenes
Scene overlap with training
Used for training
Training
18
–
Yes
Seen
18
Same scenes
No
Unseen
5
None
No
Appendix
Table 5: Scene-level partition used in CrossFly (§ B.1 ).
We introduce ActiveFly-Bench, the first benchmark to bridge cyberspace reasoning and physical-world interaction for UAV embodied perception. The benchmark decomposes active perception into three hierarchical tasks: Aerial Embodied Question Answering (Air-EQA), Observation Behavior Planning (OBP), and Fine-grained Language-guided UAV Control (FLUC), explicitly connecting high-level task understanding, behavior planning, and low-level control. The datasets are collected from both real-world and simulated outdoor environments for training and evaluation. We further develop ActiveFly, a closed-loop UAV agent that integrates visual-language reasoning with fine-grained control, and deploy it on a physical UAV platform. Experiments with representative VLMs and VLA models show that current UAV agents still struggle with behavior planning, viewpoint adjustment, and robust task completion in active perception. These results establish ActiveFly-Bench as a new testbed for embodied aerial intelligence.
Aerial manipulators extend robotic manipulation into 3D workspaces that are difficult for ground-based robots to access, creating new opportunities for general-purpose manipulation. However, extending Vision-Language-Action (VLA) models to aerial robots introduces distinct challenges due to the tight coupling between manipulation and flight, continuously changing observations, and safety-critical physical interactions. These challenges demand diverse training data and systematic policy evaluation, yet collecting demonstrations and evaluating policies directly on physical aerial platforms are costly, difficult to scale, and hard to repeat under controlled conditions. We present AeroManip-VLA, a scalable benchmark for aerial VLA data generation and policy evaluation. AeroManip-VLA provides a GPU-accelerated simulation framework with low-level payload-aware flight and manipulation control in massively parallel environments. Building on this framework, we combine reusable reinforcement learning policies with expert task rules to automatically generate demonstrations without human teleoperation across diverse objects, environments, and randomized initial conditions. The generated data include basic skills such as grasping and placing, as well as long-horizon tasks that require both navigation and manipulation. We further introduce automated event labeling and trajectory categorization to filter demonstrations. These mechanisms enable fine-grained analysis of task progress, behavioral outcomes, and safety-related failures. Finally, we evaluate a range of imitation learning and VLA baselines across different task settings, revealing their performance characteristics and failure modes. Together, AeroManip-VLA enables scalable aerial manipulation data generation, structured trajectory analysis, and systematic VLA evaluation in simulation prior to real-world deployment.
Rui Huang, Yanlin Mu, Lidong Li +3
National University of Singapore · Beijing Institute of Technology
End-to-end Vision-Language-Action (VLA) models have shown promise in UAV navigation. However, existing approaches typically rely on historical observations to directly predict actions, often struggling in dense urban environments where severe occlusions and sharp turns result in drastic viewpoint transitions. We argue that the ability to "imagine" future states -- inherent in World Models -- is critical for robust decision-making under such partial observability. To address this, we construct a challenging Urban Canyon Traversal Benchmark, specifically designed to evaluate spatial understanding in scenarios characterized by severe occlusions and drastic viewpoint transitions. To this end, we propose WorldFly, a novel world-model-based VLA framework that employs a dual-branch coupled flow matching mechanism to jointly generate future video predictions and navigation actions, thereby explicitly guiding the agent's policy via spatial imagination. Extensive evaluations on our benchmark demonstrate that WorldFly outperforms other baselines, particularly in unseen environments, validating the effectiveness of integrating world models into embodied aerial agents.
Shengtao Zheng, Kai Li, Weichen Zhang +5
Tsinghua Shenzhen International Graduate School · BNRist, Tsinghua University