UAVs Meet Embodied Intelligence: Bridging Human Intents and Flying Dynamics Via Harnessing Physical-Digital AI Agents

1 Introduction

• Existing UAV systems often assume predefined objectives, closed-set observations, and fixed task interfaces, limiting their ability to interpret open-ended requirements and adapt as missions evolve,.Earlier work advanced from flight control toward task-specific perception, navigation, planning, and coordination,.Source: §1 · p. 1
• UAV embodied intelligence grounds intelligence in aerial embodiment, with physical observations and execution outcomes grounding digital reasoning and digital decisions guiding subsequent perception and action.The digital agent provides semantic understanding, memory, world modeling, reasoning, and task organization; the physical agent realizes sensing, flight, control, actuation, and interaction.Source: §1 · p. 1
• The paper distinguishes UAV EI capabilities from their system realization in an embodied-intelligent UAV and connects them through five capability dimensions and five architectural layers.The framework organizes research across embodiment, perception, world modeling, planning, navigation, manipulation, and collaboration.Source: §1 · p. 2
• The survey organizes research across embodiment, perception, world models, planning, navigation, manipulation, and collaboration, emphasizing their coupling in the perception-reasoning-action loop.It identifies long-horizon aerial autonomy, flight-grounded physical reasoning, adaptation to changing flight conditions, and experience-driven capability evolution as challenges.Source: §1 · p. 2

2.1 UAV EI: Five Capability Dimensions

• UAV EI comprises human intent understanding, embodied self-awareness, general environment understanding, complex task planning, and closed-loop execution.These capabilities operate jointly and depend on grounding intelligence in aerial embodiment and physical interaction.Source: §2.1 · p. 2
• The five capabilities connect human instructions to actionable tasks, account for platform state and limits, actively acquire environmental information, plan feasible action sequences, and update behavior from physical feedback.Intermediate objectives and strategies may change as a mission evolves.Source: §2.1 · p. 2

2.2 EI UAV: Five Architectural Layers

• An EI UAV realizes these capabilities through the Embodiment, Brain, Cerebellum, Interaction, and Harness layers.Together, the layers cover physical resources, general cognition, embodiment-specific skills, external cooperation, and runtime coordination.Source: §2.2 · p. 2
• The layers are tightly coupled rather than sequential: embodiment defines physical constraints, cognition supports reasoning, the Cerebellum realizes skills, Interaction enables cooperation, and Harness coordinates operation.The Harness integrates otherwise separate capabilities into an executable system.Source: §2.2 · p. 3

3 UAV Embodied Morphology

• Body, arm, and hand morphology jointly determine mobility, interaction workspace, and physical contact, while payload placement, arm motion, and contact forces affect vehicle dynamics.The framework groups morphology into aerial bodies, arms, hands, and their co-design.Source: §3 · p. 3, §3.4 · p. 5

3.1 Aerial Body Morphology

• Aerial bodies span multirotors, fixed-wing and VTOL platforms, reconfigurable and multimodal designs, soft structures, and bio-inspired forms, with trade-offs in hovering, endurance, maneuverability, accessibility, payload, and stability.Multirotors hover and maneuver well, fixed-wing platforms improve endurance and cruise efficiency, and hybrid VTOL combines hovering with efficient forward flight.Source: §3.1 · p. 3, Figure 2 · p. 4

3.2 Aerial Arm Morphology.

• Aerial arms include serial, coupled-linkage, and continuum designs, extending interaction beyond the vehicle body while coupling manipulation with flight pose and stability.Continuum arms such as Aerial Elephant Trunk support shape-adaptive reaching and wrapping in cluttered environments.Source: §3.2 · pp. 3–4

3.3 Aerial Hand Morphology.

• Aerial hands include rigid and compliant graspers, enveloping carriers, anchoring feet, and tool couplers to support grasping, transport, perching, and tool use.These interfaces must balance grasp reliability, stability, and tolerance to uncertainty under limited payload and moving-base dynamics.Source: §3.3 · pp. 4–5

3.4 Body-Arm-Hand Co-Design

• Co-design integrates body and hand, arm and hand, or tools across UAVs, combining mobility and interaction functions within coupled structures.Examples include body-level thrust vectoring for interaction forces [13], continuum reaching with wrapping, and cooperative tool transport and exchange.Source: §3.4 · p. 5

4.1 Active Perception

• Active perception selects sensing actions to improve environment and state understanding, locate specified targets, or obtain observations suitable for subsequent reasoning.Methods adjust sensing coverage, target search, viewpoint, distance, or sensor configuration.Source: §4.1 · p. 5

4.2 Embodied Visual Tracking

• Embodied visual tracking regulates UAV motion to maintain informative observations of moving targets, coupling identification, temporal estimation, motion prediction, and control.Approaches include task-specific policies, VLA methods, and foundation-model-augmented modular pipelines; UAV-specific examples address flight actions, occlusion recovery, and obstacle-aware planning [55]–[57], [63], [64], [68].Source: §4.2 · pp. 5–6, Figure 3 · p. 6

4.3 Benchmarks

• Evaluation resources include simulation platforms, task-specific benchmarks, and offline trajectory datasets, but fixed observation-action sequences do not directly evaluate closed-loop interaction.Benchmarks cover tracking, viewpoint regulation, and active target acquisition.Source: §4.3 · p. 6

5 UAV World Models

• World models and multimodal foundation models predict environmental evolution and flight consequences, providing priors and cognitive maps for latency mitigation, sim-to-real transfer, and long-horizon navigation.The section groups their uses into spatial-temporal intelligence, predictive spatial representations, generative simulation, model-based flight control, and world-action modeling with vision-language integration.Source: §5 · pp. 6–7

5.1 Spatial-Temporal Intelligence for MLLMs

• General MLLMs struggle with low-altitude UAV grounding, ego-motion perception, and tracking dynamic interactions because of ground-centric bias and weak temporal-spatial awareness.Specialized geometry-aware data, motion modeling, and hierarchical cognitive maps are among the approaches described for addressing these challenges.Source: §5.1 · p. 7

5.2 Predictive Spatial Representations

• Predictive spatial representations encode geometry and visibility or anchor generated futures to 3D priors and real observations to limit spatial or semantic drift.Mapping-Aware Dreamer uses occupancy and visibility maps; ANWM adds geometric priors, while WorldVLN periodically grounds imagined latents in real observations.Source: §5.2 · p. 7

5.3 Generative Simulation and Data Synthesis

• World models can synthesize interaction data and navigable environments, while real-world 6-DoF data exposes temporal degradation and environment-randomization generalization correlates with deployment success.FlyMirage uses an LLM and 3D Gaussian Splatting to synthesize navigable environments.Source: §5.3 · p. 7

5.4 Model-Based Reinforcement Learning for Flight Control

• Model-based reinforcement learning optimizes flight policies in latent imagination, with Dream to Fly learning camera-pointing behavior and AirDreamer achieving zero-tuning sim-to-real transfer under extensive domain randomization.AirDreamer also discovers escape maneuvers such as reversing out of unmapped physical traps.Source: §5.4 · p. 7

5.5 World-Action Models and Vision-Language-Action Systems

• World-action systems jointly connect prediction and executable actions, including FlowPilot’s 18 ms edge inference and ImagineUAV’s conversion of language-guided imagined video into planner-refined 6-DoF waypoints.FlowPilot avoids future-video decoding at test time and generates controller-trackable references; ImagineUAV refines waypoints for safety and actuation feasibility.Source: §5.5 · pp. 7–8

6.1.1 Explicit Model-Based Planning

• Explicit model-based planning represents geometry, dynamics, objectives, and constraints directly, preserving feasibility conditions while converting intent and observations into structured states, maps, goals, or costs.Classical and later approaches construct dynamically feasible trajectories and improve replanning, obstacle avoidance, or perception awareness.Source: §6.1.1 · p. 8

6.1.2 Policy-Centric Planning

• Policy-centric planning shifts online computation into learned mappings from observations or states to decisions, with structure encoded through demonstrations, supervision, rewards, or training distributions.Recurrent policies use observation history, teacher-student learning transfers privileged planning information, and safety-oriented reinforcement learning introduces task constraints.Source: §6.1.2 · p. 8

6.1.3 Structured Hybrid Planning

• Structured hybrid planning assigns difficult perception, prediction, semantic reasoning, or motion generation to learned components while retaining explicit geometry, trajectory execution, or safety mechanisms.Foundation-model guidance can also feed geometric search, waypoint refinement, or dynamics-aware execution.Source: §6.1.3 · p. 8

6.1.4 World-Model Planning

• World-model planning predicts how states or observations evolve under actions, enabling reasoning about future outcomes beyond choosing actions from current or recent states.Aerial systems use latent dynamics, spatial memory, future-observation prediction, and combinations with VLA generation or kinodynamic planning.Source: §6.1.4 · p. 8

6.1.5 Foundation-Model Planning

• Foundation-model planning incorporates open-ended visual and linguistic intent, either generating actions directly or supplying semantic goals, waypoints, or priors for downstream planning.The latter may be processed by geometric or dynamics-aware modules.Source: §6.1.5 · p. 9

6.2.1 Dynamic-Limit Flight

• Dynamic-limit flight couples sensing horizon, trajectory feasibility, perception, and control as speed rises, exposing trade-offs among latency, physical structure, and robustness near the flight envelope.Reviewed approaches range from explicit minimum-time optimization to learned policies and structured perception-to-execution systems.Source: §6.2.1 · pp. 9–10

6.2.2 Collision-Constrained Flight

• Collision-constrained flight must progress toward goals with occluded, unseen, or moving obstacles, balancing perceptual coverage and reaction latency with motion representation and explicit safety enforcement.Methods include recurrent policies, trajectory or waypoint generation, and safety filters before physical execution.Source: §6.2.2 · p. 10

6.2.3 Intent-Conditioned Flight

• Intent-conditioned flight translates sketches, visual references, language, or semantic tasks into motion compatible with geometric and dynamic execution.The reviewed pattern grounds semantic reasoning through geometric search, waypoint refinement, or dynamics-aware execution.Source: §6.2.3 · pp. 10–11

7 UAV Embodied Navigation

• Aerial navigation must connect high-level objectives to executable flight while handling changing altitude, heading, and viewpoint alongside dynamics, sensing, computation, energy, and safety constraints.The review distinguishes task formulations, navigation methods, and benchmarking resources.Source: §7 · p. 11

7.1.1 Route-Following Navigation

• Route-following navigation aligns ordered language descriptions of landmarks and motion with observations and progress, with viewpoint changes and accumulated grounding errors complicating long-distance flight.AerialVLN established outdoor language-guided route following; later work extends instruction-motion alignment toward city-scale navigation.Source: §7.1.1 · p. 11

7.1.2 Dialogue-Assisted Navigation

• Dialogue-assisted navigation uses interaction to reduce uncertainty about landmarks, destinations, spatial relations, or required behavior rather than merely processing dialogue history.AVDN combines dialogue history, visual grounding, and waypoint prediction; later approaches add structured spatial memory and clarification.Source: §7.1.2 · p. 11

7.1.3 Semantic-Goal Navigation

• Semantic-goal navigation requires active exploration to find an object, region, scene, or open-vocabulary concept whose location is unspecified.Small or occluded targets, viewpoint changes, flight time, and energy constrain aerial search; the UAV-ON benchmark established this task formulation.Source: §7.1.3 · p. 11

7.1.4 Geospatially Conditioned Navigation

• Geospatially conditioned navigation aligns an external reference such as satellite imagery, maps, coordinates, or overhead views with egocentric observations before generating local motion.UAV-VLA studies satellite-image and language conditioning, while LinkS2Bench isolates alignment between UAV observations and satellite imagery.Source: §7.1.4 · p. 11, Figure 4 · p. 12

7.2.1 Task-Specific Policies

• Task-specific policies map observations and task conditions to navigation actions through a dedicated model, with generalization tied to vocabulary, observation model, action interface, and training environment.Inputs can include images, language, dialogue, or platform state; outputs range from discrete actions to waypoints and motion commands.Source: §7.2.1 · p. 12

7.2.2 Hierarchical Methods

• Hierarchical methods divide navigation across scales, using high-level task progress, semantic memory, or subgoals and lower-level local motion responsive to observations.This organization is used for long-range navigation, recovery, and semantic search.Source: §7.2.2 · pp. 12–13

7.2.3 AI Agent Methods

• AI agent methods use pretrained language or multimodal models in iterative reasoning loops with memory, spatial representations, or tools to ground intermediate decisions into executable motion.See, Point, Fly converts model reasoning into image-space waypoints for downstream geometric execution.Source: §7.2.3 · p. 13

7.2.4 VLA Methods

• VLA methods integrate multimodal understanding and action generation in a learned model, directly specializing representations for outputs such as waypoints, velocities, or flight commands.Generated actions still must satisfy flight dynamics, collision constraints, latency, and real-time control requirements.Source: §7.2.4 · p. 13

7.2.5 World-Model Methods

• World-model methods use predicted future observations, states, or action consequences for lookahead navigation rather than selecting actions only from current observations.Prediction quality remains sensitive to large viewpoint changes and environmental variation.Source: §7.2.5 · p. 13

7.3 Benchmarks

• Navigation resources divide into task-execution benchmarks, grounding or diagnostic resources, and configurable simulation platforms.Grounding resources isolate capabilities such as geographic localization or embodied reasoning.Source: §7.3 · p. 13
• AerialVLN, TravelUAV, and CityNav evaluate route following; AVDN tests dialogue-assisted navigation, UAV-ON semantic-goal search, and HUGE-Bench multi-stage tasks.LinkS2Bench evaluates geospatial grounding rather than complete navigation.Source: §7.3 · p. 13
• Closed-loop navigation changes the UAV state with each action, unlike offline evaluation on fixed observations or trajectories.Configurable platforms can support different interaction regimes depending on the instantiated task.Source: §7.3 · p. 13
• Evaluation measures include task completion, path efficiency, goal proximity, trajectory consistency, collision, and task progress.The most informative UAV embodied navigation evaluation combines intent grounding with efficient, safe, closed-loop physical execution.Source: §7.3 · p. 13

7.4 Embodied Grasping and Aerial Manipulation

• Floating-base dynamics, UAV–manipulator coupling, and aerodynamic disturbances make aerial grasping distinct from fixed-base manipulation.Robust systems must address interaction-data scarcity, nonlinear kinodynamic constraints, and sim-to-real transfer gaps.Source: §7.4 · p. 13

7.4.1 Data Regimes and Foundation Model Transfer

• AIRVLA provides a 12-DoF teleoperated aerial-manipulation simulation benchmark, where current physical foundation models grasp basically but degrade catastrophically in long-horizon spatial navigation.Flight transfer also must handle payload mass shifts, altitude droops, and inertia variations that violate terrestrial quasi-static assumptions; AirVLA uses payload-aware inference guidance and temporal chunking for composite navigate-then-grasp without expensive retraining.Source: §7.4.1 · pp. 13–14

7.4.2 Kinodynamic Modeling and Integrated Planning

• Regime-conditioned diffusion models capture residual-dynamics transitions across flight, grasping, and release, while integrated planning models the aerial system as a varying collision ellipsoid.Imitation priors and soft penalties guide trajectory optimization; throwing cable-suspended payloads requires bridging offline trajectory optimization and online model predictive control across taut and slack cable states.Source: §7.4.2 · p. 14

7.4.3 Learning-Based Adaptive Control and Interaction

• Adaptive control methods address contact torque disturbances and center-of-mass changes through dynamics estimation, contextual meta-RL, and direct end-effector pose control.The approaches target payload uncertainty, heterogeneous object properties, and severe underactuation in lightweight hardware.Source: §7.4.3 · p. 14

7.4.4 Task-Specific Execution and High-Speed Grasping

• Fruit picking uses sub-task imitation learning to address phase-dependent data distributions, while Flying Hand uses end-effector-centric teleoperation to mitigate cross-platform coupling issues.Swooper captures moving objects at 1.5 m/s through synchronized flight and gripper control, using progressive RL curriculum training to avoid early collision penalties.Source: §7.4.4 · p. 15

8 UAV Embodied Collaboration

• Collaborative UAV systems extend single-agent capabilities through complementary viewpoints, sensing, mobility, and skills, with progress toward heterogeneous teams.The section covers collaborative perception, navigation, task planning, motion coordination, and aerial manipulation.Source: §8 · p. 15

8.1.1 Collaborative Perception; 8.1.2 Collaborative Navigation

• Collaborative perception moves from feature sharing and representation alignment toward active acquisition and use of complementary information.Collaborative navigation spans dual-altitude roles, leader–follower strategies, air-ground search and handoff, and decentralized exploration under task-allocation, connectivity, and communication constraints.Source: §8.1.1 · p. 15, §8.1.2 · p. 15

8.1.3 Collaborative Task Planning; 8.1.4 Collaborative Motion Coordination

• Task planning increasingly coordinates heterogeneous agents through capability descriptions, negotiation, online assessment, and execution feedback rather than fixed roles.Collaborative motion work combines language-guided generation with constraints and refinement, and increasingly addresses physically coupled systems and payloads.Source: §8.1.3 · pp. 15–16, §8.1.4 · p. 16

8.1.5 Collaborative Aerial Manipulation and Operation; 8.2 Methodological Analysis

• Collaborative aerial manipulation extends cooperation to contact tasks such as construction, docking, and tool exchange, with stronger agent–payload–environment coupling.Shared challenges include joint reasoning, heterogeneous information exchange, and physically executable decisions.Source: §8.1.5 · p. 16, §8.2 · p. 16

8.2.1 The Role of Foundation Models in Collaboration

• Foundation models mainly provide high-level cognition, while specialized planners, controllers, or learned policies handle physical execution.Across collaboration roles, models support planning, navigation reasoning, behavior generation, and knowledge transfer.Source: §8.2.1 · p. 16

8.2.2 Information Sharing and Alignment Strategies

• Feature sharing retains detailed perceptual information but requires alignment across diverse sensing, while unified task-oriented representations improve interoperability but may discard task-specific details.The central design question is what information to share, at what abstraction level, and under which task constraints.Source: §8.2.2 · pp. 16–17

8.2.3 Safety Assurance Mechanisms

• Safety mechanisms include geometric constraints, optimization-based refinement, and learned policies; deterministic projection, road-feasible planning, and validation can limit semantic-error propagation but not incorrect localization or false verification.Safety is framed as an architectural concern linked to semantic reasoning, predictive models, and safety-critical execution.Source: §8.2.3 · pp. 17–18

8.3 Benchmark and Simulation Platforms; 8.3.2 Collaborative Task and Mission Benchmarks; 8.3.3 Collaborative Simulation Platforms

• Collaborative evaluation includes perception benchmarks, mission benchmarks, and simulation platforms, with resources moving toward complete perception–reasoning–action loops.AGOS-Bench separates aerial discovery, candidate guidance, and ground verification, while collaborative simulation platforms support closed-loop mission evaluation.Source: §8.3 · p. 18, §8.3.2 · p. 18, §8.3.3 · p. 18

9 Challenges and Future Directions

• Current UAV embodied intelligence remains largely optimized for bounded missions and predefined operating conditions.More general aerial autonomy requires addressing long-duration operation, physical understanding, adaptation, and continual capability development.Source: §9 · p. 18

9.1.1 Long-Horizon Aerial Autonomy

• Long-horizon autonomy requires selectively retaining information and linking observations across time, space, viewpoints, and mission states.Simply increasing model context length is not presented as sufficient for coherent understanding over extended operations.Source: §9.1.1 · p. 18

9.1.2 Flight-Grounded Physical Reasoning

• Flight-grounded reasoning must predict environmental evolution, interaction outcomes, and action consequences while accounting for embodiment, dynamics, payloads, and disturbances.This moves beyond semantic understanding toward physical reasoning for aerial systems.Source: §9.1.2 · pp. 18–19

9.1.3 Changing Flight Conditions Adaptation

• Changing weather, sensing, payload, battery, and vehicle conditions motivate rapid adaptation from limited experience without extensive retraining.Adaptation must maintain stability, safety, and physical feasibility.Source: §9.1.3 · p. 19

9.1.4 Experience-Driven Capability Evolution; 9.2 Harnessing Physical-Digital AI Agents

• Capability evolution requires consolidating and refining experience into reusable skills and workflows while preserving previously acquired knowledge.The proposed Harness links digital reasoning and physical execution through runtime orchestration and feedback, with action and evolution loops that verify outcomes and reuse interaction experience.Source: §9.1.4 · p. 19, §9.2 · p. 19, Figure 6 · p. 19

• UAV embodied intelligence is moving toward integrated systems coupling semantic understanding, memory, physical reasoning, planning, and closed-loop execution.Source: §10 · p. 19
• The framework connects capability development with realization across Embodiment, Brain, Cerebellum, Interaction, and Harness layers.Source: §10 · p. 19
• Progress across morphology, perception, world models, planning, navigation, manipulation, collaboration, and benchmarks reflects tighter coupling between high-level intelligence and physical interaction.Source: §10 · p. 19
• Harnessing parallel physical-AI agents links reasoning and capability organization with sensing, dynamics, control, and real-world feedback as a pathway to evolving autonomy.Source: §10 · p. 19