ImpedanceDiffusion: Diffusion-Based Global Path Planning for UAV Swarm Navigation with Generative Impedance Control

Abstract: Safe swarm navigation in cluttered indoor environment requires long-horizon planning, reactive obstacle avoidance, and adaptive compliance. We propose ImpedanceDiffusion, a hierarchical framework that leverages image-conditioned diffusion-based global path planning with Artificial Potential Field (APF) tracking and semantic-aware variable impedance control for aerial drone swarms. The diffusion model generates geometric global trajectories directly from RGB images without explicit map construction. These trajectories are tracked by an APF-based reactive layer, while a VLM-RAG module performs semantic obstacle classification with 90% retrieval accuracy to adapt impedance parameters for mixed obstacle environments during execution. Two diffusion planners are evaluated: (i) a top-view long-horizon planner using single-pass inference and (ii) a first-person-view (FPV) short-horizon planner deployed via a two-stage inference pipeline. Both planners achieve a 100% trajectory generation rate across twenty static and dynamic experimental configurations and are validated via zero-shot sim-to-real deployment on Crazyflie 2.1 drones through the hierarchical APF-impedance control stack. The top-view planner produces smoother trajectories that yield conservative tracking speeds of 1.0-1.2 m/s near hard obstacles and 0.6-1.0 m/s near soft obstacles. In contrast, the FPV planner generates trajectories with greater local clearance and typically higher speeds, reaching 1.4-2.0 m/s near hard obstacles and up to 1.6 m/s near soft obstacles. Across 20 experimental configurations (100 total runs), the framework achieved a 92% success rate while maintaining stable impedance-based formation control with bounded oscillations and no in-flight collisions, demonstrating reliable and adaptive swarm navigation in cluttered indoor environments.

• The paper presents ImpedanceDiffusion, a framework combining image-conditioned diffusion modeling and variable impedance control for efficient, map-free UAV swarm navigation, achieving a 92% success rate in real-world tests.
• The system uses a two-planner structure: a top-view planner for smooth, direct trajectories, and an FPV planner offering higher local clearance and speed, both with reliable performance across various indoor scenarios.
• Experimental results show that the FPV planner improves velocity and clearance near human obstructions but introduces potential safety concerns in dynamic environments.

ImpedanceDiffusion is a hierarchical navigation framework for aerial drone swarms that replaces classical search-based global planning with an image-conditioned diffusion model, while retaining reactive Artificial Potential Field (APF) tracking and adding semantic-aware variable impedance control. The system is designed for map-free operation in cluttered indoor environments containing both rigid obstacles and humans, and is validated through zero-shot sim-to-real deployment on Crazyflie 2.1 drones.

Framework architecture

The framework comprises four modules: (1) a conditional UNet-based diffusion model that generates global trajectories directly from RGB images plus start/goal pixel inputs; (2) an APF local planner that tracks the diffusion trajectory on the leader drone by treating successive waypoints as temporary goals under combined attractive and repulsive forces; (3) a VLM–RAG module using Molmo-7B-O (4-bit quantized) to classify obstacle types from a top-view image, with FAISS-based retrieval of impedance parameters from a custom database; and (4) a virtual impedance-based formation controller for follower drones.

The diffusion planner operates in pixel space: the input $x_0$ is a three-channel mask encoding start pixel, goal pixel, and trajectory mask. Training uses a squared-cosine noise schedule with DDPM posterior sampling during inference, where start and goal channels are inpainted at each denoising step to enforce boundary consistency. The loss combines weighted path reconstruction (MSE over trajectory masks) and endpoint accuracy terms. Both planners are trained purely on A*-generated trajectories — 10,000 samples from ProcTHOR for the top-view planner and 13,000 from Matterport-derived indoor environments for the FPV variant — with best performance at 100 diffusion steps over 30 epochs. Notably, no fine-tuning is performed before real-world deployment.

Variable impedance control

The impedance layer distinguishes two mechanisms. Drone–drone formation impedance models leader–follower coupling as a mass–spring–damper system, with follower targets offset geometrically around the leader. Parameters are modulated near humans via hysteresis thresholds on minimum human distance, reverting to hard-obstacle defaults beyond an exit radius. Drone–obstacle interaction impedance activates within a deflection radius, computing penetration depth $\delta$ and its discrete-time derivatives to form a virtual normal force $$F_n = k_o \delta + d_o \dot{\delta} + m_o \ddot{\delta}$$, converted to a position correction via constant-acceleration integration. The authors concede that the acceleration term has limited influence due to small penetration variations but retain it for completeness.

The impedance parameter database was constructed from 200 real-world swarm flight experiments, storing per-obstacle-class virtual mass, stiffness, damping, separation distance, deflection limits, and path tolerance for five categories (cylinder, chair, trolley, gate, human). Soft obstacles receive lower stiffness and higher damping than rigid ones; gates share grouped parameters for coordinated passage.

Experimental results

Twenty configurations across eight scenario types were run five times each (100 total runs), yielding a 92% success rate, with failures attributed to hardware or communication issues rather than planning or control instability. The VLM–RAG module achieved 90% retrieval accuracy (18/20 correct classifications). Both planners achieved a 100% trajectory generation rate across all scenarios.

The comparative analysis between the two planners reveals a clear trade-off:

Metric	Planner 1 (top-view)	Planner 2 (FPV)
Inference time	1.4–2.5 s	2.5–3.4 s
Avg. collision ratio	0.348	0.246
Avg. cumulative turning	9.41 rad	11.03 rad
Tracking speed near hard obstacles	1.0–1.2 m/s	1.4–2.0 m/s

Planner 1 produces smoother, more direct trajectories with faster single-pass inference, while Planner 2 achieves greater local clearance despite being deployed cross-modally (trained on FPV data, inferred on top-view images via two-stage inference), which reduces APF-induced corrective forces and permits higher speeds. Path lengths and goal errors are comparable between planners.

A notable safety-relevant observation arises in dynamic scenarios: Planner 2 maintains approximately 1.6 m/s near a moving human versus Planner 1's 0.8 m/s. The paper acknowledges that higher velocity near moving humans reduces reaction margin and may increase risk — a tension between clearance-driven speed gains and safe human-robot interaction that the framework does not explicitly resolve. Impedance behavior itself behaved as intended: rigid compliance (small deflections) near hard obstacles, soft compliance (larger deflections, reduced speed down to ~0.6 m/s) near humans, with bounded, decaying oscillations and no in-flight collisions.

Limitations and open questions

Several limitations are stated explicitly. All perception, VLM inference, and diffusion inference run on an external RTX 4090 workstation, so onboard autonomy and decentralization remain unresolved. Impedance adaptation is discrete class-based switching rather than continuous policy learning, limiting scalability in dense heterogeneous scenes. The reported collision ratios measure overlap with inflated pixel-space safety regions rather than physical collisions, so they quantify plan-level clearance only. Top-view images were captured under consistent daylight conditions, leaving robustness to illumination variation untested. Finally, whether the higher speeds enabled by FPV conditioning remain acceptable near dynamic humans is left open.

Conclusion

ImpedanceDiffusion demonstrates that image-conditioned diffusion models can serve as reliable map-free global planners for swarm navigation when coupled with reactive APF tracking and semantically adaptive impedance control, achieving a 92% success rate over 100 real-world runs without any sim-to-real fine-tuning. The principal contribution beyond prior work such as ImpedanceGPT is obstacle-class-dependent compliance selected via VLM–RAG retrieval, enabling differentiated interaction behavior within a single trajectory through mixed environments. The remaining gaps — onboard computation, continuous impedance adaptation, and speed regulation near dynamic humans — define the immediate open questions for this line of work.