Technical Summary: SkySense

1. Problem Statement

The paper addresses the critical challenge of localizing unauthorized, non-cooperative Unmanned Aerial Vehicles (UAVs) within 6G Integrated Sensing and Communication (ISAC) networks. While Channel State Information (CSI) fingerprinting offers a promising passive sensing solution, existing deep learning-based approaches face three fundamental limitations:

• Spatial Ambiguity: Distinct physical positions can induce highly similar CSI fingerprints due to multipath effects. Conventional discriminative models often fail here, yielding invalid "average" coordinates (mean collapse) when faced with overlapping mappings.
• Data Scarcity: Practical deployment is hindered by the extreme cost of acquiring massive, centimeter-level Real-Time Kinematic (RTK) labeled datasets required for fully supervised training.
• Inference Latency: While generative diffusion models can probabilistically resolve spatial ambiguities, their reliance on long-chain Markov iterative denoising processes creates prohibitive latency, making them unsuitable for highly dynamic, real-time UAV tracking.

2. Methodology: The SkySense Framework

The authors propose SkySense, an end-to-end semi-supervised generative framework designed to overcome these bottlenecks. The architecture operates in two distinct stages and supports distributed multi-base station (BS) fusion.

A. Self-Supervised Spatial Representation (Stage I)

To mitigate the dependency on labeled data, the framework exploits the kinematic continuity of UAV flight trajectories.

• Mechanism: A self-supervised encoder ($E_\theta$) is pre-trained on massive unlabeled CSI sequences.
• Objective: It employs a temporal contrastive learning strategy. By constructing a spatial similarity prediction matrix based on learned latent features and aligning it with a temporal target matrix (modeled via Gaussian decay based on timestamps), the encoder learns robust spatial representations without requiring physical coordinates. This ensures the model captures the underlying spatial manifold of the flight path.

B. Conditional Consistency Generative Backbone (Stage II)

To resolve spatial ambiguity and reduce latency, the framework utilizes a Consistency Model (a derivative of diffusion architectures) as the core generative backbone ($F_\phi$).

• Generative Formulation: Instead of modeling the full reverse diffusion process, the model learns a direct mapping from arbitrary noisy states to the clean physical origin. It parameterizes perturbation using a continuous noise scale $\sigma$.
• Training Objective: The backbone is jointly fine-tuned with the pre-trained encoder using a limited set of labeled data. The loss function ($L_{total}$) combines:
  1. Structural Consistency ($L_{cons}$): Ensures that predictions from different noise levels along the same trajectory map to the identical clean coordinate.
  2. Physical Anchor ($L_{anchor}$): Anchors the generative predictions to the scarce ground-truth coordinates.
• Inference Efficiency: Unlike traditional diffusion models requiring hundreds of steps, the consistency formulation allows the inference trajectory to be compressed to 1–2 steps, bypassing the iterative denoising bottleneck.

C. Distributed Multi-BS Fusion

The framework supports a lightweight distributed fusion mechanism. During online inference, multiple base stations independently generate coordinate proposals based on their local CSI. These predictions are aggregated via a spatial consensus mechanism (averaging) to form a unified estimate, effectively suppressing multipath ambiguity from a multi-view geometry perspective.

3. Key Contributions

• Semi-Supervised Generative Paradigm: The paper introduces a novel framework that combines self-supervised representation learning with conditional consistency generation, addressing both the data scarcity and spatial ambiguity issues simultaneously.
• Latency Optimization: By adopting a consistency model over standard diffusion models, the framework achieves real-time inference capabilities (1–2 steps) while maintaining the probabilistic benefits of generative modeling.
• Physics-Aware Pre-training: The self-supervised strategy leverages the physical principle that temporally adjacent observations are spatially proximate, providing a robust initialization that reduces reliance on labeled data.
• Distributed Fusion: A lightweight mechanism is proposed to aggregate predictions from multiple base stations without the prohibitive overhead of joint multi-BS training.

4. Experimental Results

The framework was evaluated on a real-world open-source ISAC UAV dataset containing 3-BS radar surveillance measurements in complex urban environments.

• Accuracy under Label Scarcity: With only 1% of labeled data (RTK coordinates), SkySense achieved a mean 3D localization error of 9.77 cm in a 3-BS fusion setup.
• Comparison with Baselines:
  • SkySense significantly outperformed fully supervised discriminative baselines (e.g., FC-AE-GPR) and semi-supervised discriminative baselines (e.g., SSLUL).
  • Under extreme label scarcity (0.1%), SkySense maintained decimeter-level accuracy, whereas discriminative models suffered from precision collapse.
  • Compared to standard Diffusion Probabilistic Models (DDPM), SkySense avoided catastrophic performance degradation (DDPM error >245 cm in semi-supervised settings) and eliminated the need for long inference chains.
• Ablation Studies:
  • Backbone Architecture: The U-Net variant of the backbone performed best, confirming the necessity of hierarchical structures for spatial feature preservation.
  • Consistency vs. Diffusion: The consistency objective proved critical; standard DDPM failed to establish robust mappings with limited anchors.
  • Multi-BS Fusion: Moving from a single-BS to a 3-BS setup reduced the error from ~20.67 cm to 9.77 cm, demonstrating the effectiveness of spatial consensus in resolving ambiguity.
• Inference Efficiency: SkySense maintained robust accuracy even when compressed to 1–2 inference steps, whereas truncating DDPM trajectories caused exponential error growth.

5. Significance and Claims

The authors claim that SkySense represents a significant advancement for practical ISAC deployments by solving the "trilemma" of data scarcity, spatial ambiguity, and inference latency.

• Practicality: The framework enables centimeter-level localization with minimal annotation effort (1% labels), making it feasible for scenarios where acquiring large labeled datasets is prohibitively expensive.
• Real-Time Capability: By reducing inference to 1–2 steps, the system becomes computationally viable for highly dynamic UAV tracking, a limitation of previous generative approaches.
• Robustness: The probabilistic modeling of the conditional distribution effectively resolves the mean collapse problem inherent in discriminative models when facing multipath spatial ambiguities.

The paper concludes that this approach provides a scalable foundation for ubiquitous UAV localization, with future work aiming to utilize synthetic datasets and Sim-to-Real transfer to further eliminate reliance on empirical measurements.