Locaris: Token-Based Indoor Localization

• Locaris is an indoor localization framework that uses tokenized Wi-Fi measurements and a decoder-only transformer to directly map raw telemetry to spatial coordinates.
• The system integrates low-rank LoRA modules with a compact MLP regression head, enabling efficient few-shot adaptation and robust performance across varying environments.
• Empirical results demonstrate sub-meter accuracy, resilience to missing access point data, and scalable, calibration-free deployment for diverse real-world applications.

Locaris is an indoor localization framework that utilizes a compact, telemetry-agnostic, transfer-learning enabled decoder-only transformer to map raw Wi-Fi signals directly to spatial coordinates. Distinct from fingerprinting and conventional model-based positioning approaches, Locaris abstracts each access point (AP) measurement as a token, allowing seamless ingestion of heterogeneous, incomplete telemetry such as RSSI and FTM. This enables calibration-free, robust regression from unstructured Wi-Fi data, offering scalable, cross-environment performance especially in scenarios where extensive calibration is infeasible (Bhatia et al., 13 Oct 2025).

1. Conceptual Framework and Problem Domain

Locaris addresses the inherent variability and sensitivity of indoor Wi-Fi positioning, which is affected by environmental dynamics, propagation characteristics, and hardware heterogeneity. Traditional localization methods require labor-intensive site calibration, fixed feature engineering, and degrade rapidly under changing conditions or hardware. Locaris introduces a decoder-only LLM paradigm, treating each AP measurement as an independent “token.” This modeling allows the system to natively process variable-length input sequences, accommodate missing APs, and exploit raw telemetry without preprocessing or imputation. The learned map from tokenized Wi-Fi signals to spatial position shifts indoor positioning from handcrafted pipelines to data-driven regression.

2. Model Architecture and Technical Approach

Locaris is instantiated using a pre-trained decoder-only transformer (e.g., LLaMA-3.2-1B), which remains largely frozen during adaptation. Only low-rank LoRA modules (inserted into the attention projections) and a compact multi-layer perceptron (MLP) regression head are fine-tuned for localization. Each AP measurement—irrespective of modality (RSSI, FTM, etc.)—is serialized as a token, and the entire sequence is fed into the model. After processing through $L$ transformer layers, the final hidden state $z = h_T$ of the last token summarizes the context, which is decoded to a 2D spatial location via

$\hat{y} = \mathrm{MLP}(z) \in \mathbb{R}^2$

The model is supervised by minimizing the mean squared error loss

$\mathcal{L} = \lVert y - \hat{y} \rVert^2$

where $y$ is the ground truth location.

During few-shot adaptation, only a small number of calibration samples are required to re-tune the LoRA layers and regression head, enabling rapid deployment to unseen devices or environments without comprehensive retraining.

3. Data Representation and Ingestion

Locaris employs schema-free, prompt-driven data representation. Measurements are ingested as tokenized sequences—e.g.,

AP1 RTT: <v> AP2 RTT: <v> ... AP5 RSS: <v>

The sequential ordering encodes potential relations such as multipath propagation and environmental dependencies. Variable-length input handling is intrinsic: missing APs or diverse modalities are naturally represented without padding. This approach permits direct modeling of environmental heterogeneity and signal irregularities encountered in realistic deployments.

4. Transfer Learning and Calibration Efficiency

Locaris capitalizes on transfer learning by fine-tuning a small parameter subset atop a general-purpose transformer. Unlike full model retraining, LoRA-driven adaptation enables the reuse of common signal patterns across devices, vendor hardware, and building layouts. This leads to rapid calibration with only a handful of labeled samples, minimizing human effort and reducing computational demands. Cross-environment generalization is empirically demonstrated: models trained in one setting retain high accuracy when ported, with sub-meter errors achievable after few-shot fine-tuning.

5. Empirical Evaluation and Performance

Locaris was systematically compared to k-nearest neighbor, feedforward MLP, and tree-based regression baselines across standard datasets (SODIndoorLoc, FTM-RSSI). The main findings include:

Metric	Locaris Performance	Baseline Performance
Mean Absolute Error	~2.84 m	3–4 m
Sub-meter Accuracy	Achievable with multi-modality	Unachievable (most baselines)
Robustness (Missing APs)	Maintains performance; graceful degradation	Rapid error increase
Worst-case Error	Lower 95th percentile error	Higher variability

Fusion of RSSI and FTM signals yields sub-meter error under favorable conditions. AP ablation experiments demonstrate robustness to dropped or missing measurement sources. In few-shot adaptation scenarios, Locaris achieves comparable accuracy to baselines that require substantially more calibration data, such as 50–75% coverage for non-transformer models.

6. Practical Implementation and Deployment Considerations

Locaris’s lightweight adaptation mechanism (fine-tuning LoRA modules and regression head) reduces both computational and memory cost, facilitating execution on resource-constrained platforms. The calibration-free design, prompt-driven input, and robustness to incomplete data enable wide applicability:

• Navigation in retail, healthcare, and transit environments
• Large-scale positioning without exhaustive site surveys
• Dynamic adaptation to hardware and layout changes

A plausible implication is that systems adopting Locaris methodology can generalize quickly across deployments, avoid rigid preprocessing, and integrate new telemetry modalities as needed. The parameter-efficient approach also positions Locaris for mobile or embedded deployments.

7. Future Directions

Potential future work highlighted includes:

• Extending Locaris to direct 3D localization tasks
• Incorporating additional sensor modalities (visual, inertial, environmental) for richer spatial context
• Scaling model size and dataset diversity to enhance generalizability
• Systematic analysis of transformer-based regression robustness for Wi-Fi and multi-modal positioning

This suggests continued research may further reduce calibration cost, expand environmental coverage, and improve positioning fidelity in adversarial or poorly instrumented scenarios (Bhatia et al., 13 Oct 2025).

---

Locaris establishes a technical foundation for calibration-free, adaptable indoor localization by leveraging token-based LLM architectures with transfer learning. Its empirical superiority and architectural flexibility make it a benchmark for future indoor positioning research.