LocDreamer: World Model-Based Learning for Joint Indoor Tracking and Anchor Scheduling

Accurate, resource-efficient localization and tracking enables numerous location-aware services in next-generation wireless networks. However, existing machine learning-based methods often require large labeled datasets while overlooking spectrum and energy efficiencies. To fill this gap, we propose LocDreamer, a world model (WM)-based framework for joint target tracking and scheduling of localization anchors. LocDreamer learns a WM that captures the latent representation of the target motion and localization environment, thereby generating synthetic measurements to imagine arbitrary anchor deployments. These measurements enable imagination-driven training of both the tracking model and the reinforcement learning (RL)-based anchor scheduler that activates only the most informative anchors, which significantly reduce energy and signaling costs while preserving high tracking accuracy. Experiments on a real-world indoor dataset demonstrate that LocDreamer substantially improves data efficiency and generalization, outperforming conventional Bayesian filter with random scheduling by 37% in tracking accuracy, and achieving 86% of the accuracy of same model trained directly on real data.

💡 Research Summary

LocDreamer is a novel framework that tackles two intertwined challenges in indoor wireless localization: accurate target tracking and energy‑efficient anchor scheduling. Traditional approaches either rely on large labeled datasets or assume that all anchors are active at every time step, leading to high power and spectrum consumption. LocDreamer leverages a world model (WM) in the form of a Deep State Space Model (DSSM) to learn a compact latent representation of target motion and the surrounding radio environment. The DSSM consists of (i) a recurrent neural network that maintains a deterministic hidden state, (ii) a dynamics module that combines a physics‑based motion model with a learnable residual MLP to predict the prior latent state, (iii) a set‑transformer encoder that infers the posterior latent state from distance measurements and anchor positions, and (iv) a decoder that reconstructs distances using the Euclidean geometry for the mean and an MLP for the noise variance. Training maximizes a variational evidence lower bound (ELBO), which balances reconstruction of synthetic measurements with regularization of latent dynamics.

Once the WM is pretrained on a well‑measured source environment, it can “imagine” measurements for arbitrary, unseen anchor configurations. These imagined measurements serve as a self‑supervised signal for two downstream components: (1) the DSSM tracker itself, which continues to refine its latent dynamics using the synthetic data, and (2) an actor‑critic reinforcement‑learning (RL) agent that learns an anchor‑scheduling policy. The RL policy receives the current latent prior and hidden state as its observation and outputs a binary activation vector for the anchors. The reward is defined as the log‑likelihood of the imagined measurements given the selected anchors, encouraging the agent to activate only the most informative anchors. Because the reward is computed on imagined data, the policy can be trained without any additional real‑world measurements, enabling rapid adaptation to new anchor layouts or environmental changes.

Experiments on a real‑world indoor tracking dataset evaluate LocDreamer against two baselines: a conventional Bayesian filter with random anchor activation, and the same DSSM trained directly on the real measurements (i.e., without imagination). Results show that LocDreamer improves tracking accuracy by 37 % over the Bayesian baseline while using far fewer active anchors. Compared with the fully supervised DSSM, the imagination‑driven version attains 86 % of the accuracy, demonstrating strong data efficiency and generalization.

Key contributions include: (i) a unified maximum‑likelihood objective that jointly optimizes tracking and scheduling, (ii) a WM‑based imagination pipeline that generates synthetic measurements for unseen anchor deployments, (iii) a hybrid physics‑plus‑learned dynamics model that captures complex indoor propagation effects, and (iv) an RL‑based scheduler that directly minimizes resource usage while maximizing measurement informativeness.

The paper also discusses limitations. The current implementation assumes a 2‑D planar scenario and uses only distance (range) measurements, leaving out richer modalities such as angle‑of‑arrival or RSSI. Moreover, the quality of the imagined data depends heavily on the pretraining dataset; bias in the source environment could propagate to the synthetic measurements and degrade performance in truly novel settings. Future work is suggested to extend the model to three‑dimensional spaces, incorporate multimodal sensor data, and develop online adaptation mechanisms that continuously refine the WM as new real measurements become available. Overall, LocDreamer showcases how world‑model imagination can dramatically reduce the need for extensive labeled data while delivering both high tracking precision and substantial energy savings in indoor wireless localization systems.