⚡ electrical engineering

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

LocDreamer is a world model-based framework that improves data efficiency and resource management in indoor localization by using synthetic measurements to jointly train a target tracking model and a reinforcement learning-based anchor scheduler.

Original authors: Geng Wang, Zhouyou Gu, Shenghong Li, Peng Cheng, Jihong Park, Branka Vucetic, Yonghui Li

Published 2026-02-10

📖 3 min read☕ Coffee break read

CC BY 4.0

Original authors: Geng Wang, Zhouyou Gu, Shenghong Li, Peng Cheng, Jihong Park, Branka Vucetic, Yonghui Li

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are playing a high-stakes game of Hide and Seek in a massive, dark mansion.

You are the "Seeker" (the tracking system), and you are trying to find a "Hider" (the target) who is moving around. To find them, you have several "Flashlights" (the anchors) placed around the house.

The Problem: The Flashlight Dilemma

In a perfect world, you would turn on every single flashlight at once to see exactly where the Hider is. But there’s a catch:

Battery Life: Turning on all the lights drains your batteries instantly.
Distraction: If too many lights are flashing, it creates a chaotic glare (signal interference), making it harder to see clearly.
The Unknown: What if you move to a new mansion where the lights are in different spots? You don't have a manual for this new house, and you can't afford to waste time testing every light.

Current technology usually either uses too much energy (all lights on) or uses "dumb" logic (turning on lights at random), which often misses the Hider entirely.

The Solution: "LocDreamer" (The Mental Simulator)

The researchers created LocDreamer. Instead of just reacting to what it sees, LocDreamer has a "Dream Engine" (a World Model).

Think of LocDreamer as a Seeker who has spent enough time in one house to understand the "rules" of movement. It knows that people don't usually teleport through walls and that they tend to walk in certain patterns.

Because it understands these rules, it can close its eyes and "dream" about the new mansion. It imagines: "If I were in that new room, and I turned on Flashlight A and Flashlight B, what would the Hider's shadow look like?"

How it Works (The Three Steps)

The Learning Phase (The Memory): First, the system watches a little bit of real movement in a known area. It learns the "physics" of the world—how things move and how light bounces off walls.
The Dreaming Phase (The Imagination): Now, here is the magic. Even if you move to a brand-new building with a totally different layout, LocDreamer doesn't panic. It uses its "Dream Engine" to simulate thousands of fake scenarios. It "imagines" different ways the Hider might move and different ways the new lights might shine. It practices in its head!
The Smart Selection (The Pro Scheduler): While dreaming, it learns a strategy: "I don't need all five lights. If I just use Light #1 and Light #4, I get a perfect view of the hallway. I'll keep the others off to save battery."

Why This Matters (The Results)

By "practicing in its dreams," LocDreamer becomes incredibly efficient:

It’s a Battery Saver: It only turns on the most useful "flashlights," saving energy and reducing signal clutter.
It’s a Fast Learner: It can walk into a new environment and be an expert almost immediately, without needing a human to teach it every single corner of the room.
It’s Highly Accurate: In tests, it was 37% more accurate than traditional methods that just picked lights at random, and it performed almost as well as systems that had "real-world" experience in that specific room.

In short: LocDreamer is like a professional athlete who can win a game in a brand-new stadium because they have spent so much time practicing in their mind.

Technical Summary: LocDreamer

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

1. Problem Statement

Next-generation wireless networks require accurate and resource-efficient indoor localization and tracking. Existing approaches face two primary challenges:

Resource Inefficiency: Most tracking methods assume all available anchors are active simultaneously, leading to excessive energy consumption, spectrum usage, and signaling overhead.
Poor Generalization and Data Dependency: Machine learning (ML) models for tracking and anchor scheduling typically require massive labeled datasets specific to a particular environment. When anchor configurations change or the environment shifts, these models fail to generalize without extensive new data collection.
Static Modeling Limitations: Traditional Bayesian filters (like EKF) rely on handcrafted mathematical assumptions that struggle to adapt to complex, non-linear indoor dynamics (e.g., multipath and NLoS).

2. Methodology

The authors propose LocDreamer, a framework that leverages a World Model (WM) to enable "imagination-driven" learning. Instead of relying solely on real-world measurements, the system learns the underlying dynamics of the environment and "dreams" synthetic measurements to train its components.

A. Deep State Space Model (DSSM) - The World Model:
The backbone is a DSSM that captures the latent dynamics of the target and the environment. It consists of:

Sequence Model: An RNN that maintains a deterministic hidden state ( $h_t$ ) to capture temporal dependencies.
Dynamic Model: A hybrid model combining a physics-based motion model with an MLP-based learned residual to predict the next latent state ( $z_t$ ).
Encoder Model: Uses a Set Transformer to infer the posterior latent state from noisy measurements, ensuring permutation invariance regarding the number/order of active anchors.
Decoder Model: Reconstructs distance measurements by combining a physics-based distance formula with a learnable noise model.

B. Reinforcement Learning (RL) Anchor Scheduler:
An Actor-Critic (AC) agent is trained to select the most "informative" subset of anchors ( $K_t$ ).

The Actor outputs a binary scheduling policy ( $\alpha_t$ ).
The Critic evaluates the policy based on a reward derived from the reconstruction loss of the DSSM. High reward is given to anchor selections that allow the DSSM to most accurately reconstruct the measurements (implying high-quality, informative data).

C. Two-Stage Training Process:

DSSM Pre-training (Source Environment): The DSSM is trained on a small set of real measurements from a known anchor configuration to learn the fundamental physics and dynamics of the system.
Imagination-Driven Joint Training (Unseen Environment): The DSSM is used to "imagine" trajectories and synthetic measurements for entirely new anchor deployments. The DSSM and the RL scheduler are then fine-tuned together using these synthetic "dreams," allowing the system to adapt to new configurations without new real-world data.

3. Key Contributions

Unified Framework: Formulates joint tracking and scheduling as a single maximum likelihood estimation problem optimized via a World Model.
Imagination-Driven Learning: Introduces a method to train both tracking and scheduling models in unseen environments using synthetic data generated by the DSSM, significantly improving data efficiency.
Self-Supervised Adaptation: Enables the system to generalize to new anchor deployments without requiring additional site surveys or real-world measurements in the new environment.

4. Experimental Results

The framework was evaluated using a real-world UWB indoor dataset. Key findings include:

Superior Accuracy: LocDreamer (with scheduling) outperformed conventional EKF with random scheduling by 37% in tracking accuracy.
High Generalization: When tested on unseen anchor configurations, LocDreamer achieved 86% of the accuracy of a model trained directly on real data from those same anchors.
Physical Consistency: The learned scheduling policy naturally prioritized anchors with better geometric diversity (lower GDOP), proving the model learned meaningful physical principles through imagination.
Efficiency: The model successfully reduced the number of active anchors (from all available to a subset $K_t=3$ ) while maintaining high performance.

5. Significance

LocDreamer represents a significant step toward AI-native 6G networks. By decoupling the learning process from the need for constant real-world data collection, it provides a scalable solution for massive IoT deployments. It bridges the gap between the high accuracy of deep learning and the resource constraints of practical wireless hardware, offering a path toward autonomous, self-configuring localization systems in complex, dynamic environments.