🤖 AI

Think over Trajectories: Leveraging Video Generation to Reconstruct GPS Trajectories from Cellular Signaling

This paper introduces Sig2GPS, a novel approach that reconstructs high-precision GPS trajectories from coarse cellular signaling data by reframing the problem as a map-visual video generation task, utilizing a specialized dataset and reinforcement learning to outperform existing methods in accuracy and scalability.

Original authors: Ruixing Zhang, Hanzhang Jiang, Leilei Sun, Liangzhe Han, Jibin Wang, Weifeng Lv

Published 2026-03-30

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Ruixing Zhang, Hanzhang Jiang, Leilei Sun, Liangzhe Han, Jibin Wang, Weifeng Lv

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

The Big Problem: The "Fuzzy" Phone Signal

Imagine you are trying to track a friend's journey across a city. You have a map, but instead of seeing their car moving smoothly along the roads, you only see a series of blurry, glowing dots. These dots represent the cell towers their phone connected to.

The Reality: Cell towers cover huge areas (like a giant umbrella). If your phone connects to "Tower A," you know your friend is somewhere under that umbrella, but you don't know if they are on Main Street or in the park next door.
The Goal: We want to turn those fuzzy, disconnected dots into a smooth, continuous line that shows exactly where they drove, just like a high-precision GPS.

The Old Way: The "Robot Accountant"

Traditionally, engineers tried to solve this by building a complex, multi-step machine (a "pipeline").

Step 1: Clean up the noise (remove "ping-pong" effects where the phone jumps between towers).
Step 2: Match the dots to the nearest road on a map.
Step 3: Guess the route based on traffic rules and probability.

The Flaw: This is like trying to solve a puzzle by measuring every single piece with a ruler, calculating the angles, and then gluing them together. It's slow, requires a lot of human tweaking, and often breaks if the city layout changes. It's very "math-heavy" but lacks "common sense."

The New Idea: "Think Over Trajectories" (The Artist)

The authors of this paper had a different idea. They noticed that when a human expert looks at those fuzzy dots on a map, they don't do math. They just draw. They look at the dots and say, "Oh, the person must have turned left here and gone down that street," and they sketch the line.

They asked: "Can we teach a computer to 'draw' the route just like a human artist?"

Instead of calculating coordinates (X, Y), they treated the problem as Video Generation.

The Input: A static image of a map with the fuzzy cell tower dots drawn on it.
The Task: Ask an AI video generator to "animate" a car driving from the start to the end, following the dots.
The Output: A short video where a line is drawn, showing the exact path the car took.

How They Taught the AI (The Training)

To teach the AI to do this, they used a two-step process:

1. Supervised Fine-Tuning (The Art Class)
They showed the AI thousands of examples where they had both the fuzzy dots and the real GPS path. The AI learned to mimic the "drawing" style of the real paths. It learned that if dots are close together, the car is likely moving slowly, and if they are far apart, the car is speeding.

2. Reinforcement Learning with Rewards (The Critic)
Just watching videos isn't enough; the AI might draw a path that looks good but is physically impossible (like driving through a building). So, they added a "Critic" system (called Traj-GDPO).

Distance Reward: "Did the car end up near the right place?"
Direction Reward: "Did the car drive forward, or did it magically reverse?"
Continuity Reward: "Is the line broken, or is it one smooth path?"

If the AI draws a bad path, the Critic gives it a low score. If it draws a good path, it gets a high score. The AI learns to maximize its score, effectively "polishing" its drawing skills until the path is perfect.

Why This is a Game-Changer

Speed: The old "Robot Accountant" method took minutes to process one trip because it had to do so many calculations. The new "Artist" method generates the whole path in one go, in seconds.
Common Sense: Because the AI is "drawing" on a map, it naturally understands road rules. It won't draw a car driving through a river or a building because the video model has learned what roads look like.
Flexibility: It works in different cities without needing to be reprogrammed. If you show it a map of a new city, it just "draws" the path based on the roads it sees, just like a human would.

The Result

The paper shows that this new method is much more accurate than the old industrial methods. It successfully turns "fuzzy cell tower guesses" into "crystal clear GPS routes."

In a nutshell: Instead of trying to solve a complex math equation to find a route, the researchers taught an AI to watch a map and draw the route, using a "critic" to make sure the drawing makes sense. It's a shift from "calculating" to "visualizing."

1. Problem Definition: Sig2GPS

The paper addresses the Sig2GPS problem: reconstructing high-precision GPS trajectories from coarse cellular signaling data.

Input: Cellular signaling records, which consist of timestamps and serving base station identifiers (cell IDs). These records offer broad coverage but low spatial resolution (typically hundreds of meters to kilometers).
Output: Fine-grained, continuous GPS trajectories (latitude/longitude coordinates) that adhere to road network topology.
Challenges:
- Coarse Resolution: Signaling data only indicates the serving cell, not the precise location.
- Complex Pipelines: Traditional industrial solutions rely on multi-stage engineering workflows (ping-pong effect mitigation, map matching, route inference) that are latency-heavy, brittle, and require extensive expert heuristics.
- Discrepancy: Existing learning-based methods often treat trajectories as numeric sequences or discrete tokens, failing to capture the spatial continuity and topological constraints inherent in map-based reasoning.

2. Methodology: The "Think Over Trajectory" Paradigm

The authors propose a novel paradigm that reframes trajectory reconstruction as an Image-to-Video Generation task, mimicking how human experts visualize signaling traces on a map to sketch routes.

A. Core Concept

Instead of regressing coordinates directly, the model operates in the map-visual domain:

Conditioning Input ( $c$ ): A map tile (from OpenStreetMap) overlaid with the cellular signaling trace rendered as a polyline.
Generation Target ( $y$ ): A video sequence where the ground-truth GPS path is progressively "drawn" onto the map frame-by-frame.
Model Architecture: A flow-based video generation model (specifically Wan2.2-TI2V-5B) is used. This architecture treats generation as integrating a learned velocity field over a continuous time variable.

B. Two-Stage Training Pipeline

The framework employs a two-stage training strategy to ensure both visual coherence and trajectory accuracy:

Stage 1: Supervised Fine-Tuning (SFT)

Goal: Initialize the model to understand the mapping between signaling traces and GPS paths.
Data: A paired dataset of ~20,000 signaling-taxi pairs (matched via spatiotemporal consistency) is converted into (Map+Signal, GPS-Video) pairs.
Objective: Standard conditional flow-matching loss ( $\ell_2$ regression on velocity fields) to minimize pixel-level reconstruction error.

Stage 2: Trajectory-Aware Reinforcement Learning (Traj-GDPO)

Motivation: SFT often results in visually plausible but topologically incorrect paths (e.g., wrong turns, direction reversals).
Mechanism: Reinforcement Learning from Verifiable Rewards (RLVR) is applied to optimize fine-grained trajectory criteria.
Reward Design: Three verifiable rewards are computed on the generated video:
1. Distance Reward ( $R_{dist}$ ): Measures geodesic error between generated and ground-truth points at key frames.
2. Direction Reward ( $R_{dir}$ ): Ensures the overall displacement vector aligns with the ground truth (preventing direction reversals).
3. Continuity Reward ( $R_{cont}$ ): A hard constraint ensuring the generated path is a single connected component with exactly two endpoints.
Optimization Algorithm (Traj-GDPO):
- To handle heterogeneous reward scales, the authors propose Trajectory-aware Group Decoupled Policy Optimization.
- It decouples reward components, normalizes them within a group of rollouts (Group Relative), and then aggregates them.
- KL Regularization: Crucially, the KL penalty is anchored to the fixed SFT initialization policy ( $p_{\theta_{initial}}$ ) rather than the previous policy. This prevents "mode collapse" and ensures the model does not drift away from the supervised solution manifold during RL.

3. Key Contributions

New Paradigm: Introduces "Think Over Trajectory," the first attempt to leverage video generation for trajectory data mining. It shifts the reasoning space from numeric sequences to visual map drawing.
Methodological Innovation: Proposes Traj-GDPO, a specialized RL fine-tuning method that uses verifiable, multi-component rewards to optimize video generation for spatial constraints.
Efficiency & Scalability: Demonstrates a "one-step" generation paradigm that eliminates complex multi-stage engineering pipelines, significantly reducing inference latency.
Cross-City Transferability: Shows that the visual representation of road networks allows the model to generalize across different cities without retraining on specific city data.

4. Experimental Results

Experiments were conducted on large-scale real-world datasets (Beijing for Sig2GPS; Chengdu and Xi'an for Next GPS Prediction).

Sig2GPS Performance:
- The proposed method significantly outperformed strong baselines, including deep learning models (GRU, Transformer, SigFormer) and a production-grade industrial baseline (Rule $_{sig}$ ).
- Metrics: Achieved the lowest MAE (214.96m vs. 306.12m for Rule $_{sig}$ ) and RMSE, and the highest proportion of points within 100m error (L100: 36.71% vs. 11.56%).
- Efficiency: Reduced inference time from >2 minutes (industrial pipeline) to <30 seconds per trajectory.
Next GPS Prediction:
- The model was extended to predict future GPS points given past history.
- It outperformed state-of-the-art methods (DeepMove, AgentMove, TrajMLLM) in both Chengdu and Xi'an, demonstrating superior modeling capacity for spatial continuity.
Cross-City Transfer:
- A model trained on Xi'an data achieved lower RMSE on the Chengdu test set than all baselines trained specifically on Chengdu, proving strong transferability.
Ablation Studies:
- Removing the RL stage (SFT only) degraded performance, confirming that pixel-level loss is insufficient for topological correctness.
- The Direction Reward was identified as the most critical component for preventing directional errors.
- Anchoring the KL term to the fixed SFT policy was essential to prevent generation collapse.

5. Significance and Impact

Practical Utility: The approach offers a practical, scalable interface for trajectory data mining, transforming low-quality signaling data into high-value GPS trajectories without complex engineering pipelines.
Generalization: By operating in the visual map domain, the model implicitly learns road network constraints (topology, driving rules) that are transferable across different geographic regions.
Future Directions: This work suggests that video generation models can serve as powerful engines for spatial reasoning tasks, potentially extending to other mobility analytics like trajectory recovery, anomaly detection, and long-horizon forecasting.

In summary, the paper successfully demonstrates that reframing trajectory reconstruction as a visual video generation problem, combined with verifiable reward-based RL, yields state-of-the-art accuracy and efficiency, surpassing both traditional engineering pipelines and previous deep learning approaches.