[Discussion] Localization in Open Spaces — How to Improve Accuracy in SLAM?

[EricChen0104]

Hi everyone!

I'm working on a Hybrid-SLAM project where a simulated robot explores an unknown maze environment using a full autonomy stack:

🧠 Architecture Overview

• SLAM: Occupancy grid map + localization
• Global Path Planner: [D* Lite]
• Local Planner: [Dynamic Window Approach (DWA)]
• User Interaction:
  • Left-click to set target goals
  • Right-click to dynamically insert obstacles (replanning happens in real-time)

It works quite well overall. Here's a brief summary of how the system behaves:

• The robot explores while building a map
• It uses D* Lite to plan long-term routes across the map
• DWA helps it handle short-term obstacle avoidance and dynamic changes

---

🆘 The Problem: Localization Fails in Open Spaces

In maze-like environments (with walls and corridors), the localization is stable and accurate.

However, when the robot enters wide open spaces (like rooms with few or no walls), the position estimate begins to drift significantly. Eventually, the robot gets completely lost or stuck in incorrect estimates.

---

❓What I've Tried

• Standard log-odds grid mapping
• A custom particle filter-based localization system
• Increasing sensor update rates
• Adding a small amount of motion noise

Still, the localization remains unstable in open spaces. I suspect it's due to a lack of structural features or poor data association.

---

🔍 What I'm Looking For

I'm wondering if any of you have faced similar issues. I'd really appreciate:

• General advice on localization in low-feature environments
• Recommended techniques (e.g., scan matching, visual landmarks, adaptive resampling)
• Papers or GitHub projects tackling this issue
• Feedback on how to combine multiple localization strategies

---

📎 Repo Link

GitHub Repo (Hybrid-SLAM)

---

Feel free to reply here, fork the repo, or even open PRs with suggestions. I'd love to improve this system and learn from the community. Thanks in advance!