💻 computer science

Multi-Session Ground Texture SLAM in Low-Dynamic Environments

This paper investigates techniques for multi-session ground texture SLAM in low-dynamic environments, demonstrating that Kullback-Leibler Divergence is the most effective method for improving trajectory accuracy while also introducing a new dataset with multi-session ground texture images and high-accuracy pose annotations.

Original authors: Kyle M. Hart, Brendan Englot

Published 2026-05-20

📖 6 min read🧠 Deep dive

CC BY 4.0

Original authors: Kyle M. Hart, Brendan Englot

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 Picture: A Robot That Gets Lost in a Changing Room

Imagine a robot that needs to navigate a large warehouse. Usually, robots use "landmarks" to know where they are, like spotting a red fire hydrant, a tall bookshelf, or a specific painting on the wall.

But what if the robot is in a giant, empty room with no walls, no furniture, and no decorations? The only thing it can see is the floor. In this case, the robot has to navigate by looking at the texture of the carpet or concrete beneath its wheels.

The problem? Floors change over time.

The "Low-Dynamic" Change: Imagine the floor is a carpet. Over months, people walk over it, vacuuming it, or spilling coffee on it. The carpet doesn't change instantly (like a person walking across the room), but it slowly wears down, gets dusty, or gets patched up.
The Robot's Dilemma: If the robot remembers the floor from last month, it might get confused today because the carpet looks different. It might think, "I've never seen this patch of carpet before!" and get lost.

This paper is about teaching a robot how to keep its map accurate even when the floor slowly changes appearance over time.

The Experiment: The "Tape" Test

To test their ideas, the researchers built a special dataset. They didn't just use a normal floor; they simulated years of wear and tear in a few weeks.

The Setup: They had a robot with a camera pointing straight down at a carpeted floor.
The Trick: They ran the robot for a "session" (a trip across the room). Then, they put strips of tape all over the floor to simulate wear and tear.
The Repeat: They ran the robot again. Then added more tape. Then ran it again.
The Goal: Could the robot realize, "Oh, that patch of carpet I saw in Session 1 is the same spot as in Session 5, even though it's now covered in tape?"

They also used a high-tech motion-capture system (like the ones used in movies to track actors) to know the exact true location of the robot, so they could grade the robot's performance.

The Three Solutions Tested

The researchers tried three different "brain tricks" to help the robot recognize the floor despite the changes.

1. The "Visual Overlap" Check (Method 2)

The Idea: Before the robot tries to match a new picture of the floor to an old one, it checks: "Is it even physically possible for these two pictures to be looking at the same spot?"
The Analogy: Imagine you are looking at a map. If you are facing North and your friend is facing South, you can't be looking at the same tree. This method is like the robot checking its compass to see if the view could overlap.
The Result: It helped a little, but not enough. The robot still got confused by the tape.

2. The "Symmetry" Check (Method 3)

The Idea: This method looks at the "pattern" of light and dark spots in two images. If the floor hasn't changed much, the pattern of light and dark should be symmetrical (balanced) between the two images.
The Analogy: Think of a mirror. If you hold up a photo of a face next to its reflection, the features line up perfectly. If the face has a big scar (the tape), the symmetry breaks. The robot checks if the "mirror image" of the floor looks balanced.
The Result: This didn't work very well. The floor changes were too subtle for this math to catch.

3. The "Kullback-Leibler Divergence" (KLD) Check (Method 1)

The Idea: This is the winner. Instead of looking at specific shapes, the robot looks at the overall color distribution of the floor. It creates a "color recipe" (a histogram) for the floor from the very first day.
The Analogy: Imagine you have a jar of M&Ms. On Day 1, you count the colors: 50% Red, 20% Blue, 30% Green. This is your "Baseline Recipe."
- On Day 5, you take a new photo of the floor. You count the colors again. Maybe now it's 45% Red, 25% Blue, 30% Green because some red carpet wore away.
- The robot calculates the difference between the Day 1 recipe and the Day 5 recipe.
- The Magic: If the difference is small, the robot says, "Okay, this is the same floor, but it's a little worn. I'll trust this match, but I'll be a little cautious." If the difference is huge (like if the floor was repainted), the robot says, "This is too different! I don't trust this match."
The Result: This was the most successful method. It allowed the robot to keep its map accurate even as the floor changed significantly.

Why This Matters (According to the Paper)

Robots Can Stay Longer: Without this trick, a robot would need a human to come in and manually reset its map every time the floor got a little dirty or worn. With this trick, the robot can operate for a "lifetime" without human help.
It's Fast: The researchers checked how long the robot took to think. Adding this "color recipe" check didn't slow the robot down at all. It was like adding a tiny, instant calculation that didn't require extra brainpower.
It Catches Mistakes: The paper showed that the old way of doing things made many "false friends" (thinking it was in the right place when it wasn't). The new KLD method stopped these mistakes.
Bonus Feature: The researchers found they could use the "difference score" to make a Heat Map. If a specific spot on the floor has a high "difference score," it means that spot is very worn. This could help building managers know exactly where the floor needs repair.

Summary

The paper introduces a new way for robots to navigate using only the floor texture. By using a statistical method called Kullback-Leibler Divergence (which compares the "color recipe" of the floor over time), the robot can tell the difference between "this is the same floor, just a bit dirty" and "this is a totally different floor." This allows the robot to keep working accurately for months or years without needing a human to fix its map.

Technical Summary: Multi-Session Ground Texture SLAM in Low-Dynamic Environments

Problem Statement
Monocular visual Simultaneous Localization and Mapping (SLAM) is widely used for ground robot navigation, typically relying on outward-facing features like buildings or vegetation. However, in environments with few external features (e.g., large warehouses), robots must rely on the unique texture of the ground beneath them. While existing ground texture SLAM systems function well in static environments, they struggle with low-dynamic changes—gradual alterations in surface appearance such as wear, repair, or seasonal shifts that occur between operational sessions. Without the ability to adapt to these changes, robots require manual map reinitialization, limiting their autonomy and operational lifespan. Furthermore, no existing systems or datasets specifically address multi-session operations where ground texture evolves over time.

Methodology
The authors propose an investigation into three specific techniques designed to improve loop closure performance in a ground texture SLAM framework (based on Hart et al. [10]) when operating across multiple sessions with changing ground textures. The core goal is to adjust the confidence of loop closure factors added to the pose graph to prevent erroneous corrections caused by texture changes.

Kullback-Leibler Divergence (KLD): This method treats the ground texture as a probability distribution. During the first session, the system builds a baseline histogram of pixel intensities for each color channel. In subsequent sessions, the KLD score is calculated between the current image's intensity distribution and the baseline. This score is used to bias the covariance matrix of potential loop closures: higher KLD scores (indicating significant texture change) result in higher covariance (lower confidence), thereby down-weighting the influence of potentially mismatched loop closures during optimization.
Visual Overlap: This approach uses estimated robot poses and camera fields of view to project rectangles onto the ground plane. A loop closure candidate is only considered valid if the projected field of view of the current observation overlaps with that of the candidate. This aims to filter out candidates that are geometrically unlikely to share visual data.
Joint Intensity Histogram Symmetry: This method computes a 2D joint intensity histogram between a current image and a candidate image. It evaluates the symmetry of this histogram; similar textures yield symmetric histograms, while changing textures yield asymmetric ones. This symmetry score is used to bias the loop closure covariance, similar to the KLD approach.

Key Contributions

First-of-its-Kind Analysis: The paper presents the first investigation of multi-session, low-dynamic-change ground texture SLAM. It evaluates three distinct methods for handling texture evolution, identifying Kullback-Leibler Divergence as the most effective technique for maintaining trajectory accuracy.
New Dataset: The authors introduce a novel dataset specifically for this domain. It features:
- Multiple operational sessions where the ground texture undergoes low-dynamic changes (simulated by progressively taping a carpeted floor to mimic wear).
- High-accuracy ground truth pose data captured via a motion capture system, enabling objective evaluation of SLAM accuracy rather than relying on relative comparisons.
Environmental Monitoring Tool: The paper demonstrates that the KLD score can serve as a secondary utility for environment maintainers, generating "wear heat maps" to identify areas of significant surface degradation.

Results
The methods were evaluated using the new dataset, comparing pose estimation accuracy (Root Mean Square Error) against ground truth.

Accuracy: The KLD method (both color and grayscale versions) significantly outperformed all other approaches. It achieved a position RMSE of 0.086 m and orientation RMSE of 1.572 degrees. In contrast, the baseline system treating all sessions as a single continuous run (Original Single) failed as textures changed, resulting in a position RMSE of 0.540 m. The Visual Overlap and Joint Intensity Histogram methods also underperformed compared to KLD.
Loop Closure Quality: Analysis revealed that the KLD method effectively eliminated false positive loop closures (erroneous matches) that plagued the baseline system. While it did exclude some true loop closures (false negatives), the excluded candidates were, on average, less accurate than those retained, leading to a net improvement in map quality.
Computational Efficiency: The KLD approach added negligible computational overhead. The scoring process is $O(1)$ with respect to the number of stored observations, resulting in processing times comparable to the original single-session system.
Baseline Comparison: Standard forward-looking SLAM systems (e.g., ORB-SLAM3) failed to produce continuous trajectories in this ground-texture-only environment, highlighting the necessity of domain-specific solutions for downward-facing cameras.

Significance
The paper claims that by integrating Kullback-Leibler Divergence to bias loop closure confidence, robots can operate autonomously for longer durations in ground-texture-dependent environments without user intervention, even as the ground surface degrades. The work establishes a new benchmark for multi-session ground texture SLAM through the release of a specialized dataset and demonstrates that adapting statistical similarity metrics can effectively mitigate the impact of low-dynamic environmental changes. Additionally, the derived wear heat maps offer a practical, non-intrusive method for monitoring the health of operational surfaces.