⚡ electrical engineering

Calibration-Free Floor Detection via Gravity-Aware SOAP Descriptors: Theory and Embedded Deployment for Reactive Humanoid Locomotion and Obstacle Avoidance

This paper introduces a calibration-free, gravity-aware SOAP descriptor that achieves superior floor detection accuracy (90.0% F1) and enables real-time reactive locomotion and obstacle avoidance on a humanoid robot by leveraging IMU-aligned geometry without requiring labeled data or environment-specific parameters.

Original authors: Eiaki Von Roeder Morooka, Natalia Kamysheva, Patrik Asikainen, Yuto Omae, Hirotaka Takahashi

Published 2026-08-12

📖 9 min read🧠 Deep dive

CC BY 4.0

Original authors: Eiaki Von Roeder Morooka, Natalia Kamysheva, Patrik Asikainen, Yuto Omae, Hirotaka Takahashi

Original paper licensed under CC BY 4.0 (https://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 a robot trying to walk through a room it has never seen before. It doesn't have a map, and it hasn't been trained on pictures of that specific room. To move safely, it needs to answer one simple question instantly: "Is the ground under my feet, or is that a wall, a table, or a chair?" This is the challenge of calibration-free floor detection. In the world of robotics, "calibration" usually means a human has to spend time teaching the robot the specific height of the floor or the layout of the room. But for a robot to be truly autonomous, it needs to walk into any new space and know where the floor is immediately, without a tutorial.

To solve this, robots often use LiDAR, a sensor that shoots out laser beams to create a 3D map of the world, like a bat using sonar but with light. The robot then has to look at this cloud of laser points and figure out which ones belong to the flat, walkable floor and which ones belong to vertical walls or random objects. Traditionally, robots have used simple math tricks, like checking if a group of points is flat or if they are at a certain height. However, these tricks can get confused by messy data, like sparse laser hits or weird angles. This paper explores a new way to solve this puzzle by borrowing a tool from a completely different field: chemistry. Specifically, it uses a method called SOAP (Smooth Overlap of Atomic Positions), which was originally designed to understand how atoms clump together in molecules. The researchers asked: "Can we use this 'atomic' math to understand the 'atomic' points in a robot's laser scan?"

The researchers, working with a humanoid robot named the Unitree G1, discovered that they could adapt this chemistry tool to act as a "gravity-aware" detector. Instead of just looking at the shape of the points, they aligned the math with the direction of gravity (using the robot's built-in tilt sensor). This allowed the robot to distinguish between a flat floor and a vertical wall with surprising accuracy, even without any prior training or maps. They found that this method, when combined with a standard geometric check, could identify the floor in 90% of cases on a difficult benchmark dataset, beating the previous best "no-training" method. They also proved they could run this complex math on the robot's own small computer chip in real-time, fast enough to stop the robot before it bumped into anything.

The Chemistry of Walking

The core idea of this paper is a bit like teaching a robot to "feel" the floor using a sense it didn't have before. The authors took a mathematical descriptor called SOAP, which was invented to describe how atoms sit next to each other in a crystal or a molecule. In chemistry, this helps scientists predict how a material will behave. The researchers realized that a 3D laser scan of a room is just a giant cloud of points, similar to a cloud of atoms. They decided to treat the robot's laser points as if they were atoms and used the SOAP math to describe the "neighborhood" around each point.

However, there was a catch. The original SOAP math is designed to be totally rotation-proof; it doesn't care which way is up. But for a robot walking, knowing which way is "up" is everything. A floor is flat relative to gravity, while a wall is vertical. If the math doesn't know which way is up, it can't tell the difference between a floor and a wall that happens to be rotated.

To fix this, the team created a gravity-aware version of SOAP. They used the robot's IMU (a sensor that measures tilt and acceleration, like the one in your phone) to align the math with the real world's "up" and "down." This allowed the descriptor to specifically look for points that are arranged horizontally, like a floor, while ignoring points arranged vertically, like a wall.

The "Magic Score"

The result of this math is a "score" for every single point in the laser scan.

If the points around a spot look like a flat, horizontal sheet, the score is high.
If the points look like a vertical wall or a messy pile of furniture, the score is low.

The authors found that this score was incredibly good at spotting the floor. On a standard test dataset called S3DIS (which contains 68 different rooms), their method correctly identified the floor 90.0% of the time. This was better than the previous best method that didn't require training data, which only got 86.6% right. The key to this success was that the SOAP method is very good at handling "sparse" data. If the laser only catches a few points on a floor near a table leg, the old methods often get confused and think it's a wall. The SOAP method, because it uses a smooth "cloud" of probability rather than just counting neighbors, stays calm and correctly identifies the floor even when the data is a bit messy.

Running on a Robot's Brain

One of the biggest hurdles for this kind of math is speed. Doing complex chemistry-style calculations on a robot's tiny computer usually takes too long. If the robot has to wait a second to figure out where the floor is, it might already have walked into a wall.

The team solved this by writing a super-fast version of the math using PyTorch (a popular AI software library) that runs directly on the robot's NVIDIA Jetson Orin NX chip. They managed to make it run 20 times faster than a standard computer program.

Speed: It takes only 0.08 seconds to process a full scan of the room.
Frequency: This is fast enough to update the robot's "vision" 2 times every second (2 Hz).

This speed allowed them to deploy the system on a Unitree G1 EDU humanoid robot. This robot is about 1.3 meters tall and weighs 35 kg. The robot uses this "floor detector" as its primary way to avoid obstacles. When the robot hears a voice command like "walk forward for 30 seconds," it starts moving. Every time it gets a new laser scan, it checks the SOAP score. If the path ahead is clear, it keeps walking. If the score drops (meaning the floor is blocked or it's about to walk off a cliff), it stops and turns 90 degrees to look for a new path.

Real-World Tests and "Ghost" Obstacles

The researchers tested their robot in real life with several challenges:

Clear Paths: The robot walked 5 meters down a clear hallway 9 times and never crashed.
Obstacles: They placed a wicker basket and an office chair in the robot's path. The robot detected them and turned around to avoid them.
The "Glass" Test: They tested the robot against glass partitions. Glass is tricky because lasers often pass right through it, leaving no points in the scan. The robot's logic was: "If I see no points where I expect to see the floor, I must assume there is a wall or glass there." This "no-data is a block" rule saved the robot from walking into invisible glass.
The "Thin" Failure: The system isn't perfect. They tried to trick the robot with a thin, translucent plastic spray bottle. Because the bottle is see-through and thin, the laser only hit it a few times. The robot didn't see it until it actually bumped into it. The authors admit this is a limitation: very thin, transparent objects that don't reflect enough laser light can slip past the detector.

Why This Matters

This paper is significant because it's the first time this specific "chemistry" tool (SOAP) has been used as a real-time "eyes" for a humanoid robot. It proves that you don't need a massive database of labeled photos or a pre-built map to walk safely. You just need to know which way is up and use the right math to look at the world.

The authors also showed that this single math tool does double duty. It doesn't just find the floor; it can also tell the difference between a table, a wall, and a drop-off (a cliff). This creates a rich "geometric map" of the room that the robot can use for more advanced tasks later, like navigating around complex furniture or even planning a path through a crowded room.

In a simulation, they even tested if this "floor score" could be used to steer the robot smoothly, like a magnetic field pulling it toward the floor and pushing it away from walls. In those simulations, the robot took 4.5 times fewer steps to reach a goal compared to older, clunkier navigation methods. While they haven't put this smooth-steering system on the real robot yet (because the robot's software needs to get faster at acknowledging commands), the math works.

The Bottom Line

The paper concludes that by borrowing a tool from chemistry and aligning it with gravity, they created a floor detector that is:

Calibration-free: It works in any room without setup.
Fast: It runs in real-time on a small robot computer.
Accurate: It beats the previous best "no-training" methods.
Robust: It handles messy laser data better than older methods.

The team suggests that while the system is great for walking forward and avoiding big obstacles, it still struggles with very thin, transparent objects. They also note that the robot's current software limits how smoothly it can turn, but the "brain" (the SOAP detector) is ready for more advanced navigation tasks. This work opens the door for robots to enter new buildings and start walking immediately, without needing a human to teach them the layout first.

Technical Summary: Calibration-Free Floor Detection via Gravity-Aware SOAP Descriptors

Problem Statement
Autonomous indoor locomotion for humanoid robots requires a floor estimator that operates on every LiDAR scan in new, unseen environments without per-environment tuning (calibration). Existing geometry-based baselines, such as PCA surface normals combined with RANSAC plane fitting, achieve moderate accuracy (86.6% F1 on S3DIS Area 5) but degrade under sparse, non-uniform, and noisy sampling typical of head-mounted LiDARs. Conversely, trained deep learning methods (e.g., PointNet++, KPConv) offer higher accuracy but require labeled point clouds and suffer from domain shift when deployed in new settings. The challenge is to develop a method that is calibration-free (no labels, no height priors), robust to sparse sampling, and computationally efficient enough for real-time execution on embedded humanoid hardware.

Methodology
The authors propose a novel approach adapting the Smooth Overlap of Atomic Positions (SOAP) descriptor, originally developed for computational chemistry, to room-scale LiDAR point clouds.

Gravity-Aware $m$ -Band SOAP: Standard SOAP descriptors are fully $SO(3)$-invariant, discarding directional information. The authors introduce a gravity-aligned frame using the robot's IMU. Instead of contracting over all magnetic quantum numbers ( $m$ ) to achieve full rotational invariance, they retain the $m$ -resolved coefficients. This restricts invariance to rotations about the vertical axis ($SO(2)$), preserving directional information relative to gravity.
Floor Scoring: The method computes a per-point score based on the ratio of the power in the $\ell=2, m=0$ band to the total power in the $\ell=2$ band. A horizontal surface concentrates power in the $m=0$ channel, while vertical walls and irregular surfaces distribute power across other $m$ channels.
RANSAC Refinement: To distinguish the floor from other horizontal surfaces (e.g., table tops), the per-point SOAP scores are combined with RANSAC plane fitting. Points exceeding a fixed score threshold are candidates for RANSAC; only points within the inlier distance of the dominant plane are classified as floor.
Multi-Class Output: The same descriptor coefficients support a five-class classification (Floor, Cliff, Table, Wall, Object) without additional compute, utilizing the $\ell=1, m=0$ ratio for vertical asymmetry and signed distance to the RANSAC plane.
Embedded Deployment: The descriptor is implemented as a vectorized PyTorch module with optional CUDA-accelerated spherical harmonics. This runs on an NVIDIA Jetson Orin NX, achieving 2 Hz operation (0.08 s/scan), which is 20× faster than a CPU implementation.

Key Contributions

Gravity-Aware $m$ -Band Descriptor: A novel application of SOAP where $m$ -resolved coefficients are computed in a gravity-aligned frame. This yields an $SO(2)$-invariant feature where the $\ell=2, m=0$ ratio serves as a calibration-free floor score. The descriptor provides a 315-dimensional feature space per point, enabling multi-class surface classification and continuous similarity fields at the same compute cost as a scalar floor decision.
Calibration-Free Pipeline: A floor classification pipeline combining the SOAP $m$ -band score with RANSAC plane refinement. It requires no labeled training data, no environment-specific height thresholds, and only the physical assumption of gravity alignment.
Real-Time Humanoid Deployment: The first deployment of SOAP descriptors as a real-time perception layer on a humanoid robot (Unitree G1 EDU). The system operates as the primary obstacle-avoidance mechanism at 2 Hz, integrated with a voice-commanded reactive walk policy.

Results

S3DIS Area 5 Benchmark: On 68 rooms of the S3DIS Area 5 dataset, the SOAP + RANSAC pipeline achieved 90.0% F1 (96.2% precision, 84.5% recall), outperforming the PCA-normal + RANSAC baseline (86.6% F1). The improvement is attributed to the smooth Gaussian density representation's robustness to sparse sampling (1.8% F1 shift under subsampling vs. 7.2% for PCA normals).
Real-Robot Trials:
- Clear-Path Locomotion: 9/9 trials in a 5m corridor completed without collision.
- Obstacle Avoidance: 4/4 trials with small, sparse obstacles (wicker basket, office chair) resulted in 0 collisions.
- Ablation: In a matched SOAP-on vs. SOAP-off test, the SOAP-enabled system avoided a basket 100% of the time, while the geometric-only fallback (SOAP-off) collided in 2/2 trials.
- Latency: The GPU-vectorized implementation runs in 0.08 s per scan on the Jetson Orin NX, fitting within the 0.5 s control cycle budget.
Simulation: A SOAP-based potential field steering controller reduced simulated path length by 4.5× compared to a classical artificial potential field (APF) in scenarios where the classical APF failed due to local minima.

Significance and Claims
The paper claims this work represents the first deployment of SOAP descriptors as a real-time perception layer on a humanoid robot. It demonstrates that a descriptor derived from computational chemistry can be successfully transferred to room-scale LiDAR perception (a scale change of ~9 orders of magnitude) to solve a critical robotics problem.

The authors position the method as a "third route" in geometry-based floor detection: it improves upon the strongest calibration-free baseline (PCA+RANSAC) without requiring the labeled data of deep learning methods. The significance lies in the descriptor's ability to provide a rich, continuous geometric understanding of the scene (multi-class surface types and similarity fields) at a low computational cost, enabling reactive locomotion in previously unseen environments without pre-built maps or calibration. The authors note that while the current system is limited to reactive primitives, the continuous similarity field it generates is theoretically compatible with gradient-based steering for future goal-directed navigation.