KISS-IMU: Self-supervised Inertial Odometry with Motion-balanced Learning and Uncertainty-aware Inference

KISS-IMU is a novel self-supervised inertial odometry framework that eliminates the need for ground truth data by leveraging LiDAR-based ICP registration and pose graph optimization as supervisory signals, while employing motion-balanced training and uncertainty-aware inference to ensure robustness across diverse robotic platforms and environments.

The Gist

Imagine you are trying to navigate a blindfolded person through a dense, foggy forest. You can't see the trees (no GPS), and you can't see the ground (no cameras). All you have is a pedometer strapped to their ankle (the IMU) that counts steps and a gyroscope that tells you which way they are turning.

This is the job of Inertial Odometry (IO): figuring out where a robot is just by feeling its own movements.

The Problem: The "Drunk Navigator"

The problem with relying only on the pedometer is that it's a bit unreliable. If the robot runs, the pedometer might overcount. If the robot slips, it might undercount. Over time, these tiny errors add up, and the robot thinks it's in a different country than it actually is. This is called "drift."

To fix this in the past, scientists taught robots using Ground Truth. Imagine a teacher standing on a hill with a perfect map, shouting, "You are here! No, you're there!" The robot learns by comparing its guess to the teacher's answer.

• The Catch: You can't have a teacher with a perfect map in every forest, cave, or on every planet. It's too expensive and hard to set up. If you only train the robot in one specific forest, it gets confused when you take it to a different forest.

The Solution: KISS-IMU

The authors of this paper created KISS-IMU. The name stands for Keep IMU Stable and Strong. Think of it as a new training method that doesn't need a teacher with a map. Instead, it uses a clever "self-check" system.

Here is how it works, broken down into three simple concepts:

1. The "Self-Check" (Self-Supervised Learning)

Instead of a teacher shouting the answer, KISS-IMU uses a Laser Scanner (LiDAR) as a temporary guide.

• How it works: The robot scans the trees around it. It takes a picture of the forest, moves a bit, and takes another picture. It then tries to match the two pictures together (like a puzzle).
• The Magic: If the robot's internal pedometer says it moved 5 meters, but the laser scanner says the trees only shifted 3 meters, the robot knows, "Oops, I was wrong." It uses this mismatch to correct itself.
• Why it's special: It doesn't need a human to tell it the right answer. It just needs to make sure its movement guess matches the shape of the world it sees.

2. "Keeping it Stable" (Motion-Balanced Learning)

Imagine you are learning to ride a bike. If you only practice on a perfectly straight, flat road, you will get really good at going straight. But the moment you hit a sharp turn or a bumpy hill, you'll crash.

• The Issue: Most training data is "boring." Robots spend 90% of their time walking straight. They rarely practice sharp turns or sudden stops. So, the AI gets lazy and only learns the boring stuff.
• The KISS-IMU Fix: The authors invented a "Gym Coach" (called a Gaussian Mixture Model). This coach looks at the robot's training data and says, "Hey, you've done 1,000 straight walks, but only 5 sharp turns! Let's focus on the sharp turns."
• The Result: The robot learns to handle every type of movement, not just the easy ones. It becomes "Stable" because it's not biased toward one type of motion.

3. "Keeping it Strong" (Uncertainty-Aware Inference)

Now imagine the robot is in a storm. The wind is blowing, the ground is slippery, and the laser scanner is getting confused by rain.

• The Issue: A standard robot would blindly trust its sensors even when they are lying.
• The KISS-IMU Fix: The robot learns to ask, "How sure am I?" It calculates its own confidence score.
  • If the ground is smooth and the laser sees clearly, the robot says, "I'm 100% sure, let's go!"
  • If it's slippery and the laser is blurry, the robot says, "I'm not sure, I'll trust my memory a bit less and be more careful."
• The Result: The robot becomes "Strong" because it knows when to trust its sensors and when to be cautious. It adapts to the weather and the terrain in real-time.

The Big Win

The paper tested this on regular robots and even quadruped robots (like robot dogs) running through craters and rough terrain.

• Old methods: Needed a perfect map to train. If you took them to a new place, they failed.
• KISS-IMU: Trained itself using only the robot's own movement and a laser scanner. It worked perfectly in new, unseen environments, even with very little training data.

Summary Analogy

Think of KISS-IMU as a student who:

1. Doesn't need a teacher: They learn by checking their own homework against the textbook (Self-supervised).
2. Studies the hard stuff: They don't just memorize the easy questions; they force themselves to practice the rare, difficult problems (Motion-balanced).
3. Knows their limits: They know when they are tired or confused and adjust their effort accordingly (Uncertainty-aware).

This makes the robot a much better explorer, capable of navigating the unknown without needing a human to hold its hand.

Technical Summary

Here is a detailed technical summary of the paper "KISS-IMU: Self-supervised Inertial Odometry with Motion-balanced Learning and Uncertainty-aware Inference."

1. Problem Statement

Inertial Odometry (IO) using Deep Neural Networks (DNNs) has shown promise but suffers from a critical bottleneck: heavy reliance on ground truth data (e.g., motion capture systems) for training.

• Scalability & Generalization: Acquiring centimeter-accurate ground truth in diverse, real-world environments (especially for quadruped robots or aggressive maneuvers) is resource-intensive and often impossible.
• Overfitting: Supervised methods tend to overfit to specific motion patterns and environments seen during training, leading to poor performance in unseen scenarios (e.g., different terrains or motion dynamics).
• Limitations of Existing Self-Supervised Methods: Previous self-supervised approaches often require joint training of multiple modalities (e.g., Visual-Inertial), creating tightly coupled architectures that inherit the same generalization limitations as supervised methods.

2. Methodology: KISS-IMU Framework

The authors propose KISS-IMU (Keep IMU Stable and Strong), a self-supervised framework that learns only the IMU network without ground truth. It leverages LiDAR solely as a lightweight supervisory signal via geometric constraints.

A. Core Architecture

• IMU Network: A CNN-GRU encoder-decoder architecture processes raw IMU measurements (angular velocity and linear acceleration).
• Dual Output: The network predicts:
  1. Corrections ($\hat{\sigma}$): Bias and scale corrections for IMU measurements.
  2. Uncertainties ($\hat{\eta}$): Learned noise covariances for angular velocity and acceleration.
• Integration: Corrected measurements are pre-integrated to estimate relative pose changes ($\Delta R, \Delta v, \Delta p$). The learned uncertainties are propagated through the integration process to update the state covariance.

B. Self-Supervised Training (The "Stable" Component)

To eliminate ground truth dependency, the system generates pseudo-labels using LiDAR data:

1. Pseudo-Label Generation:
   • ICP (Iterated Closest Point): Provides relative pose between consecutive LiDAR scans.
   • PGO (Pose Graph Optimization): Optimizes global consistency using both LiDAR and IMU constraints.
   • Selective Fusion: A Symmetric Overlap Score is calculated for both ICP and PGO results. The transformation with the higher geometric consistency (overlap) is selected as the pseudo-label for that segment.
2. Motion-Balanced Learning (GMM Reweighting):
   • Problem: Sequential datasets are often imbalanced (e.g., dominated by straight-line motion), causing the network to ignore rare but critical maneuvers (sharp turns, jumps).
   • Solution: The authors use a Gaussian Mixture Model (GMM) to cluster motion patterns based on IMU descriptors.
   • Reweighting: A class-balanced weighting strategy is applied. Rare motion clusters receive higher weights during loss calculation, ensuring the network learns diverse dynamics rather than just dominant patterns.
3. Loss Function: Combines standard pose errors (rotation, velocity, position) with uncertainty-aware losses (incorporating the learned covariance) and the GMM-based motion weights.

C. Uncertainty-Aware Inference (The "Strong" Component)

During inference, the system employs Sensor Confidence-Aware Adaptive Weights within the PGO framework:

• LiDAR Confidence: Weighted by the Symmetric Overlap Score (higher overlap = higher confidence).
• IMU Confidence: Weighted by the inverse of the propagated covariance (lower uncertainty = higher confidence).
• Result: The PGO dynamically adjusts constraints based on real-time sensor reliability, ensuring robustness in varying conditions (e.g., LiDAR degradation in featureless environments or IMU saturation during aggressive motion).

3. Key Contributions

1. Novel Self-Supervised Paradigm: The first framework to train an IMU-only odometry network without ground truth, using LiDAR registration and PGO purely for supervision without joint learning of visual features.
2. Motion-Aware Balanced Training: Introduction of GMM-based motion clustering and reweighting to prevent bias toward dominant motion patterns, significantly improving generalization to unseen environments.
3. Uncertainty-Driven Adaptive Inference: A mechanism to dynamically weight sensor constraints during PGO based on learned IMU uncertainty and geometric LiDAR consistency, enhancing robustness.
4. Comprehensive Evaluation: Validated across diverse platforms (wheeled robots, quadruped robots) and datasets (Botanic Garden, DiTer++, In-House), including extreme conditions where ground truth is unobtainable.

4. Experimental Results

• Datasets: Evaluated on Botanic Garden (wheeled), DiTer++ (quadruped, diverse terrains), and an In-House dataset (crater-filled planetary analog, extreme motion).
• Performance:
  • Generalization: KISS-IMU significantly outperformed state-of-the-art methods (TLIO, AirIMU, AirIO) on unseen sequences, particularly when training data was limited (e.g., 20% of the dataset).
  • Robustness: In the "In-House" dataset (extreme complexity, no ground truth available), supervised methods were infeasible, while KISS-IMU successfully generated consistent maps.
  • Ablation Studies:
    • Removing GMM balancing led to performance drops in unseen environments (overfitting to dominant motions).
    • Removing adaptive weighting reduced robustness in varying sensor conditions.
  • Metrics: Achieved competitive Absolute Pose Error (APE) and Relative Pose Error (RPE), often ranking 1st or 2nd across all test scenarios.

5. Significance

• Scalability: By removing the need for motion capture systems, KISS-IMU enables the deployment of high-precision inertial odometry in real-world, unstructured environments (e.g., disaster zones, planetary exploration) where ground truth is impossible to obtain.
• Decoupled Learning: The approach proves that IMU networks can be learned independently of visual modalities, avoiding the complexity and generalization limits of multi-modal joint training.
• Practical Applicability: The framework is particularly valuable for quadruped robots and dynamic systems where motion patterns are highly irregular and difficult to model with traditional filters or supervised learning.

In summary, KISS-IMU represents a significant step forward in making inertial odometry scalable, generalizable, and robust for real-world robotic applications by leveraging self-supervised geometric constraints and intelligent motion balancing.