← Back ICRA 2026

Bootstrapping Self-Supervised Learning of Binary Classification Using Error Bounds: A Case Study on a Robotic Insertion Task

Zebin Duan, Norbert Krüger, Juan Heredia, Thorbjørn Mosekjær Iversen, Frederik Hagelskjær

PDF

AI summary

Key figure (auto-extracted from paper)

A confidence-driven data engine dynamically switches between fast ML predictions and slow physical verification, enabling reliable self-supervised learning with guaranteed error bounds while minimizing setup time.

Self-supervised learning confidence bounds Wilson Score robotic insertion online learning error guarantees

Problem

Flexible manufacturing requires rapid robotic deployment, but conventional methods demand extensive manual setup, data collection, and tuning. Existing ML-based verification lacks guaranteed error bounds, making it unsafe for critical tasks without costly fallbacks.

Approach

The system uses a radius-neighbor classifier with UMAP dimensionality reduction to predict insertion success, applying Wilson-Score bounds to quantify confidence. When confidence falls below a threshold, it falls back to an expensive physical verification, collecting the labeled data to continuously improve the model online.

Key results

Guaranteed upper error bounds via Wilson-Score in an online robotic system
Dynamic reduction of expensive verification frequency as model confidence increases
Direct tuning of error tolerance and learning speed through confidence parameters
Superior reliability over Binomial Interval baselines with limited training data

Why it matters

Enables rapid, reliable deployment of self-improving robotic systems in flexible manufacturing by guaranteeing error rates without extensive initial data collection or manual tuning.

Abstract

Flexible manufacturing requires rapid deployment of solutions and minimal setup time to remain competitive. An essential attribute is the ability to control error levels, as failures can range from minor performance degradation to severe equipment damage. However, conventional deployment often involves extensive setup, data collection, model training or parameter tuning, and system testing, resulting in significant delays that hinder commercial feasibility. We propose a data engine which gathers data and improves its performance while executing the task. The data engine consists of two classifiers, a fast model prediction and expensive verification. First, a model prediction is performed and based on the confidence level of the prediction, the expensive verification can be used. By adjusting the confidence level, users can control the level of tolerable error. Our method is implemented on a real-world robotic insertion task, which uses force data for the model prediction. The system applies UMAP dimensionality reduction and uses Wilson-Score to compute the confidence bounds of the prediction. Results demonstrate the ability to learn and reduce the need for expensive verifications over time, while staying within the set error-rate. The results highlight the potential of confidence bounds in self-improving models to enhance reliability in robotic classification task.

Index terms

Continual Learning Industrial Robots Force and Tactile Sensing