A Feedback-Control Framework for Efficient Dataset Collection from In-Vehicle Data Streams

TL;DR

This work introduces Feedback Control Data Collection (FCDC), a closed-loop framework that treats data collection from in-vehicle streams as a dynamical system guided by online density estimation to regulate sample retention. By embedding samples, online estimating the dataset distribution, and using a value function to modulate sampling, FCDC actively balances exploration and redundancy, transforming data collection into a self-regulating component of the learning loop. Two value-function designs, complementary Gaussian and reciprocal Gaussian, are analyzed, with empirical results showing improved data balance (up to 25.9%) and reduced storage (up to 39.8%) on real vehicle data, and synthetic experiments illustrating convergence toward a target distribution. The approach offers a principled way to achieve diverse, representative datasets in data-centric AI for autonomous driving, while pointing to future work on richer, multimodal distributions and end-to-end model-performance evaluations.

Abstract

Modern AI systems are increasingly constrained not by model capacity but by the quality and diversity of their data. Despite growing emphasis on data-centric AI, most datasets are still gathered in an open-loop manner which accumulates redundant samples without feedback from the current coverage. This results in inefficient storage, costly labeling, and limited generalization. To address this, this paper introduces Feedback Control Data Collection (FCDC), a paradigm that formulates data collection as a closed-loop control problem. FCDC continuously approximates the state of the collected data distribution using an online probabilistic model and adaptively regulates sample retention using based on feedback signals such as likelihood and Mahalanobis distance. Through this feedback mechanism, the system dynamically balances exploration and exploitation, maintains dataset diversity, and prevents redundancy from accumulating over time. In addition to demonstrating the controllability of FCDC on a synthetic dataset that converges toward a uniform distribution under Gaussian input assumption, experiments on real data streams show that FCDC produces more balanced datasets by 25.9% while reducing data storage by 39.8%. These results demonstrate that data collection itself can be actively controlled, transforming collection from a passive pipeline stage into a self-regulating, feedback-driven process at the core of data-centric AI.

Figures (7)

• Figure 1: Data Flow Schematic of the Feedback Control Data Collection framework. The input Stream $\mathcal{S}$ consisting of data points $x[k]$ at the corresponding step $k$ as the exogenous disturbance. The Embedding function $\phi$ transforms the data into an embedding space in which the distribution is estimated using the distribution estimation $\mathcal{E}$. Based on the estimation and the target distribution $\theta^\star$, the Value Function $\mathcal{V}$ updates the data collection strategy in the Collection Control $\mathcal{F}$.
• Figure 2: Evolution of the collection probability over the number of processed samples for newly arriving data points, based on $\psi_\mathrm{C}$ (top) and $\psi_U$ (bottom).
• Figure 3: Comparison of selected data points between the ground truth (), $\psi_\mathrm{C}$ () and $\psi_\mathrm{R}$ ().
• Figure 4: Q–Q diagrams comparing $10^5$ samples: random collection (), FCDC with $\psi_\mathrm{C}$ (), and FCDC with $\psi_\mathrm{R}$ (), each compared against a theoretical uniform distribution in both dimensions. Deviation of the diagonal means a deviation to the theoretical uniform distribution.
• Figure 5: Error $\Delta_\mathrm{uni}$ of collected samples $N$ to a theoretical uniform over the number of collected samples using random collection (), FCDC with $\psi_\mathrm{C}$ (), and FCDC with $\psi_\mathrm{R}$ ().

Theorems & Definitions (1)

• proof