Preface

Ali Jadbabaie, John Lygeros, George J. Pappas, Pablo A. Parrilo, Benjamin Recht, Claire J. Tomlin, Melanie N. ZeilingerL4DC 2021

Introduction

The L4DC 2021 preface outlines the mission of a rapidly growing conference series dedicated to bridging the gap between traditional model-based control theory and data-driven machine learning. By creating a dedicated forum for "Learning for Dynamics and Control," the organizers aim to address the unique challenges of applying artificial intelligence to physical, real-time systems where safety and dynamical foundations are paramount.

The Interdisciplinary Gap

The authors identify a significant "silo" problem in modern engineering and computer science. On one hand, control theory has spent decades perfecting model-based design for physical systems but struggles to incorporate the massive influx of real-time, high-dimensional data from embedded devices. On the other hand, machine learning has revolutionized fields like computer vision and natural language processing but often lacks the tools to handle complex dynamical processes or the rigorous safety requirements of decision-making in the physical world.

Existing venues often failed to bridge this divide: major control conferences (such as CDC) frequently felt invisible to ML researchers, while top-tier ML conferences (such as NeurIPS or ICML) showed limited interest in the nuances of dynamical systems. L4DC was established to fill this niche, providing a "natural home" for researchers working at the intersection of control, optimization, and learning.

Technical Pillars and Methodology

As a preface to the conference proceedings, the document describes the organizational framework used to build this interdisciplinary community. The conference scope highlights several key technical pillars essential for the future of autonomous systems:

  • System Identification: Merging classical modeling with modern data techniques.
  • Safe Reinforcement Learning: Ensuring physical systems do not fail while learning in real-world environments.
  • Physics-Constrained Learning: Incorporating known physical laws into ML models to improve sample efficiency and reliability.
  • Distributed and Statistical Learning: Handling data across networked systems and quantifying uncertainty in control loops.

Due to the global pandemic, the 2021 event transitioned to a virtual format. The organizers utilized platforms like Zoom, YouTube, and Gather.Town to maintain the interactive nature of poster sessions and the critical "rebuttal phases" during peer review.

Key Insights: Beyond Pattern Recognition

A core realization for the L4DC community is that machine learning for dynamics must move beyond simple pattern recognition. It must evolve into data-driven decision making, where the temporal and spatial constraints of physics are treated as first-class citizens. The authors argue that the most significant generator of data in the coming decade will not be social media or text, but embedded devices sensing and controlling the physical world.

The success of the conference suggests that the previous marginalization of these researchers in other venues was a significant barrier to progress. The authors use the term "rapprochement" to describe the necessary coming-together of control, ML, and optimization—suggesting that these fields are fundamentally dependent on one another for the next generation of robotics and autonomy.

Growth and Results

The preface provides a quantitative snapshot of the field's rapid expansion as of 2021:

  • High Selectivity: The conference received 140 submissions, of which only 14 (10%) were accepted for oral presentation, with an additional 90 accepted as posters.
  • Growing Engagement: Despite the virtual format, the event secured nearly 600 registrations, a 50% increase over the inaugural 2019 event's 400 attendees.
  • Global Reach: The virtual format allowed for a peak of 500 concurrent viewers on YouTube, demonstrating a global interest that exceeds physical venue capacities.

Challenges and Future Directions

While the virtual format was successful, the authors acknowledge that the social aspects of community building are harder to replicate online. A primary goal moving forward is to transition L4DC from a series of installments into a permanent, annual fixture in the academic calendar.

Ultimately, the authors note that truly bridging these "distant areas" of research remains an ongoing challenge. It requires nurturing a new generation of junior researchers who are "bilingual"—equally proficient in the rigors of control theory and the scalability of machine learning.

Summary updated

BibTeX
@inproceedings{jadbabaie2021preface,
  author = {Ali Jadbabaie and John Lygeros and George J. Pappas and Pablo A. Parrilo and Benjamin Recht and Claire J. Tomlin and Melanie N. Zeilinger},
  title = {{Preface}},
  booktitle = {L4DC},
  year = {2021},
}
Figure