Controller Synthesis for Safety of Physically-Viable Data-Driven Models

Mohamadreza Ahmadi, Arie Israel, Ufuk TopcuarXiv 2018 · DOI · arXiv

Abstract

We consider the problem of designing finite-horizon safe controllers for a dynamical system for which no explicit analytical model exists and limited data only along a single trajectory of the system are available. Given samples of the states and inputs of the system, and additional side information in terms of regularity of the evolution of the states and conservation laws, we synthesize a controller such that the evolution of the states avoid some pre-specified unsafe set over a given finite horizon. Motivated by recent results on Whitney's extension theorem, we use piecewise-polynomial approximations of the trajectories based on the data along with the regularity side information to formulate a data-driven differential inclusion model that can predict the evolution of the trajectories. For these classes of data-driven differential inclusions, we propose a safety analysis theorem based on barrier certificates. As a corollary of this theorem, we demonstrate that we can design controllers ensuring safety of the solutions to the data-driven differential inclusion over a finite horizon. From a computational standpoint, our results are cast into a set of sum-of-squares programs whenever the certificates are parametrized by polynomials of fixed degree and the sets are semi-algebraic. This computational method allows incorporating side information in terms of conservation laws and integral constraints using Positivstellensatz-type arguments.

Introduction

In safety-critical applications—such as medical robotics or unmanned aerial vehicles—formal guarantees that a system will avoid unsafe states are paramount. Traditionally, this requires a precise analytical model in the form of differential equations. However, modern systems are often too complex to model perfectly, and "data-hungry" machine learning methods often require thousands of trials to learn a safe policy. This paper addresses the challenge of synthesizing provably safe controllers when only limited data from a single trajectory is available and no explicit model exists.

Methodology

The proposed framework bridges the gap between sparse data and formal safety through a multi-step pipeline:

  1. Piecewise-Polynomial Approximation: B-splines are used to approximate the system's trajectory from discrete data points.
  2. Differential Inclusions: Recognizing that any data-driven model has errors, the authors use Whitney’s Extension Theorem to define a set of possible system behaviors (a differential inclusion). If the real system is sufficiently smooth (regular), it is guaranteed to live within this set.
  3. Incorporating Side Information: Physical truths, such as conservation of energy or mechanical limits, are incorporated using Positivstellensatz-type arguments. This mathematically restricts the set of possible system behaviors to those that are physically viable.
Differential inclusion model showing uncertainty regions (cones) derived from sparse trajectory data for state and velocity.
  1. Barrier Certificates and SOS Programming: To prove safety, the framework searches for a "barrier certificate"—a function that acts as a boundary the system cannot cross into an unsafe zone. This search is formulated as a Sum-of-Squares (SOS) program, turning a difficult nonlinear problem into a convex optimization problem that can be solved efficiently.

Key Results

The method was evaluated using simulated dynamical systems to test performance under extreme data scarcity:

  • Nonlinear 1D System: Using only 20 samples from a single trajectory, the framework predicted safety boundaries and synthesized a controller. As the degree of the polynomial used for the barrier certificate increased, the predicted "safe time" $T$ approached the true theoretical limit (reaching $T=2.49$ vs. the actual $T=2.66$).
Safety analysis of a 1D system: the predicted trajectory (blue) is kept below the unsafe threshold (red) based on sparse data samples (circles).
  • Aircraft Landing Model: In a 3D longitudinal model of a DC9-30 aircraft, the authors simulated failure scenarios like sudden lift loss. Using just 25 non-uniform data samples, the synthesized controller successfully intervened to guide the aircraft back to a safe trajectory, preventing it from exceeding safe landing speeds.

Discussion and Future Work

By embracing uncertainty and using physics as a filter, the authors demonstrate that formal safety guarantees are possible even with minimal data. Instead of finding a single "perfect" model, modeling the uncertainty itself as a set allows for robust safety proofs.

Future work will address noisy data by incorporating spline smoothing to handle probabilistic uncertainty. Additionally, while SOS programming is efficient for systems with a moderate number of states, scaling the approach to high-dimensional systems remains a primary computational challenge. Extending this framework to hybrid dynamics (e.g., systems with impacts) is also a promising next step, as these systems provide clear physical "side information" that can further constrain the model.

Summary updated

Abstract

We consider the problem of designing finite-horizon safe controllers for a dynamical system for which no explicit analytical model exists and limited data only along a single trajectory of the system are available. Given samples of the states and inputs of the system, and additional side information in terms of regularity of the evolution of the states and conservation laws, we synthesize a controller such that the evolution of the states avoid some pre-specified unsafe set over a given finite horizon. Motivated by recent results on Whitney's extension theorem, we use piecewise-polynomial approximations of the trajectories based on the data along with the regularity side information to formulate a data-driven differential inclusion model that can predict the evolution of the trajectories. For these classes of data- driven differential inclusions, we propose a safety analysis theorem based on barrier certificates. As a corollary of this theorem, we demonstrate that we can design controllers ensuring safety of the solutions to the data-driven differential inclusion over a finite horizon. From a computational standpoint, our results are cast into a set of sum-of-squares programs whenever the certificates are parametrized by polynomials of fixed degree and the sets are semi-algebraic. This computational method allows incorporating side information in terms of conservation laws and integral constraints using Positivstellensatz-type arguments.
BibTeX
@misc{ahmadi2018controller,
  author = {Mohamadreza Ahmadi and Arie Israel and Ufuk Topcu},
  title = {{Controller Synthesis for Safety of Physically-Viable Data-Driven Models}},
  year = {2018},
  eprint = {1801.04072},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/1801.04072},
}
Figure