Iterative Safe Set Expansion

53 concepts47 papers1 proposal

When controlling a complex system—whether it is a quadrotor navigating turbulent winds or an industrial robot arm avoiding obstacles—safety is paramount. Traditional robust control handles uncertainty by preparing for the absolute worst-case scenario. However, this creates a fundamental dilemma: if the initial model of the system's dynamics or environment is highly uncertain, the guaranteed "safe operating region" shrinks to a tiny, overly conservative zone. Staying within this zone prevents crashes, but it also prevents the system from ever exploring or optimizing its performance.

The core insight of Iterative Safe Set Expansion is that safety does not have to be static. Instead of restricting a system to a permanently small operational envelope, we can treat the safe set as a dynamic boundary that grows as the system gathers data. By systematically exploring the edges of what is currently known to be safe, the system can reduce its model uncertainty, update its mathematical safety certificates, and safely expand its operational envelope over time.

The Anatomy of the Expansion Loop

How can a system expand its safe zone without crossing into danger? The process operates as a tightly coupled loop consisting of three main phases: boundary exploration, uncertainty reduction, and certificate updating.

  1. Boundary Exploration: Rather than optimizing for performance immediately, the system deliberately guides its states toward the perimeter of the currently verified safe set. As shown by Li Wang et al. in Safe Learning of Quadrotor Dynamics Using Barrier Certificates, prioritizing data collection near these boundaries provides the most informative data about the transition between safe and unsafe regimes.
  2. Uncertainty Reduction: When the system operates near the boundary, it collects new state transition data. This data is used to refine a statistical model of the system's dynamics, often using Gaussian Processes or robust regression frameworks. As demonstrated by Anqi Liu et al. in Robust Regression for Safe Exploration in Control, directly predicting uncertainty bounds via minimax optimization allows the system to tighten its worst-case estimates. As the model's confidence improves, regions that were previously deemed too risky due to high uncertainty are revealed to be safe.
  3. Certificate Updating: Once the uncertainty shrinks, the mathematical boundaries of the safe set can be formally pushed outward. This "expand-then-maximize" strategy, pioneered in algorithms like SAFEOPT by Yanan Sui et al. in Safe Exploration for Optimization with Gaussian Processes, ensures that the searchable parameter space is sufficiently understood before any high-performance maneuvers are attempted.

Bootstrapping from a Local Seed

An expansion loop requires a starting point—a "seed" region that is guaranteed to be safe from the very beginning. How do we find this initial set without global knowledge of the dynamics?

A common approach is to leverage local linear control theory around a known equilibrium point. For instance, Priyabrata Saha et al. in Neural Identification for Control demonstrate that neural controllers can bootstrap from a small neighborhood around an equilibrium point, using an iterative training process to learn safe recovery behaviors from progressively more distant states. Similarly, John M Dolan et al. in Synthesis and verification of robust-adaptive safe controllers initiate their synthesis process with a small safe region derived from a local linear controller, systematically expanding it through successive optimization steps to capture a much larger invariant set than traditional robust approaches could manage.

Alternatively, the seed can be constructed from historical data or a single successful demonstration. Lukas Brunke et al. in RLO-MPC: Robust Learning-Based Output Feedback MPC for Improving the Performance of Uncertain Systems in Iterative Tasks show that a safe operating region can be built directly from the convex hull of previously visited states during successful executions, allowing the controller to reliably expand its reachable workspace as it gains experience.

Mechanisms of Growth: Flow, Backups, and Inflation

Depending on how the safe set is mathematically represented, researchers have developed distinct mechanisms to compute and drive this regional growth.

Forward Simulation and Flow Invariance

If the safe set is defined via a Lyapunov function, the region can be expanded by identifying states whose trajectories naturally "flow" into the already certified zone. Felix Berkenkamp et al. in The Lyapunov Neural Network: Adaptive Stability Certification for Safe Learning of Dynamical Systems introduced the Lyapunov Neural Network (LNN), which uses forward simulations to discover these states. If a state is guaranteed to terminate inside the known safe zone, its safety certificate can be updated, allowing the boundary to dynamically reshape itself to the true geometry of the system's dynamics.

Feasible Backup Trajectories

In Model Predictive Control (MPC), a state is considered safe if there exists a valid sequence of future control actions (a backup trajectory) that can bring the system to a stop or a known invariant set. Kim Peter Wabersich et al. in Linear model predictive safety certification for learning-based control leverage this by having the system incrementally enlarge its safe operating region as it discovers and memorizes new feasible backup trajectories during exploration.

This idea can also be extended to backward propagation. Brijen Thananjeyan et al. in ABC-LMPC: Safe Sample-Based Learning MPC for Stochastic Nonlinear Dynamical Systems with Adjustable Boundary Conditions expand a controller's range by actively seeking paths from unknown start states into previously verified safe zones. In a similar vein, Albert Wilcox in LS3: Latent Space Safe Sets for Long-Horizon Visuomotor Control of Sparse Reward Iterative Tasks proposes a recursive training objective that labels a state as safe if it is predicted to lead to an existing safe state, propagating success confidence backward to guide the agent.

Alternating Optimization and Inflation

When geometry dominates—such as finding collision-free configurations for a robotic arm—the safe set can be modeled as a convex polytope. Hongkai Dai et al. in Certified Polyhedral Decompositions of Collision-Free Configuration Space utilize an alternating optimization strategy called C-IRIS that switches between a certification step (finding hyperplanes that separate the robot from obstacles) and an inflation step (expanding the polytope boundaries). This regional inflation can expand the certified safe volume by orders of magnitude within a small number of iterations.

Why It Works: Decoupling Exploration from Optimization

The intuitive reason why iterative safe set expansion works so effectively is that it decouples two conflicting objectives: exploring to find the boundaries of safety and optimizing to maximize performance.

If an agent tries to maximize an objective function within an unexpanded, conservative safe set, it often gets trapped in local optima because the true global optimum lies outside its current comfort zone. As highlighted by Marcello Fiducioso et al. in Safe Contextual Bayesian Optimization for Sustainable Room Temperature PID Control Tuning, dynamic safe set expansion allows optimization algorithms to discover high-performance parameters that were initially invisible under conservative safety boundaries.

By explicitly dedicating phases of operation to expanding the safe envelope based on uncertainty reduction (as seen in the symbolic abstraction work of Adnane Saoud et al. in Learning-based symbolic abstractions for nonlinear control systems), the system safely increases its exploration capacity over time, paving a secure path toward global optimality.

Key Foundational Work

Extending Consistent Domains of Numeric CSP (1999) — contributed 1 concept elements to this direction.
Generalized Interval Projection: A New Technique for Consistent Domain Extension (2007) — contributed 1 concept elements to this direction.
History & citationVersion 1 · 2026-10-04

Current · Permanent link

View frozen version 1

Cite this direction

Latest version (a living page):

@misc{lacuna_iterative_safe_set_expansion_7c3d6c6d,
  title = {iterative safe set expansion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/iterative-safe-set-expansion/rdr_7c3d6c6d5d33417aa33a0d0801284de7},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-21},
}

This exact version (version 1, frozen text):

@misc{lacuna_iterative_safe_set_expansion_7c3d6c6d_v1,
  title = {iterative safe set expansion},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/iterative-safe-set-expansion/rdr_7c3d6c6d5d33417aa33a0d0801284de7/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-21},
}
About

These directions overlap with this one through shared papers.

See something wrong? Report a mistake