Safety Filtering with Control Barrier Functions

You are reading version 1 of this direction, frozen 2026-10-04. The latest version is version 2.

Version 1 of 2253 papers301 conceptsfrozen 2026-10-04

In robotics, there is a fundamental tension between performance and safety. We want robots to move fast, navigate complex environments, and learn from experience, but we cannot afford for them to crash into obstacles or damage themselves. Traditionally, safety was handled by being conservative—slowing down or keeping a wide berth.

A growing body of research has converged on a more elegant solution: the Safety Filter. Instead of building safety directly into a complex navigation algorithm, researchers are using Control Barrier Functions (CBFs) to create a mathematical "shield" that sits between a robot's high-level brain and its low-level motors. This shield only intervenes when a collision is imminent, allowing the robot to be as "reckless" as it wants until the very last safe moment.

The Intuition of a Barrier

Imagine a robot moving toward a wall. A standard controller might calculate the distance and slow down gradually. A Control Barrier Function approach, however, defines a "safe set"—a mathematical region where the robot is allowed to exist.

As explained by Andrew Singletary and Aaron Ames in Safety-Critical Kinematic Control of Robotic Systems, the CBF acts as a certificate. If the robot’s state is inside the safe set, the CBF ensures there exists a control action that keeps it there. It doesn't just look at where the robot is, but where it is going and how fast. By incorporating kinetic energy and system dynamics, the barrier ensures the robot never reaches a point where a collision becomes physically unavoidable.

Minimal Intervention via Quadratic Programming

The core mechanism that makes this practical is Quadratic Programming (QP). In every control cycle (often hundreds of times per second), the robot solves a small optimization problem:

  1. Input: A "nominal" command (e.g., from a human pilot, a neural network, or a path planner).
  2. Constraint: A linear inequality derived from the Control Barrier Function that defines the boundary of safety.
  3. Objective: Find the control signal that is mathematically closest to the nominal command while strictly satisfying the safety constraint.

This "minimal intervention" philosophy is what allowed Eric Feron and his team to build the Robotarium. The facility allows remote researchers to run potentially "reckless" swarm code on expensive hardware because the underlying safety layer—the Safety Barrier Certificate—guarantees no two robots will ever collide, regardless of what the user's code commands.

Beyond Simple Circles

Early safety filters often treated obstacles as simple circles or spheres, which led to "timid" behavior in tight spaces. Recent work has focused on making these filters "configuration-aware."

For instance, Fan Ding and colleagues in Configuration-Aware Safe Control for Mobile Robotic Arm use convex sets to ensure that if a robot's joints are safe, the entire connecting link is also safe. Similarly, Manan Tayal introduced Collision Cone Control Barrier Functions, which use the geometry of relative velocity to handle moving obstacles more naturally than traditional distance-based methods. By modeling obstacles as polygons rather than circles, Polygonal Cone CBFs allow robots to navigate narrow corridors and brush past walls without triggering a "panic" response from the safety filter.

Bridging the Gap to AI

One of the most exciting frontiers for CBFs is their integration with generative AI and Large Language Models (LLMs). While neural networks are excellent at planning complex behaviors, they are notoriously difficult to verify.

Researchers like Wei Xiao have developed frameworks like SafeDiffuser, which embeds CBF constraints directly into the denoising process of a diffusion model. This ensures that the trajectories generated by the AI are physically safe from the moment they are conceived.

Furthermore, Lukas Brunke and his team have shown that LLMs can be used to translate "common sense" risks—like "don't move the water over the laptop"—into formal CBF constraints in Semantically Safe Robot Manipulation. This allows the robot to benefit from high-level reasoning while maintaining the low-level mathematical guarantees of the safety filter.

The Challenge of Deadlock

While safety filters are highly effective at preventing crashes, they can sometimes be too effective, leading to "deadlock." This happens when the safety constraints and the goal-seeking behavior of the robot cancel each other out, leaving the robot stuck in a local minimum.

Changliu Liu and Jaskaran Grover characterized this in The Before, During, and After of Multi-Robot Deadlock, describing it as a physical force equilibrium. Their work suggests that by slightly rotating the robot's intent when it nears a barrier, we can "nudge" it out of these traps without sacrificing the underlying safety guarantee.

By decoupling "what to do" (the nominal controller) from "how to stay safe" (the CBF filter), this research direction provides a modular, mathematically rigorous path toward deploying autonomous systems in the messy, unpredictable real world.

Papers in this version

Showing 60 of 253 papers.

History & citationVersion 2 · 2026-10-04

Current · Permanent link

  1. Version 2 (latest) · 2026-10-04 · 232 papers
  2. Version 1 · 2026-10-04 · 253 papers
Cite this direction

Latest version (a living page):

@misc{lacuna_optimization_based_safety_filtering_e25cbd59,
  title = {optimization-based safety filtering},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/optimization-based-safety-filtering/rdr_e25cbd59d82946c6bd354eccc93fcc18},
  note = {Latest version of a living research direction; cite a version for a fixed text. Accessed 2026-11-20},
}

This exact version (version 1, frozen text):

@misc{lacuna_safety_filtering_with_control_barrier_fu_e25cbd59_v1,
  title = {Safety Filtering with Control Barrier Functions},
  author = {{Lacuna}},
  howpublished = {Lacuna research direction},
  year = {2026},
  url = {https://lacuna.tiptreesystems.com/direction/safety-filtering-with-control-barrier-functions/rdr_e25cbd59d82946c6bd354eccc93fcc18/v/1},
  note = {Version 1, frozen 2026-10-04. Accessed 2026-11-20},
}