Safety Filter

A layer between a policy and the actuators that only minimally overrides an action when it would otherwise be unsafe.

A safety filter splits ‘completing the task’ from ‘staying safe’ into two separate layers. A task policy — a hand-written controller, a reinforcement-learning policy, or a VLA model — proposes a nominal action, and the filter checks whether executing it would keep the system inside a defined safe set. If it's safe, the action passes through unchanged; if not, it's replaced with the closest safe alternative. Three implementations are common: control barrier functions (CBFs) solve a small quadratic program every step to find the action closest to the nominal one that keeps a safety function h(x) from dropping below zero; Hamilton-Jacobi (HJ) reachability analysis precomputes which states will lead to a hazard no matter what control is applied afterward; and MPC-based shielding predicts online whether a future trajectory can still return to the safe region. A 2023 survey by Kai-Chieh Hsu, Haimin Hu, and Jaime Fisac unifies these under a ‘monitor and intervene’ framework. Learned policies carry no inherent safety guarantee on their own, so adding this outer layer is what provides a provable constraint.

ExampleFor a mobile robot avoiding people, define h(x) = ‖p − p_human‖² − d² (p is the robot's position, d is the minimum allowed distance, and h ≥ 0 means safe). Each cycle it solves min‖u − u_nom‖² subject to ḣ ≥ −α·h, where u_nom is the velocity from the navigation policy and α > 0 sets how early it starts slowing as it approaches. Far from people the constraint is inactive and u simply equals u_nom.

Also called

Safety Shield

Related

Control Barrier Function · Hamilton-Jacobi Reachability Analysis · Model Predictive Control · Safe Reinforcement Learning · Embodied Safety · Quadratic Programming

Sources

The Safety Filter: A Unified View of Safety-Critical Control in Autonomous Systems (Hsu, Hu, Fisac, arXiv 2309.05837)
Control Barrier Functions: Theory and Applications (Ames et al., arXiv 1903.11199)