Source-linked AI summary
The Safety Filter: A Unified View of Safety-Critical Control in Autonomous Systems
Kai-Chieh Hsu, Haimin Hu, Jaime Fernández Fisac
TL;DR
Autonomous robots need reliable safety guarantees across increasingly diverse deployment settings, while existing model-based and data-driven approaches have complementary limitations. The paper reviews safety filters and unifies them through a common technical and theoretical framework. It identifies shared modular structure across filter classes and uses that view to frame more scalable synthesis, robust monitoring, and efficient intervention as future directions.
Problem
Safe operation remains difficult as autonomous robots enter new domains, with model-based methods facing generalizability and scalability challenges and data-driven methods lacking well-understood guarantees.
Method
The paper provides a unified review that compares safety-filter approaches through common terminology, shared structure, and a theoretical framework.
Results
The review exposes a shared modular structure across seemingly disparate safety-filter classes and provides a common basis for comparing them.
Takeaways & Limitations
Safety filters can separate task-oriented control from safety enforcement by preempting unsafe actions while preserving viable task performance.
Takeaways & Limitations
A least-restrictive safety filter is generally intractable for practical robotic systems, so current methods trade off permissiveness, robustness, generality, and computational scalability.
Abstract
from arXiv · showhide
Recent years have seen significant progress in the realm of robot autonomy, accompanied by the expanding reach of robotic technologies. However, the emergence of new deployment domains brings unprecedented challenges in ensuring safe operation of these systems, which remains as crucial as ever. While traditional model-based safe control methods struggle with generalizability and scalability, emerging data-driven approaches tend to lack well-understood guarantees, which can result in unpredictable catastrophic failures. Successful deployment of the next generation of autonomous robots will require integrating the strengths of both paradigms. This article provides a review of safety filter approaches, highlighting important connections between existing techniques and proposing a unified technical framework to understand, compare, and combine them. The new unified view exposes a shared modular structure across a range of seemingly disparate safety filter classes and naturally suggests directions for future progress towards more scalable synthesis, robust monitoring, and efficient intervention.
1. INTRODUCTION
The review addresses safe control for autonomous robots by organizing model-based and data-driven approaches around safety filters and a shared technical framework. It introduces common terminology, formal safety requirements, and a structure for comparing methods under uncertainty.
- Motivation: Robotic autonomy requires rigorous safe control across diverse environments, motivating both model-based methods and data-driven safety schemes.Examples include control barrier functions, model predictive control, reachability analysis, learned certificates, adversarial reinforcement learning, and self-supervised safety analysis.
- Purpose: The article serves as an introduction, a synthesis of recent safety-filter advances, and a reflection on challenges and future research directions.It targets newcomers, researchers and practitioners, and the broader technical community.
- Novelty: Unlike surveys focused on particular methodologies, this review provides a unified conceptual and theoretical framework for identifying common features across safety-control research.The framework is intended to distill key ideas from a broad set of recent advances.
- Organization: The review introduces common terminology and notation, surveys safety-filter classes, and develops a theoretical foundation centered on the Universal Safety Filter Theorem.It also motivates safety-performance separation and discusses future challenges.
- System Model: The system model represents robot and environment state x, control input u, and disturbance d in discrete time as x_k+1 = f(x_k, u_k, d_k).Disturbances capture predictive uncertainty from exogenous factors, changing operating conditions, and modeling error; discrete-time reasoning requires attention to safety between steps.
- Safety Specifications: Safety-critical requirements forbid entry into a failure set F and generally require avoiding it for all future times.The complement F^c is the constraint set, but long-horizon safety is expressed as x_k ∉ F for every k ≥ 0.
Safe Set vs. Constraint Set
The explicit constraint set excludes current failure states, whereas the safe set also excludes states from which future violations cannot be prevented.
- Safe Set vs. Constraint Set: The constraint set F^c is not generally a safe set because some non-failure states already make future safety violations unavoidable.A drone flying toward an obstacle too quickly to avoid it illustrates this distinction.
Safety vs. Stability
Safety and stability are distinct: stability tools can sometimes establish safe invariant regions, but useful safety requires sets large enough to support operation. Robust and probabilistic formulations differ in how they model uncertainty and tolerate failure.
- Safety vs. Stability: Stability and safety are distinct properties, so neither generally guarantees the other.A stable system may be unsafe, while a safe system may not be stable.
- Safety vs. Stability: Finding a practically useful safe set is harder than finding some safe set, motivating maximal safe sets and optimal safety policies.Trivially safe behaviors such as never moving can prevent operation rather than support task performance.
- Safety vs. Stability: Robust safety requires guarantees for all uncertainty realizations in a bounded set, whereas probabilistic safety imposes chance constraints under a distribution.Probabilistic formulations may allow catastrophic failures with small nonzero probability.
- Safety vs. Stability: A safe set is a region from which a robot can ensure safety at all future times.Controlled invariance provides the mechanism for maintaining membership in the set.
- Safety vs. Stability: Probabilistic guarantees fit high-uncertainty, limited-stakes settings better than high-stakes domains where rare long-tail failures are consequential.The paper contrasts robot vacuums with autonomous driving and notes that extreme events can escape statistical assessment.
- Safety vs. Stability: A least-restrictive safety filter can preserve viable task performance by overriding only controls that would immediately leave the maximal safe set.The filter can be designed independently of task objectives, but such filters are generally intractable for practical robotic systems.
Safe Learning
Safety filters separate safety enforcement from task learning, allowing safety-agnostic policies to explore while a filter prevents physical or environmental damage. The ideal separation is conceptually powerful but difficult to obtain tractably in most practical systems.
- Safe Learning: Safety filters let conventional safety-agnostic learning algorithms explore and update task policies without damaging the physical system or its environment.The filter supplies the safety mechanism while learning focuses on task behavior.
- Safe Learning: A task-agnostic, least-restrictive filter can shield the system from failure states while preserving optimal viable task performance.Unsafe actions are preempted, while safe actions remain available to the task policy.
- Safe Learning: Least-restrictive safety filters are generally intractable for practical robotic systems, forcing tradeoffs among permissiveness, robustness, generality, and computational scalability.The review uses these tradeoffs to organize prominent safety-filter classes and develop a unified framework.
3. SAFETY FILTERS
Safety filters share a runtime structure built around monitoring candidate controls, intervening when needed, and using a fallback safety strategy. They can be compared by monitoring, intervention, fallback synthesis, and the guarantees they provide.
- 3. SAFETY FILTERS: Every safety filter can be understood through complementary runtime functions: monitoring the proposed action and intervening to maintain safety.Intervention may modulate or completely override the candidate control input.
- 3. SAFETY FILTERS: Many safety filters use a safety-oriented fallback policy to inform both monitoring and intervention.The fallback can serve as a potential alternative to the task policy.
- 3. SAFETY FILTERS: The review categorizes safety filters by monitoring methodology, intervention scheme, fallback-policy synthesis, and safety assurances.These four characteristics organize the main classes summarized in Figure 3.
- 3. SAFETY FILTERS: Monitoring approaches are value-based, evaluating safety-encoding functions, or rollout-based, predicting the system’s potential evolution explicitly.The distinction concerns how the filter assesses the safety of the task policy’s proposed control.
- 3. SAFETY FILTERS: Intervention commonly switches between the task and fallback policies or optimizes control for task performance subject to safety.The intervention scheme modifies the candidate action only as needed to satisfy the safety requirement.
Games of Kind and Games of Degree
Safety-filter methods can be organized around how they characterize safety, synthesize fallback strategies, monitor risk, and intervene. The review contrasts maximally permissive but computationally demanding approaches with more tractable, potentially restrictive or weakly guaranteed alternatives.
- Conceptual foundations: Isaacs distinguished games of kind, which have categorical winners, from games of degree, which optimize over a continuum of outcomes.He also showed that games of kind can be represented implicitly as games of degree.
- Maximal safe sets: HJ reachability formulates worst-case safety as a pursuit-evasion game and seeks the maximal safe set Ω∗, also called the viability kernel.Its safety value function can be computed by backward induction over a compact state-space domain, typically using a numerical grid.
- Runtime filtering: The HJ value function and policy support value-based monitoring with switch-type intervention, overriding the task policy near the boundary of Ω∗.This intervention prevents the safety value from becoming negative but can produce chatter and erratic motion when task policies repeatedly leave the safe set.
- Scalability limits: HJ computation has complexity and memory requirements exponential in state dimension, limiting general nonlinear applications to systems with no more than six continuous state variables.Decomposition, hardware design, learning, optimization, and runtime aggregation of local fallback policies have been explored to improve tractability.
- Alternative fallback strategies: CBFs seek a smaller safe set Ω⊂Ω∗ and provide smooth intervention through a differential inequality that keeps the encoded safety function from decreasing too quickly.Finding valid, computationally tractable, and data-efficient CBFs remains difficult because existing synthesis methods often work only for restricted dynamics.
- Unified comparison: Safety filters are compared by their monitoring methodology, intervention scheme, fallback-strategy synthesis, and resulting safety assurances, while tube MPC and learned critics retain scalability or uncertainty limitations.Tube MPC can be computationally intensive for high-dimensional systems, and neural-network critics provide no guarantees unless distributional assumptions support probabilistic guarantees.
4. UNIFIED SAFETY FILTER THEORY
The unified theory models safety filters as modular combinations of a fallback policy, safety monitor, and intervention scheme. This framework yields general safety guarantees and recovers several established filters as special cases.
- The formalism covers robust safety filters under uncertainty and exposes a shared modular structure for reconceptualizing existing and designing new filters.The analysis focuses on robust uncertainty while noting analogous probabilistic counterparts.
- A safety filter combines a fallback policy, safety monitor, and intervention scheme to preclude monitored-unsafe actions whenever the fallback policy passes its own check.
- A safety monitor certifies candidate actions when the fallback policy is guaranteed to maintain safety after those actions.The monitor provides a one-way safety implication rather than a general performance assessment.
- Theorem 1 guarantees the safety condition for all possible task policies when the initial information state passes the fallback-policy safety check.The filtered dynamics are defined by applying the intervention scheme to the intended action.
- The framework rederives least-restrictive HJ, CBF, robust model-predictive shielding, and unmapped-environment model-predictive filters as safety-preserving corollaries.The HJ and CBF corollaries guarantee all-time safety from initially safe states; the unmapped-environment filter uses braking within currently known free space.
- CBFs provide smooth intervention but accept suboptimal safe sets and lack general synthesis methods, motivating combinations with approaches offering constructive mechanisms.The review highlights approximations and online tuning as directions for improving CBF-based filters.
5. DISCUSSION
The review unifies a growing family of task-agnostic safety filters that convert arbitrary control policies into safe ones. It compares their shared structure and identifies opportunities for further safety-enforcing control research.
- The review presents a unified perspective on task-agnostic tools that filter arbitrary control policies into safe policies.
- The unified perspective exposes shared structure and compares the main features of safety-filter approaches.
- The paper concludes by discussing new research opportunities in safety-enforcing control.
1. Safety filters that account for runtime learning.
Safety filters for partial observability or unknown intent often use static information states, which can make monitoring and intervention overly conservative. Future filters must tractably incorporate runtime learning capabilities.
- Static information-state assumptions in partially observable or unknown-intent settings can produce overly conservative monitoring and intervention.
- Safe autonomy in information space requires filters that tractably account for the system’s runtime learning capabilities.
2. Safety filters that account for interaction.
Many safety filters assume that other agents are nonresponsive or adversarial, avoiding reliance on their actions for safety. Less conservative assurances for interactive applications remain an open technical and regulatory priority.
- Many safety filters assume nonresponsive or actively adversarial behavior from other agents to avoid relying on them for safety.
- Interactive contexts such as autonomous driving and robotic caregiving require practically acceptable assumptions for less conservative safety assurances.
3. Safety filters with high-dimensional representations.
High-dimensional sensory inputs and learned latent representations create an open challenge for safety-critical autonomy: connecting them to rigorous guarantees grounded in dynamical models.
- Autonomous robots increasingly use high-dimensional sensory input.
- Modern autonomy frameworks also rely on learned latent representations.
- Reconciling these representations with rigorous dynamical-model safety guarantees remains an open research challenge.
DISCLOSURE STATEMENT
The authors report no known affiliations, memberships, funding, or financial holdings that might affect this review’s objectivity.
- The authors report no affiliations that might affect the review’s objectivity.
- The disclosure reports no memberships that might affect objectivity.
- The authors report no funding or financial holdings that might affect the review’s objectivity.