Introduction
As autonomous systems and robotics increasingly rely on learning-based controllers, such as Reinforcement Learning (RL), the challenge of ensuring safety becomes paramount. While these controllers can achieve high performance in complex tasks, they often lack the formal guarantees required for deployment in safety-critical environments. To address this, researchers have developed "safety filters"—modular components that sit between a controller and a system to intercept and modify unsafe commands.
However, most existing safety filters assume "perfect state feedback," meaning they operate under the assumption that the robot knows its exact position and velocity at all times. In real-world applications, robots must rely on noisy, partial measurements from sensors like cameras or GPS. This paper introduces the Robust Predictive Output-Feedback Safety Filter (RPOF-SF), a framework designed to provide safety guarantees for uncertain nonlinear systems where the full state is not directly observable.
The Problem: The Gap Between Theory and Reality
The primary bottleneck in deploying safety filters on physical hardware is the discrepancy between idealized state-feedback models and the reality of sensor noise and system disturbances. If a safety filter does not account for the uncertainty in its own state estimate, it may certify a command as safe that actually leads to a constraint violation or a collision.
Bridging this gap requires a filter that can simultaneously handle nonlinear system dynamics, external disturbances, and measurement noise. The RPOF-SF addresses this by integrating robust estimation with predictive control to ensure that safety is maintained even when the robot's knowledge of its environment is imperfect.
The RPOF-SF Framework
The proposed architecture consists of two primary components working in tandem to "wrap" around an arbitrary, potentially unsafe, primary controller:
1. Robust State Observer
The observer (which can be implemented as a Luenberger-like observer or a Moving Horizon Estimator) processes noisy sensor data to produce a state estimate. Crucially, it also provides a mathematically guaranteed error bound. This bound quantifies the maximum possible difference between the estimated state and the true state of the system.
2. Predictive Safety Filter
This module functions as an optimization-based supervisor. At each time step, it receives a suggested input from the primary controller. It then solves an optimization problem to find the closest possible input to the suggested one that is guaranteed to keep the system safe over a finite future horizon.
To manage uncertainty, the filter employs Constraint Tightening. By using the error bounds from the observer, the filter creates a "buffer zone" around obstacles and system limits. If the robot is highly uncertain about its position, the buffer zone grows, forcing the filter to be more cautious. As the estimate becomes more certain, the constraints relax, allowing for more agile movement.
Mathematical Foundations: Incremental Stability
To extend these guarantees to nonlinear systems, the authors utilize the concepts of incremental Input-to-State Stability (i-ISS) and incremental Input/Output-to-State Stability (i-IOSS). These mathematical tools allow the framework to track how estimation errors and disturbances propagate through the system's nonlinear dynamics over time.
By leveraging these incremental Lyapunov functions, the authors proved (Theorem 1) that if the safety filter finds a feasible solution at the initial time step, it will remain feasible and keep the system safe for all future time steps, a property known as recursive feasibility.
Experimental Validation
The framework was evaluated using a simulated uncertain nonlinear mass-spring-damper system. The task was to control the system using a raw sinusoidal input—an inherently "unsafe" command that ignores the system's physical boundaries.
Key Results: * Constraint Satisfaction: While the raw controller immediately caused the system to violate position and velocity limits, the RPOF-SF successfully modified the inputs to keep the system within the safe zone. * Robustness: The filter maintained safety despite the presence of additive state disturbances and measurement noise. * Efficiency: Qualitative analysis showed the system operating near the edge of the tightened constraints, suggesting that the filter is not overly conservative and allows the system to remain as performant as safety allows.
Discussion and Future Work
While the RPOF-SF provides a rigorous path toward safe robotics, it faces certain challenges. The reliance on worst-case error bounds can lead to conservatism; if the worst-case scenario is highly improbable, the robot may move more slowly than necessary. Additionally, solving nonlinear optimization problems in real-time requires significant computational resources, which may be a hurdle for low-power hardware.
Future research will focus on incorporating probabilistic learning to transition from "worst-case" safety to "high-probability" safety, potentially reducing conservatism. The authors also intend to explore using the RPOF-SF to guide RL agents during the training phase, enabling safer and more efficient exploration of complex environments.
Summary updated