When deploying autonomous systems—such as self-driving cars, quadrotors, or robotic manipulators—into the real world, safety is non-negotiable. Traditionally, control theorists guarantee safety by defining a "safe set" (e.g., keeping a car within its lane and away from other vehicles) and enforcing hard constraints to ensure the system never exits this set.
To account for real-world unpredictability like sensor noise, wind gusts, or erratic human behavior, standard control frameworks add a safety buffer. However, traditional approaches rely on static, worst-case safety margins. If this buffer is too small, a sudden disturbance can push the system out of the safe zone. If it is too large, the system becomes overly conservative—such as an autonomous vehicle refusing to merge into highway traffic because it cannot find a "perfectly safe" gap.
Worse yet, rigid safety boundaries frequently trigger optimization failures. When physical hardware limits (like motor saturation) conflict with strict, unyielding safety constraints, the underlying mathematical solver finds no valid solution. The system reaches a state of control infeasibility, effectively locking up when it needs to act most.
To solve this, a powerful paradigm shift has emerged: dynamic safety margins. Instead of treating safety boundaries as rigid, fixed envelopes, modern control frameworks parameterize these boundaries, allowing them to expand, contract, and soften in real time based on system states, environmental context, and measured uncertainty.
Parameterizing the Barrier: Adaptive Control Barrier Functions
To understand how safety margins can be made dynamic, we must first look at Control Barrier Functions (CBFs). Intuitively, a CBF acts as a mathematical fence. If $h(x) = 0$ represents the boundary of the safe set, a CBF enforces a condition that prevents the system from crossing it:
$$\dot{h}(x) \geq -\alpha(h(x))$$
Here, $\alpha$ is a function that dictates the "decay rate"—effectively controlling how fast the system is allowed to approach the safety boundary. In traditional applications, $\alpha$ is a fixed, pre-tuned parameter.
The core insight of dynamic safety margins is that the decay rate should be an optimization variable, not a constant. As shown by Bike Zhang et al., treating the safety constraint's decay rate as a flexible decision variable within a quadratic program guarantees point-wise feasibility. When a robot approaches an obstacle and its motors hit their physical limits, a rigid CBF would fail to find a valid command. By allowing the decay rate to scale dynamically, the controller can adaptively relax or tighten its safety margins just enough to maintain a continuous, valid path forward without sacrificing collision avoidance.
This concept extends naturally to high-order and probabilistic settings. Van Koevering et al. and Liu et al. demonstrate that treating the parameters governing both the shape and aggressiveness of a safety barrier as decision variables allows the controller to co-optimize safety alongside the control input at every time step. If the environment becomes tight or complex, the safety margins morph dynamically to preserve a feasible solution space.
Dynamic Tube Scaling in Predictive Control
A parallel evolution has occurred within Model Predictive Control (MPC), specifically in Tube-based Robust MPC. In this framework, the controller plans a nominal trajectory through an environment and wraps it in a geometric "tube." The size of this tube represents the maximum possible deviation the system might experience due to disturbances. The controller then tightens its operational constraints by the width of the tube to ensure that even with worst-case noise, the physical system remains safe.
Historically, calculating these tubes required conservative, infinite-time assumptions, resulting in unnecessarily large buffers that restricted system performance.
Static Tube MPC:
[ Start ] ====================( Fixed Tube Width )====================> [ Goal ]
Dynamic Tube MPC:
[ Start ] ======( Narrow Tube )======> ( Wide Tube ) =====( Narrow )====> [ Goal ]
(Low Uncertainty/Tight Space) (High Noise) (High Precision)
To unlock the full capability of the hardware, Lopez et al. proposed Dynamic Tube MPC, which treats the geometric bounds of the safety envelope as online decision variables. This allows the safety tube to adapt dynamically to the local environment: shrinking when navigating narrow passages where high precision is required, and expanding in open spaces where the system can tolerate more variance.
By leveraging modern control frameworks like System Level Synthesis (SLS), the authors of recent work have found ways to optimize these tubes with unprecedented flexibility. Sieber et al. and Aboudonia and Lygeros demonstrate that by decoupling the nominal trajectory planner from the error-correcting feedback loop, a system can jointly optimize feedback gains, tube dimensions, and constraint tightening levels online.
To prevent this optimization from over-approximating error, Schwenkel et al. introduced a dynamic error-bounding mechanism. By using an auxiliary, stable scalar system to track how much past disturbances have actually excited the system, the controller dynamically scales its safety margins based on historical reality rather than a theoretical, worst-case future.
Context-Aware and Learned Safety Boundaries
How does a system know exactly how much to expand or contract its safety margins? The answer lies in pairing these flexible control structures with online learning and real-time disturbance estimation.
If a system has to assume a static, worst-case disturbance (like maximum possible wind), it will always behave conservatively. However, by incorporating real-time estimators, the safety boundary can adapt to the environment's actual risk level:
- Disturbance-Driven Adaptation: Yang and Xie integrated a disturbance observer directly into the parameterization of a CBF. Instead of maintaining a fixed buffer for estimation errors, the safe operating region dynamically reshapes itself based on the direction and magnitude of external forces, significantly smoothing control inputs.
- Uncertainty-Aware Learning: When operating in entirely unknown environments, Sihua Zhang et al. utilize Gaussian Process (GP) regression to learn system disturbances online. This feeds into an Input-to-State Safety structure that automatically relaxes safety margins when the system is far from hazards and tightens them aggressively as it approaches a boundary or enters an area with high model uncertainty.
- Parameter Drift Tracking: Physical properties can change mid-operation—such as a drone picking up a heavy payload or experiencing a partial motor failure. Rui Chen demonstrated that by treating safety index synthesis as an online geometric optimization problem, safety logic can adapt to drifting physical parameters in milliseconds, preserving total feasibility where static safety indices cause catastrophic control failures.
By shifting from rigid, worst-case boundaries to parameter-conditioned, state-dependent safety envelopes (as explored by Borquez et al.), autonomous agents gain the ability to balance safety and performance fluidly. They modulate their conservativeness in real time based on their own confidence, the environment's geometry, and the behavior of surrounding agents (e.g., Berducci et al.).
Why It Works: The Intuition of Feasibility Preservation
Ultimately, dynamic safety margins work because they resolve the fundamental conflict between safety guarantees and numerical optimization.
In a traditional safety-filtered system, the controller is forced to solve a constrained optimization problem where the safety boundaries are hard walls. If a disturbance pushes the system too close to a wall, or if actuator saturation prevents the system from generating enough force to stay away from it, the mathematical constraints become mutually exclusive. The optimization problem returns an error, and the system fails.
By transforming safety margins into time-varying, state-dependent, or optimized parameters, the control loop gains a mathematical "release valve." When the system is safe and operating normally, the margins remain tight or optimal to maximize performance. If the system enters a high-risk, highly constrained state, the framework can dynamically scale the decay rates, adjust the tube shapes, or interpolate toward a guaranteed safe reference point (as seen in the blending mechanisms of Constante Flores et al.).
By making the safety boundaries as dynamic as the physical systems they govern, this approach ensures that a valid, mathematically sound path forward always exists.