Activated Backstepping for Safety-Critical Systems

• Activated Backstepping is a safety-critical control synthesis method that integrates a state-dependent activation function into traditional CBF backstepping to selectively apply corrections.
• The method minimizes conservatism by enforcing corrective actions only when the system’s state deviates from a virtual safe controller, resulting in larger admissible safe sets.
• Its efficacy is demonstrated in benchmarks such as the inverted pendulum and automated vehicle navigation, showcasing improved geometric interpretation and controller smoothness.

Searching arXiv for papers on “Activated Backstepping” and closely related backstepping-CBF work. Activated Backstepping is a safety-critical control synthesis method for control-affine systems with high-relative-degree safety constraints. The term is introduced explicitly in “Activated Backstepping with Control Barrier Functions for the Safe Navigation of Automated Vehicles” (Gacsi et al., 28 Aug 2025), where it is presented as an extension of control barrier function backstepping. Its defining feature is the incorporation of an activation function into the barrier construction so that the backstepping correction is applied only when necessary. When the current state already evolves “safely enough,” the barrier reduces to the original constraint function itself; when it does not, an additional correction term is activated. In the formulation reported in that work, this yields less conservative safe sets than standard CBF backstepping while preserving CBF validity and locally Lipschitz safe control synthesis (Gacsi et al., 28 Aug 2025).

1. Definition and terminological scope

Activated Backstepping belongs to the broader backstepping family for nonlinear systems, but it is not simply a synonym for any augmented or modified backstepping design. In the precise usage introduced in (Gacsi et al., 28 Aug 2025), it is a constructive method for synthesizing valid control barrier functions for systems with high-relative-degree safety constraints by extending CBF backstepping with a state-dependent activation mechanism.

The immediate methodological background is “Safe Backstepping with Control Barrier Functions” (Taylor et al., 2022), which gives a recursive CBF-backstepping framework for higher-order and strict-feedback systems. In that earlier construction, a top-level barrier $h_0$ is propagated through the cascade by subtracting quadratic penalties in the backstepping errors,

$$h(z_r)=h_0(z_0)-\sum_{i=1}^r \frac{1}{2\mu_i}\left\|\xi_i-k_{i-1,\xi}(z_{i-1})\right\|_2^2.$$

Activated Backstepping retains the virtual-controller logic of this lineage but changes the barrier construction so that the penalty is not applied uniformly across the state space (Taylor et al., 2022).

The term should also be distinguished from other nearby backstepping variants. Dynamic backstepping for pure-feedback systems augments virtual controls as dynamic states so that implicit algebraic equations are driven to zero asymptotically rather than solved explicitly (Zhang et al., 2017). Unified backstepping–sliding-mode designs may “activate” a switching term in the final step while preserving the recursive backstepping structure (Vieira et al., 2019). These are related in spirit, but they are not the same construction as Activated Backstepping in the CBF sense. This suggests that the most precise usage of the term is the CBF-based method of (Gacsi et al., 28 Aug 2025), rather than a generic label for all nonclassical backstepping schemes.

2. Problem class and mathematical setting

The method is developed for control-affine nonlinear systems

$\dot{x} = f(x) + g(x)u,$

with state $x \in \mathbb{R}^n$, input $u \in \mathbb{R}^m$, and smooth $f,g$. Safety is encoded through a set

$S = \{x \in \mathbb{R}^n : h(x) \ge 0\},$

where $h$ is a continuously differentiable function. A standard CBF $h$ must satisfy

$$\sup_{u \in \mathbb{R}^m} \dot{h}(x,u) > -\alpha(h(x))$$

for an extended class-$$h(z_r)=h_0(z_0)-\sum_{i=1}^r \frac{1}{2\mu_i}\left\|\xi_i-k_{i-1,\xi}(z_{i-1})\right\|_2^2.$$0 function $$h(z_r)=h_0(z_0)-\sum_{i=1}^r \frac{1}{2\mu_i}\left\|\xi_i-k_{i-1,\xi}(z_{i-1})\right\|_2^2.$$1, on an open set containing $$h(z_r)=h_0(z_0)-\sum_{i=1}^r \frac{1}{2\mu_i}\left\|\xi_i-k_{i-1,\xi}(z_{i-1})\right\|_2^2.$$2. Any locally Lipschitz controller satisfying

$$h(z_r)=h_0(z_0)-\sum_{i=1}^r \frac{1}{2\mu_i}\left\|\xi_i-k_{i-1,\xi}(z_{i-1})\right\|_2^2.$$3

for all $$h(z_r)=h_0(z_0)-\sum_{i=1}^r \frac{1}{2\mu_i}\left\|\xi_i-k_{i-1,\xi}(z_{i-1})\right\|_2^2.$$4 renders the closed-loop system forward invariant in $$h(z_r)=h_0(z_0)-\sum_{i=1}^r \frac{1}{2\mu_i}\left\|\xi_i-k_{i-1,\xi}(z_{i-1})\right\|_2^2.$$5 (Gacsi et al., 28 Aug 2025).

The paper focuses on safety constraints of relative degree two. For an output $$h(z_r)=h_0(z_0)-\sum_{i=1}^r \frac{1}{2\mu_i}\left\|\xi_i-k_{i-1,\xi}(z_{i-1})\right\|_2^2.$$6, the standing assumption is

$$h(z_r)=h_0(z_0)-\sum_{i=1}^r \frac{1}{2\mu_i}\left\|\xi_i-k_{i-1,\xi}(z_{i-1})\right\|_2^2.$$7

so the input does not affect $$h(z_r)=h_0(z_0)-\sum_{i=1}^r \frac{1}{2\mu_i}\left\|\xi_i-k_{i-1,\xi}(z_{i-1})\right\|_2^2.$$8 directly but does affect $$h(z_r)=h_0(z_0)-\sum_{i=1}^r \frac{1}{2\mu_i}\left\|\xi_i-k_{i-1,\xi}(z_{i-1})\right\|_2^2.$$9. The safety constraint is written as

$\dot{x} = f(x) + g(x)u,$0

where $\dot{x} = f(x) + g(x)u,$1 is a smooth scalar function (Gacsi et al., 28 Aug 2025).

The construction begins from the associated single-integrator safety problem

$\dot{x} = f(x) + g(x)u,$2

under the condition

$\dot{x} = f(x) + g(x)u,$3

A smooth virtual controller $\dot{x} = f(x) + g(x)u,$4 is then chosen so that

$\dot{x} = f(x) + g(x)u,$5

This virtual controller represents a safe evolution law for the output dynamics and serves as the reference motion against which the actual output velocity is compared (Gacsi et al., 28 Aug 2025).

3. Activated barrier construction

Standard backstepping CBF constructions penalize all deviations from the virtual safe dynamics. In the formulation reviewed in (Gacsi et al., 28 Aug 2025), the standard backstepping barrier has the form

$\dot{x} = f(x) + g(x)u,$6

or an equivalent scaled version. This guarantees validity under the paper’s assumptions, but it is conservative because it shrinks the safe set by penalizing all deviations from $\dot{x} = f(x) + g(x)u,$7, even where such correction is unnecessary.

Activated Backstepping modifies this point by introducing the switching signal

$\dot{x} = f(x) + g(x)u,$8

The activated barrier is then defined piecewise as

$\dot{x} = f(x) + g(x)u,$9

or, equivalently,

$x \in \mathbb{R}^n$0

where $x \in \mathbb{R}^n$1. The paper allows choices such as $x \in \mathbb{R}^n$2 for $x \in \mathbb{R}^n$3, $x \in \mathbb{R}^n$4, and the scaled form

$x \in \mathbb{R}^n$5

Here

$x \in \mathbb{R}^n$6

(Gacsi et al., 28 Aug 2025)

The activation signal has a direct geometric interpretation. It tests whether the actual output velocity $x \in \mathbb{R}^n$7 is better or worse than the virtual safe motion $x \in \mathbb{R}^n$8, relative to the gradient of the constraint. If $x \in \mathbb{R}^n$9, the state is evolving in a direction that is at least as safe as the virtual safe controller would produce, and the barrier collapses to the original constraint function: $u \in \mathbb{R}^m$0 If $u \in \mathbb{R}^m$1, the state is not evolving safely enough, and the corrective term is activated. This is the mechanism by which the method enlarges the admissible portion of the constraint set relative to standard backstepping CBFs (Gacsi et al., 28 Aug 2025).

A basic structural property reported in the paper is

$u \in \mathbb{R}^m$2

hence

$u \in \mathbb{R}^m$3

which yields

$u \in \mathbb{R}^m$4

Thus the activated safe set remains a subset of the original constraint set, but it is generally larger than the standard backstepping safe set because the correction vanishes on states already moving safely (Gacsi et al., 28 Aug 2025).

4. Validity theorem and comparison with adjacent CBF methods

The central theorem stated in (Gacsi et al., 28 Aug 2025) is that if the output $u \in \mathbb{R}^m$5 has relative degree two, the constraint function $u \in \mathbb{R}^m$6 is a CBF for the single-integrator $u \in \mathbb{R}^m$7, and the virtual controller $u \in \mathbb{R}^m$8 satisfies

$u \in \mathbb{R}^m$9

then the activated barrier $f,g$0 is a valid CBF for the original system. The proof uses the implication

$f,g$1

On that set the barrier reduces to $f,g$2, and the single-integrator safety inequality yields

$f,g$3

which verifies the CBF condition (Gacsi et al., 28 Aug 2025).

This theorem positions Activated Backstepping relative to three neighboring constructions. High-Order CBFs often use

$f,g$4

but the paper notes that this expression is not always a valid CBF because the condition $f,g$5 can fail at some states. Rectified CBFs introduce a ReLU-like correction

$f,g$6

with

$f,g$7

and can be valid if $f,g$8 is chosen carefully so that

$f,g$9

The critique given in the paper is that if $S = \{x \in \mathbb{R}^n : h(x) \ge 0\},$0 is too large, validity fails, and if $S = \{x \in \mathbb{R}^n : h(x) \ge 0\},$1 is too small, the controller may chatter or become non-Lipschitz. Standard backstepping CBFs are smooth and valid but conservative because they always penalize deviation from the virtual controller. Activated Backstepping is proposed precisely to retain validity and smoothness while reducing that conservatism, and it does so without the $S = \{x \in \mathbb{R}^n : h(x) \ge 0\},$2-tuning burden attributed to rectified CBFs (Gacsi et al., 28 Aug 2025).

The broader safe-backstepping literature clarifies the methodological shift. In (Taylor et al., 2022), safety is propagated through a strict-feedback cascade by subtracting quadratic tracking penalties at every stage. In (Kim et al., 2023), a multicopter with mixed-relative-degree and non-strict-feedback dynamics is reformulated into a strict-feedback-like chain, and time-varying safe backstepping is used to enforce safety through a quadratic program with affine inequality constraints. Activated Backstepping may be read as a refinement within this line of work: rather than altering the cascade structure or the recursive feasibility machinery, it changes the barrier itself so that the recursive correction is state-selective rather than omnipresent (Taylor et al., 2022, Kim et al., 2023).

5. Geometric interpretation and the inverted pendulum benchmark

The inverted pendulum example in (Gacsi et al., 28 Aug 2025) serves as the paper’s clearest geometric illustration. The dynamics are

$S = \{x \in \mathbb{R}^n : h(x) \ge 0\},$3

and safety requires the angle to remain within horizontal bounds,

$S = \{x \in \mathbb{R}^n : h(x) \ge 0\},$4

This is encoded by

$S = \{x \in \mathbb{R}^n : h(x) \ge 0\},$5

For this system, the HOCBF becomes

$S = \{x \in \mathbb{R}^n : h(x) \ge 0\},$6

Its Lie derivative with respect to the input vector field is

$S = \{x \in \mathbb{R}^n : h(x) \ge 0\},$7

At $S = \{x \in \mathbb{R}^n : h(x) \ge 0\},$8, one has $S = \{x \in \mathbb{R}^n : h(x) \ge 0\},$9, but then

$h$0

which can violate the necessary condition for large $h$1. This is the specific mechanism by which the paper shows that the HOCBF is not globally valid for the pendulum (Gacsi et al., 28 Aug 2025).

Standard backstepping yields the barrier

$h$2

which defines a rotated ellipse in the $h$3-plane. The paper treats this as a safe but conservative set because it excludes many states that remain recoverably safe under the actual pendulum dynamics (Gacsi et al., 28 Aug 2025).

The activated version is

$h$4

or equivalently

$h$5

with

$h$6

The resulting interpretation is that in quadrants where the pendulum is already moving safely toward the upright region, the safe-set boundary coincides with the original angle constraint; only in unsafe-motion regions does the barrier shrink. The paper describes the outcome as an unbounded safe set that is much less conservative than the rotated ellipse from standard backstepping (Gacsi et al., 28 Aug 2025).

This benchmark also clarifies a common misconception. Activated Backstepping is not merely “backstepping plus a smooth rectifier.” The decisive element is the use of the activation signal $h$7 to encode whether the current motion is already at least as safe as the virtual safe motion. The barrier is therefore not uniformly tightened everywhere; it is tightened only where the dynamics indicate an incipient loss of safety. That is the geometric reason the method can enlarge the certified set without abandoning CBF validity.

6. Automated-vehicle realization and broader significance

The application emphasized in the title of (Gacsi et al., 28 Aug 2025) is collision-free navigation for automated vehicles modeled by the kinematic bicycle system

$h$8

where $h$9 is position, $h$0 is yaw, $h$1 is speed, $h$2 is the steering-related input, and $h$3 is longitudinal acceleration. Obstacle avoidance is encoded by

$h$4

with output $h$5, which has relative degree two when $h$6 (Gacsi et al., 28 Aug 2025).

The construction proceeds by first designing a virtual safe controller for the single-integrator dynamics of $h$7,

$h$8

using a nominal velocity command

$h$9

and then building the activated barrier from the resulting $$\sup_{u \in \mathbb{R}^m} \dot{h}(x,u) > -\alpha(h(x))$$0. The real vehicle input is obtained through the standard QP safety filter

$$\sup_{u \in \mathbb{R}^m} \dot{h}(x,u) > -\alpha(h(x))$$1

with the desired lane-keeping controller

$$\sup_{u \in \mathbb{R}^m} \dot{h}(x,u) > -\alpha(h(x))$$2

In the reported simulation, the vehicle follows the lane, detects the obstacle, deviates safely around it, and then returns to lane-keeping. The paper reports that both $$\sup_{u \in \mathbb{R}^m} \dot{h}(x,u) > -\alpha(h(x))$$3 and $$\sup_{u \in \mathbb{R}^m} \dot{h}(x,u) > -\alpha(h(x))$$4 remain positive, the vehicle avoids collision, the activated barrier deviates from $$\sup_{u \in \mathbb{R}^m} \dot{h}(x,u) > -\alpha(h(x))$$5 only when the obstacle becomes safety-critical, and the controller subsequently returns to the nominal lane-keeping action (Gacsi et al., 28 Aug 2025).

Within the broader literature, this application places Activated Backstepping in the same safety-critical trajectory as earlier CBF-backstepping work. Safe Backstepping with CBFs (Taylor et al., 2022) provides the recursive certificate-building logic for strict-feedback systems, while “Safe Control Synthesis for Multicopter via Control Barrier Function Backstepping” (Kim et al., 2023) shows that mixed-relative-degree aerial-vehicle models can be reformulated so that safety constraints on angular velocity, thrust direction, velocity, and position become enforceable through affine QP constraints. Activated Backstepping contributes a different improvement: not a new vehicle model reformulation, but a less conservative barrier geometry for the same high-relative-degree safety setting (Taylor et al., 2022, Kim et al., 2023).

The main advantages explicitly claimed for Activated Backstepping are less conservative safe sets, valid CBF construction, smooth and Lipschitz safe controllers, no $$\sup_{u \in \mathbb{R}^m} \dot{h}(x,u) > -\alpha(h(x))$$6-tuning unlike rectified CBFs, computational savings in safe regions, and good geometric interpretability (Gacsi et al., 28 Aug 2025). The main tradeoffs are equally clear: it still requires designing a suitable virtual controller $$\sup_{u \in \mathbb{R}^m} \dot{h}(x,u) > -\alpha(h(x))$$7, it depends on a single-integrator safety design, the theorem is presented for relative degree two, and the safe set remains a subset of the original constraint set rather than the full constraint set itself (Gacsi et al., 28 Aug 2025).

In that sense, Activated Backstepping can be summarized as a gated backstepping-CBF construction: it preserves the recursive virtual-control philosophy of safe backstepping, but opens or closes the backstepping penalty according to whether the current motion is already safe relative to the virtual safe dynamics. The resulting synthesis occupies a specific and technically well-defined niche within contemporary safety-critical nonlinear control.