Decentralized Control Barrier Functions

• Decentralized Control Barrier Functions are safety frameworks that enforce local and high-level logic constraints in multi-agent systems using only neighbor state information.
• They use real-time quadratic programming to adjust nominal inputs, ensuring scalable, load-shared control that maintains safety even with local disturbances.
• These methods guarantee formal safety and temporal logic satisfaction while enabling adaptive and hybrid integrations with reinforcement learning.

Decentralized Control Barrier Functions (CBFs) provide a scalable methodology for enforcing safety and high-level logic constraints in multi-agent systems using only local state, communication, and computation. This class of approaches generalizes the standard Control Barrier Function framework from centralized, single-agent, or globally coordinated multi-agent scenarios to settings in which agents act independently, solving local optimization or feedback problems using only neighbor information. Decentralized CBFs enable compositional verification of set invariance and safety-critical properties for complex behaviors such as collision avoidance, temporal logic satisfaction, cooperative navigation, and more, and have become a foundational element in formal safe multi-agent control and learning.

1. Mathematical Foundation of Decentralized CBFs

A decentralized CBF framework begins from agent-level input-affine dynamics

$$\dot x_i = f_i(x_i, t) + g_i(x_i, t) u_i + c_i(x, t),$$

where $x_i \in \mathbb{R}^{n_i}$ is the state, $u_i \in \mathbb{R}^{m_i}$ the control, and $c_i(x, t)$ an unknown bounded disturbance. Agents communicate over an undirected graph $\mathcal{G} = (\mathcal{V}, \mathcal{E})$, where typically each agent has access to only local and neighbor states. The full-row-rank input map $\operatorname{rank} g_i(x_i, t) = n_i$ is required for well-posed CBF synthesis.

For each safety or logic specification—whether expressed as collision avoidance, STL, or other property—a (possibly time-varying) scalar barrier function $b^k(\bar x_k, t)$ is defined for a subset of agents $\mathcal{V}_k \subseteq \mathcal{V}$, with the associated safe set

$$\mathcal{C}_k(t) = \{\bar x_k : b^k(\bar x_k, t) \geq 0\}.$$

Forward invariance of $\mathcal{C}_k(t)$ under decentralized feedback is certified by a local CBF condition, formulated for each agent as a constraint on its control input and those of relevant neighbors. The key technical challenge is that these constraints become coupled, leading to potential feasibility or responsibility-sharing issues in dense or congested settings.

2. Local QP-Based Controller Synthesis

The most prevalent decentralized CBF realization is through small, real-time local quadratic programs (QPs) that project a nominal or task-driven input to the closest action satisfying all active barrier constraints. For each agent, this yields an online control law

$$u_i^* = \arg\min_{u_i} \ \|u_i - u_{i, \text{nom}}\|^2, \quad \text{s.t.} \quad L_{f_i} b^k + L_{g_i} b^k u_i + \alpha_k(b^k) \geq \mathcal{R}_i(x, t),$$

where the right-hand side $\mathcal{R}_i(x, t)$ encodes the load-sharing or negotiation factor, partial derivative weights, disturbance bounds, and potentially (in higher-level protocols) predictors and correctors for neighbor actions. In the general multi-agent case, decomposition of responsibility—such as by distributing the barrier load using

$$\frac{\| \partial_{x_i} b^k \|}{\sum_{d \in \mathcal{V}_k} \| \partial_{x_d} b^k \| },$$

as in (Lindemann et al., 2020), ensures that local QPs are feasible and that each agent contributes proportionally to constraint satisfaction.

When atomic predicates require globally minimum satisfaction—e.g., for temporal logic conjunctions—smooth under-approximations such as log-sum-exp mappings are introduced. Discontinuities in active set membership can result in piecewise-constant or switching feedback controllers.

3. Temporal Logic and Formal Behavioral Guarantees

Decentralized CBFs enable each agent (or group) to enforce complex temporal logic requirements such as eventuality ($F_{[a,b]}\psi$), always ($G_{[a,b]}\psi$), or until ($\psi_1 \mathcal{U}_{[a,b]} \psi_2$) properties. This is achieved by constructing time-varying CBFs,

$b_\ell(x, t) = h_\ell(x) - \gamma_\ell(t),$

where the reference trajectory $\gamma_\ell(t)$ ensures temporal satisfaction of each STL operator. Under appropriate construction, the forward invariance of the safe set for each barrier directly implies satisfaction of the underlying temporal logic specification, even in the presence of bounded disturbances (Lindemann et al., 2020).

The architecture supports encoding both atomic and compound predicates, and the decentralization naturally aligns with task allocations involving non-overlapping agent subsets, maintaining scalability even with complex STL task graphs.

4. Scalability, Load Sharing, and Robustness

A central attribute of decentralized CBFs is their computational scalability: each agent solves only a low-dimensional convex QP, independent of the total number of agents $M$—its complexity is at most linear in the agent's own state and the size of its local neighborhood. In representative benchmarks, decentralized CBF-QP controllers maintain constant per-agent CPU time (e.g., 0.4–0.5 ms/agent in 5–8 agent settings) while centralized mixed-integer programming (MIP) solvers required computational times two orders of magnitude larger and failed to find feasible solutions under strict deadlines (Lindemann et al., 2020).

Robustness to model uncertainties is guaranteed via conservative load sharing and the inclusion of disturbance bounds in the CBF constraint. Nonsmooth analysis, specifically Filippov’s differential inclusions, establishes well-posedness even when the decentralized control law is discontinuous due to active constraint switching.

The method is provably forward-invariant, ensuring that all constructed safe sets are maintained for all time, so long as the corresponding CBF inequalities hold at each step.

5. Asymmetric and Adaptive CBF Designs

In dense agent configurations (e.g., multiple robots converging or forming tight clusters), symmetric, decentralized CBF constraints can become mutually infeasible, generating “no-solution” conditions due to over-determined inequalities. Recent advances introduce asymmetric weighting—such as speed-dependent braking responsibility allocation—so that agents with greater ability to decelerate shoulder more of the safety constraint, restoring feasibility without requiring global coordination (Etchu et al., 2023). This is formally achieved through weight functions $w_{1,ij}$ and $w_{2,ij}$ that satisfy pairwise consistency, enabling each agent to solve its own QP independently while still preserving joint safety.

Adaptive schemes, including online tuning of CBF hyperparameters using decentralized, GNN-based reinforcement learning (RL) policies, allow agents to dynamically balance conservatism and aggressiveness as a function of local congestion. These RL-tuned CBFs achieve significant improvements in navigation efficiency and are able to solve scenarios that defeat any fixed-parameter approach (Gao et al., 2023).

6. Integration with Learning and Hybrid Decision Structures

Decentralized CBFs have been successfully embedded within safe multi-agent reinforcement learning frameworks as online “shields”: at each timestep, agent actions proposed by a deep RL policy are filtered through decentralized multiple CBF-QPs, ensuring invariance of joint safe sets independent of the exploratory or nominal behavior of the learning agent (Cai et al., 2021). This integration provides strong theoretical safety guarantees with minimal impact on RL learning performance, and forms a practical bridge between adaptive, model-free controllers and formal safety specifications.

Hybrid architectures have also emerged, combining decentralized CBF enforcement with centralized or edge-based constraint management. For example, “Barriers on the EDGE” applies a Watcher node for scalable constraint activation, pushing only relevant constraints to each agent, thus reducing onboard computational burden while avoiding communication overload (Sankaranarayanan et al., 2024). Such approaches interpolate between strictly decentralized and fully centralized paradigms.

7. Theoretical Guarantees and Application Domains

Theoretical analyses across diverse problem domains consistently show that decentralized CBF controllers guarantee forward invariance of designated safe sets and, under appropriate conditions, robust satisfaction of temporal and spatial logic specifications even with additive model disturbances and switching local active sets (Lindemann et al., 2020, Li et al., 14 Jul 2025).

Application domains include multi-robot coverage, cooperative herding of mixed groups, real-time multi-vehicle merging in intelligent transportation systems, and collision-free operation in heterogeneous aerial-ground settings. Representative large-scale experiments and simulations, such as 20-vehicle merging and 20-agent cooperative groups, confirm both safety and objective satisfaction with significant improvements over existing centralized and rule-based baseline controllers (Deshpande et al., 28 Oct 2025, Li et al., 14 Jul 2025).

Summary Table: Key Features Across Recent Decentralized CBF Approaches

Approach / Domain	Decentralized QP?	Load/Constraint Sharing	Adaptive/Asymmetric?
STL Multi-Agent, Lindemann & Dimarogonas	Yes	Load sharing using partial gradients	Discontinuous, robust
Online CBFs via GNN-RL	Yes	Local CBFs, adaptive via RL-tuned parameters	Yes, learning-based
Asymmetric Weight (MARS)	Yes	Speed-weighted braking responsibility	Yes, provably robust
Edge Cluster CBFs (Aerial-Ground)	Hybrid	Centralized constraint activation, local QP	Selective decentralization
Safe MARL (MADDPG-CBF)	Yes	Multiple locally enforced CBFs/shields	Static/adaptive possible

Empirical performance, formal safety, and computational properties of decentralized CBFs in multi-agent systems are well established and continue to motivate ongoing research in optimal load-sharing, adaptive policy learning, hybrid system integration, and the handling of networked/variable communication topologies (Lindemann et al., 2020, Etchu et al., 2023, Gao et al., 2023, Li et al., 14 Jul 2025, Sankaranarayanan et al., 2024, Cai et al., 2021).