Where to Put Safety? Control Barrier Function Placement in Networked Control Systems

Purely local safety filters may lead to undesirable behavior such as late obstacle avoidance, loss of feasibility, or deadlock situations.

“While effective in ensuring safety, purely local safety filters may lead to undesirable behavior such as late obstacle avoidance, loss of feasibility, or deadlock situations [4, 5].”

• cites [4] sourceNot checked — no full text of the source
• cites [5] sourcePartly supported

  The Control Barrier Functions section (p. 4) describes a safety filter that intervenes only when necessary. The quadrotor results (pp. 5–6) report oscillations, failures to maintain safety under some settings, and cases where the drone cannot reach its goal. The paper does not establish late obstacle avoidance, loss of filter feasibility, or deadlock as stated in the claim. (LLM, conf=0.76)

  Experimental — accuracy not yet measured

The architectural implications of enforcing safety locally or remotely have not yet been systematically analyzed.

“While recent work has considered potential architectures for CBF in networked [9] or multi-rate [10] setups, the architectural implications of enforcing safety locally or remotely have not yet been systematically analyzed.”

• cites [9] sourceSupported
• cites [10] sourceSupported

Remote optimization relies on delayed or predicted state information due to communication constraints.

“While remote optimization enables complex predictive planning, it relies on delayed or predicted state information due to communication constraints [8].”

• cites [8] sourceSupported

For a remote MPC-CBF connected through a network with measurement-to-actuation latency τ_k = τ̂ + r_k, if it is feasible and computes u_k ∈ U such that condition (15) holds for all l ∈ {1,...,N}, then it renders the safe set C forward invariant.

“Lemma 2 (Disturbance Tolerance of Remote MPC-CBF).”

• Also stated in: § A. Local CBF · p. 3 · plain

  “then it renders the safe set C forward invariant.”

• AssumptionThe authors state this as an assumption: The communication network delay can be decomposed into a known constant part τ̂ and an unknown uncompensated residual r_k bounded by 0 ≤ r_k ≤ r̄.
• AssumptionThe authors state this as an assumption: The additive disturbances w_k are bounded by ||w_k|| ≤ w̄.
• AssumptionThe authors state this as an assumption: The dynamics f is locally Lipschitz with constants L_x and L_u.
• AssumptionThe authors state this as an assumption: The barrier function h is Lipschitz continuous with constant L_h.
• AssumptionThe authors state this as an assumption: The prediction errors e_k and e_{k+1} collect all deviations between the states used in the CBF constraint and the true states, resulting from additive disturbances, delay-compensating prediction, and uncompensated delay.

Comparing the disturbance bounds derived in Lemmas 1 and 2 reveals that the remote placement yields a smaller admissible disturbance tolerance.

“Comparing the disturbance bounds derived in Lemmas 1 and 2 reveals that the remote placement yields a smaller admissible disturbance tolerance.”

For discrete-time CBFs, if the loss caused by prediction errors is smaller than the prediction-error margin Δ_k, then the CBF condition holds for the true state.

“As long as the loss caused by prediction errors is smaller than the prediction-error margin”

• Also stated in: § III. ERROR MARGIN OF DISCRETE CBFS · p. 3 · plain

  “the CBF condition holds for the true state.”

• cites [12] Agrawal, Discrete control barrier functions for safety-critical control of discrete systems with application to bipedal robot navigationNot matched

  Couldn’t be matched to a paper in Crossref, so it can’t be checked.

• AssumptionThe authors state this as an assumption: The communication network delay can be decomposed into a known constant part τ̂ and an unknown uncompensated residual r_k bounded by 0 ≤ r_k ≤ r̄.
• AssumptionThe authors state this as an assumption: The additive disturbances w_k are bounded by ||w_k|| ≤ w̄.
• AssumptionThe authors state this as an assumption: The dynamics f is locally Lipschitz with constants L_x and L_u.
• AssumptionThe authors state this as an assumption: The barrier function h is Lipschitz continuous with constant L_h.
• AssumptionThe authors state this as an assumption: The prediction errors e_k and e_{k+1} collect all deviations between the states used in the CBF constraint and the true states, resulting from additive disturbances, delay-compensating prediction, and uncompensated delay.

The larger command variation δ_u for the local CBF indicates that local safety is enforced through later, more abrupt corrections.

“The larger command variation δ u for the local CBF further indicates that local safety is enforced through later, more abrupt corrections.”

For τ̂ = 0, r̄ = 0, and l = 1, the remote and local admissible disturbance bounds are equal.

“For ˆ τ = 0 and l = 1 , we have S 1 = 1 and S 0 = 0 , hence G (0 , 1) = 1 . If also ¯ r = 0 , then d u = 0 and ¯ w r = ∆ = ¯ w l .”

Predictive safety under delay yields stricter admissible disturbance levels.

“We derived disturbance-tolerance bounds for local and remote CBF implementations and showed that predictive safety under delay yields stricter admissible disturbance levels.”

The change from 4 unsafe runs in scenario I to 18 unsafe runs in scenario IV highlights the reduced disturbance tolerance of predictive safety under delay.

“The change from 4 unsafe runs in scenario I to 18 unsafe runs in scenario IV clearly highlights the reduced disturbance tolerance of predictive safety under delay.”

For a local CBF with access to the current state and nominal input, if it is feasible and returns u_k ∈ U such that condition (13) holds, then it renders the safe set C forward invariant.

“Lemma 1 (Disturbance Tolerance of Local CBF).”

• Also stated in: § A. Local CBF · p. 3 · plain

  “then it renders the safe set C forward invariant.”

• AssumptionThe authors state this as an assumption: The additive disturbances w_k are bounded by ||w_k|| ≤ w̄.
• AssumptionThe authors state this as an assumption: The dynamics f is locally Lipschitz with constants L_x and L_u.
• AssumptionThe authors state this as an assumption: The barrier function h is Lipschitz continuous with constant L_h.

The feasibility of the remote MPC for a predicted state does not guarantee the feasibility of the local CBF for the true state due to disturbances.

“Similarly, the feasibility of the remote MPC for a predicted state ˆ x k does not guarantee the feasibility of the local CBF for the true state x k due to disturbances.”

• AssumptionThe authors state this as an assumption: The combined architecture is designed for a known disturbance bound w̄ and known delay τ.

The myopic nature of the local CBF can be inferred from the minimum distance to the safety boundary being two orders of magnitude smaller than for the predictive methods, even in low delay cases, highlighting late intervention.

“However, the myopic nature of the local CBF can also be inferred, as the minimum distance to the safety boundary is two orders of magnitude smaller than for the predictive methods, even in the low delay cases, highlighting late intervention.”

In the joint architecture, the predictor must account for the local CBF corrections to maintain consistency between predicted and applied inputs.

“Note that in the joint architecture, the predictor must account for the local CBF corrections to maintain consistency between predicted and applied inputs.”

• AssumptionThe authors state this as an assumption: The combined architecture is designed for a known disturbance bound w̄ and known delay τ.

A discrete-time CBF constraint is evaluated using a model-predicted successor state, so its safety certificate is only as reliable as the prediction on which it is based.

“We leverage the fact that a discrete-time CBF constraint is evaluated using a modelpredicted successor state. Thus, its safety certificate is only as reliable as the prediction on which it is based.”

Overall, the results confirm the qualitative trade-off between robustness and performance induced by CBF placement.

“Overall, the results confirm the qualitative trade-off between robustness and performance induced by CBF placement.”

The placement of CBFs in NCS fundamentally affects both safety and performance.

“This work shows that the placement of CBFs in NCS fundamentally affects both safety and performance.”

The proposed combined architecture provides a practical compromise in uncertain networked environments.

“The proposed combined architecture provides a practical compromise in uncertain networked environments.”

Remote MPC-CBF accumulates prediction errors through delay compensation and horizon propagation, so its admissible disturbance bound decreases with τ̂, r̄, and l.

“Remote MPC-CBF accumulates prediction errors through delay compensation and horizon propagation, so its admissible disturbance bound decreases with ˆ τ , ¯ r , and l .”