WorksComputer Science
Ontology-Compiled Control Barrier Functions for Heterogeneous Robot Teams
Abstract
Safety filters for heterogeneous robot teams fail because of missing constraints, overestimated braking capabilities and conflicts between individually satisfiable constraints. We present OntoCBF, which compiles a domain ontology with a class taxonomy, typed relations, safety axioms and a dynamics profile per robot class into a control barrier function filter with one quadratic program per agent. Description-logic entailment supplies the constraints of any team online, including those that follow from chains of relations and from classes defined by what a robot carries, so the existing safety axioms cover new robot classes, payload classes and combinations. We prove that a subclass may reuse the braking capability of its parent whenever it can produce every velocity of the parent, a property separate from subsumption. For box-bounded double integrators this property is also necessary, and every loss of safety then begins with a conflict. Once three robots interact, the safe set of pairwise barriers contains states from which a collision is unavoidable, and such states persist in larger teams under any fixed capability reduction. In a simulated warehouse where humans ignore the robots, a filter written for the design classes violates safety in 89-100 % of the episodes with unseen classes. In all scenarios, including one with 32 robots, OntoCBF violates safety in 20-70 % of the episodes, each time after a conflict, with a median deepest intrusion of 10-17 % of the keep-out distance. A multi-agent PPO policy trained with the filter in the loop lowers the rate of infeasible quadratic programs, nearly all of them conflicts, by 18-41 %.
Introduction
Warehouses, hospitals and construction sites increasingly operate heterogeneous robot teams in spaces shared with people. Their safety requirements depend on what each entity is and on how the entities are related. A drone keeps farther from a person than a ground robot does, a carrier is responsible for the distance between its payload and nearby people, and ground robots avoid restricted zones that drones may cross. Industrial standards state such requirements in the same terms, through separation distances that depend on the type of machine and on its stopping performance ([18] (ISO 2016)). Safety filters based on control barrier functions (CBFs) enforce requirements of this kind on top of a learned or hand-designed controller with minimal changes to its commands ([1, 34]), and they are a common safety layer for multi-agent reinforcement learning (MARL) ([7] (Zhang et al. 2025b)).
In existing multi-agent safety filters, the constrained pairs, the keep-out distances and the braking capabilities are fixed at design time for the robot types known at that point ([34, 7, 11]). Learned graph CBFs for collision avoidance generalize across team sizes and, empirically, across robot models [19] (Zhang et al. 2025b). Semantic safety filters query language models for constraints, which leaves the completeness and consistency of the constraint set unverified [6, 28]. Robot fleets change over their lifetime, and whenever a team gains a new robot model, payload type or task combination, an engineer has to determine by hand which constraints apply to the new entities and whether the existing safety certificate still holds.
We let a domain ontology, of the kind already standardized for robotics [15], carry the safety specification. A filter for a heterogeneous team can miss a constraint for a new entity, overestimate the braking capability of a robot, or contain constraints that are each satisfiable and exclude each other. In our warehouse ontology (Figure 1), the taxonomy states that a LightDrone is a Drone and that every Drone is Aerial. A safety axiom states how far an Aerial agent stays from a Human, and a property chain relates each carrier to the people near its payload. Entailment removes the first failure, because the constraints of unlisted classes, payloads and combinations follow from these statements. The second failure depends on dynamics, because a HeavyDrone, a Drone that brakes less strongly, inherits every constraint of Drone and needs a braking capability of its own. A test on the dynamics profiles in the ontology, together with classification, decides when a subclass may reuse the capabilities, and with them the certificate, of its parent. We prove that the third failure is intrinsic to pairwise barriers.

Figure 1. (a) OntoCBF. (b) Robot classes with per-axis acceleration limits (m/s), and payload classes. LightDrone (green) is a dynamic subtype of Drone. HeavyDrone (red) brakes less strongly and Forklift has no design ancestor, so both receive their own capability. These classes and Chemical appear only in tests.
Our contributions are the following.
We formulate safe control of heterogeneous teams whose active constraints are entailed by an ontology and switch online. Our method, OntoCBF, compiles these entailments into a CBF filter with one quadratic program (QP) per agent and trains a shared multi-agent PPO (MAPPO) policy with the filter in the loop.
Theorem 4 shows that a subclass inherits exactly the constraints of its parent under a classification test, and its capabilities under dynamic subtyping, an inclusion of velocity sets that braking barriers of box-bounded double integrators require.
Propositions 5–7 show that, with capabilities assigned this way, safety is lost only after a conflict, and that some states with positive pairwise barriers lead to a collision under every input, for three robots and, under any fixed capability reduction, for large enough teams.
Experiments in a simulated warehouse with unseen classes and up to 32 robots show that OntoCBF removes the first two failures, that every remaining violation follows a conflict, and that the learned policy lowers the rate of infeasible QPs by 18–41 % (§ 6).
Related Work
Multi-agent safety filters. Our barriers and responsibility split follow Wang et al. [33, 34], and Kim et al. [20] extend the split to kinematically heterogeneous robots. Sah and Keshavan [30] give an exact pointwise test of joint feasibility for heterogeneous teams. Learned graph CBFs enforce collision avoidance in teams of any size (Zhang et al., 2025b}), are trained jointly with PPO-style policies (Zhang et al., 2025a}), and now cover heterogeneous teams through type embeddings (Yang et al., [37]) or, for unseen robot models, through features of their reachable velocity sets (Kim et al., [19]). Constrained MARL bounds expected costs (Gu et al., [14]), and shields from temporal-logic specifications act on discrete abstractions (ElSayed-Aly et al., [11]; Xiao et al. [36]). Several works check or design the compatibility of multiple CBFs under input limits (Tan and Dimarogonas, [31]; Breeden and Panagou [4]; Isaly et al. [16]; Mousavi et al. [25]), and Proposition 7 shows that pairwise braking barriers are incompatible on parts of their own safe set.
Semantic safety and reuse of certificates. Semantic safety filters derive constraints from language or vision-language models (Brunke et al., [6]; Ravichandran et al. [28]), and logical specifications have been compiled into CBFs as compositions of state constraints (Lindemann and Dimarogonas, [21]; Molnar and Ames [23]). Ontologies model safety in human–robot collaboration (Dhanda et al., [10]) and have been used to verify robot plans (Wu et al., [35]), to shape risk fields for motion planning (van der Ploeg et al., [32]) and to mask the actions of a learner (Fatima and Tolg, [12]). To our knowledge, OntoCBF is the first method that compiles entailments into control constraints with an invariance guarantee. Behavioral subtyping requires a subtype to preserve what clients of the supertype rely on (Liskov and Wing, [22]). A system that admits at least the inputs of another can reuse its controllers (Reissig et al., [29]), and barriers transfer between systems through reduced-order models (Cohen et al., [9]), learned inverse dynamics (Nadali et al., [26]) or a simulation function with a margin (Bousias et al., [3]). Part 2 of Theorem 4 is the zero-margin, class-level form of such a transfer. Parts 1 and 3 add what an ontology requires, namely a classification test for identical constraints and, for box-bounded double integrators, the necessity of dynamic subtyping for braking barriers (Table 1).
| Line of work | Team | Het. | Sem. | Spec. | Switch | New | Reuse | Learn |
| Multi-robot CBF filters ([34, 20, 30]) | ◦ | ✓ | ✗ | ✗ | ◦ | ✗ | ✗ | ✓ |
| Learned graph CBFs ([37, 19] (Zhang et al. 2025b)) | ◦ | ✓ | ✗ | ✗ | ◦ | ◦ | ✗ | ✓ |
| Safe MARL and shields ([14, 11, 8]) | ◦ | ✗ | ◦ | ◦ | ✗ | ✗ | ✗ | ✓ |
| Logic compiled into CBFs ([21, 23]) | ◦ | ✗ | ◦ | ✓ | ✓ | ✗ | ✗ | ✗ |
| Constraints from (V)LMs ([6, 28]) | ✗ | ✗ | ✓ | ✗ | ◦ | ◦ | ✗ | ◦ |
| Ontologies for robot safety ([35, 32, 12]) | ✗ | ✗ | ✓ | ✓ | ✗ | ◦ | ✗ | ◦ |
| Certificate transfer ([29, 3]) | ✗ | ✓ | ✗ | ✗ | ✗ | ◦ | ✓ | ✗ |
| OntoCBF (ours) | ◦ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ | ✓ |
Table 1. Related work compared on team certificate, class-specific dynamics, typed relational constraints, formal specification, guarantee across switches, unseen types without redesign, certificate-reuse condition and learning in the loop (✓ addressed, ◦ partly, ✗ absent).
Problem Setting
Entities and dynamics. The workspace contains controlled agents , humans , restricted zones and payloads . Each entity belongs to a class , a concept name of the ontology, and we call the restriction of to the team composition. An agent of class has dynamics with compact. In our experiments it is a planar double integrator, , , with . Humans are double integrators with unknown accelerations bounded by , an agent carries at most one payload, which sits at for a fixed offset , and zones are static.
Ontology. We use the description logic of OWL 2 ([2]), in which an ontology consists of a TBox of terminological axioms and an ABox of facts about individuals. Our TBox contains the taxonomy of Figure 1b as concept inclusions , disjointness axioms such as , the symmetric role near, the role carries from ControlledAgent to Payload, and the role nearCarried, which the property chain carries near nearCarried derives from carries and near. Payloads have their own taxonomy, Chemical HazardousPayload Payload, and the defined class HazardCarrier ControlledAgent carries.HazardousPayload contains whichever agent currently carries a hazardous payload. Each robot class is annotated with a dynamics profile, here and a nominal speed . We write if holds in every model of , which HermiT decides soundly and completely for OWL 2 DL ([13]).
Safety axioms and active constraints. A safety axiom requires that the distance between and stays at least whenever , and are entailed. For the distance is measured from the payload of . The ABox at time ,
changes when an entity enters or leaves the sensing radius of another and when an agent picks up or drops a payload. The constraints active at time are the entailed instances of the safety axioms,
and the safe set is , where is the position of , or of its payload if . Since a subclass inherits every axiom attached to its ancestors, the safe set depends on the team composition through entailment.
Barrier functions. For a constraint let , , and . We use the braking barrier of Wang et al. [33],
where is the braking capability available to the pair. If , braking at rate along stops the approach at a distance of at least , and is continuously differentiable wherever . Since is affine in the accelerations, the condition with reads
Velocity sets and certificates. For a class , is the set of state derivatives, called velocities as in viability theory, that an agent of class can produce at . For a team composition let , and let be the set of admissible velocities of the humans.
Definition 1 (Dynamic subtype). A robot class is a dynamic subtype of a robot class , written , if both classes have the same state space and for all .
Definition 2 (Joint validity). A family of functions that are continuously differentiable on an open set containing is jointly valid for on if for every there is such that for all and all .
Subsumption is entailed by the TBox, and is a property of the dynamics profiles. For box-bounded double integrators holds if and only if , so HeavyDrone Drone holds and HeavyDrone Drone fails. A certificate for a team is a family that is jointly valid on a set , together with the capabilities and the gain used to build it.
Objective. Each agent is rewarded for reaching a sequence of goals and acts on a local observation. We seek decentralized policies and a filter that maximize the expected team return subject to for all , for any team and any classes that the ontology declares, including those absent at design time and in training.
Compiling the Ontology into a Safety Filter
OntoCBF consists of a reasoner that decides which constraints are active, a CBF filter that enforces them with one QP per agent, and a shared policy trained with MARL that performs the task (Figure 1a).
Compilation and capability assignment
Offline, HermiT classifies once, and OntoCBF stores the ancestors of every class, the safety axioms indexed by role, the defined classes and the dynamics profiles. The design classes, here Drone, Rover and Carrier, are the robot classes of the team for which a certificate is established when the system is designed, and each enters the barriers with the braking capability . Every other robot class inherits from its nearest design ancestor if is a dynamic subtype of (Definition 1) and has the same subsumers as in (§ 5), a test that classification decides. By Theorem 4, then also inherits the certificate of , provided both have the same payload offset. With several nearest design ancestors, both conditions must hold for each, and the smallest capability is used. All remaining robot classes, Forklift among them, receive and need a certificate of their own. For box-bounded double integrators the dynamic-subtype test reduces to , so LightDrone inherits from Drone and HeavyDrone receives its own capability. The rule yields for every class, which makes every constraint satisfiable on its own wherever its barrier is nonnegative (Lemma 3). Joint validity of all constraints (Definition 2) is a separate property, measured in § 6.2.
Event-triggered reasoning
The reasoner runs at ABox events, which occur when a sensed near assertion appears or disappears and when the task layer asserts or retracts carries at a pickup or a drop. It closes the asserted roles under symmetry and the property chain, realizes the defined classes, so that an agent carrying a hazardous payload becomes a HazardCarrier, and instantiates every safety axiom whose role and classes are entailed for a pair. Each instance becomes a constraint with keep-out , braking capability and a split of (4) among its controlled agents. For two robots, , and (4) is split into and with and [33]. For a robot and a human, who may accelerate toward the robot, , which is positive for all our classes, and the robot alone enforces . For a robot and a zone, , and the robot enforces . We call these linear inequalities the barrier inequalities of an agent. Rule saturation is complete for our TBox, agrees with HermiT on 72 random ABoxes and takes 0.06 ms per ABox against 332 ms. It also detects inconsistent ABoxes, which would leave empty, and OntoCBF then keeps the last consistent ABox and brakes (Appendix B). Reasoning is centralized over the shared ABox, and each QP needs the states and capabilities of the entities within the sensing radius. Activation check. A relation asserted by the task layer, such as carries at a pickup, can activate a constraint with , so the reasoner computes the constraints that the assertion would activate and defers the pickup until each has m/s.
Per-agent filter
Agent collects its barrier inequalities as half-planes and solves
over its true input set , so the assigned capability enters only through the barriers. Since , (5) is solved exactly over the edges and vertices of the feasible polygon. If the polygon is empty, the agent minimizes over with . The filter evaluates (3) with enlarged by a margin m and requires the relative acceleration along inside the margin. We log the single-constraint failures and conflicts that § 5 distinguishes. A single-constraint failure is a barrier inequality with that every input in violates, and a conflict is an empty polygon whose inequalities are each satisfiable on their own.
Learning with the filter in the loop
A policy shared by all agents is trained with MAPPO (Yu et al., 2022) in the design scenario W0 only, with the filter acting on its outputs. Agent observes its own state and goal and, for every entity within , the relative state, a class embedding and the keep-out distance that the reasoner currently entails for the pair, pooled by attention. Class embeddings are tied along the taxonomy, with for classes absent from training, so an unseen class is represented through its known ancestors (Appendix § 12). The policy outputs a bounded correction to a go-to-goal controller, with capped at the authority seen in training, a cap introduced after an earlier experiment with LightDrone (Appendix § 12). The reward is goal progress plus an arrival bonus minus penalties on the intervention , on infeasible QPs, on violations and on overspeed, with the gradient taken with respect to the pre-filter action . By Proposition 6, in continuous time the penalty on infeasible QPs targets the only failure mode left under OntoCBF.
Inheritance and Conflicts
Throughout this section and its proofs in Appendix A, the agent set is fixed, uncontrolled entities keep their classes, and is the TBox of Appendix B.
Lemma 3 (Single-constraint feasibility) Let agent have the input set and capability , and consider one of its barrier inequalities from § 4.2.
If , then at every state with and some input of satisfies it.
If , there is a state with at which every input of violates the inequality, and for a zone, or a human who accelerates toward the robot, there is a state with from which every admissible input of leads into the keep-out region.
Part 1 follows from on , and part 2 from a head-on approach along a coordinate axis. Let be the set of concept names that occur in safety axioms or in TBox axioms other than inclusions between two concept names.
Theorem 4 (Certificate inheritance) Let the team composition differ from on , and for every let , let and have the same subsumers in , the same payload offset and the same capability, .
For all near and carries relations, and yield the same constraints, barriers and weights.
If for all , every family that is jointly valid for on is jointly valid for on , and every QP (5) that is feasible under is feasible under .
For double integrators with box input sets and , every barrier inequality of agent is feasible at all states with and if and only if . If , the zero-superlevel set of an inherited zone or human barrier contains states from which every admissible input of agent leads into the keep-out region.
Proof sketch. Entailment is monotone and equal subsumers in add no axiom instance (Lemma 8), so the inequalities are unchanged, and any witness or feasible QP input for remains one for because the velocity sets grow. Part 3 is Lemma 3 with capability and bound .
Part 3 shows that, for box-bounded double integrators, the capability of the parent keeps every inherited barrier inequality feasible exactly when the subclass is also a dynamic subtype. Lemma 8 applies to any entity, so a payload subclass with the same subsumers in , such as Chemical, leaves the active constraints unchanged. Typing a Payload as HazardousPayload changes them, since its carrier becomes a HazardCarrier. Appendix A extends the theorem to other state spaces through a map with , as for a unicycle, and to inclusions up to a margin.
Assumption 1 (Activation margin) Every constraint entering at time has .
As relations switch, Assumption 1 holds for sensed relations if closing speeds are bounded by and , because a constraint activates at distance (Lemma 11). At nominal speeds our pairs need at most 6.34 m, for a HeavyDrone and a human, and we use m. For relations asserted by the task layer, the activation check enforces it.
Proposition 5 (Invariance under switching). Suppose that changes finitely often on bounded intervals, that Assumption 1 holds, and that for all . If the inputs applied along an absolutely continuous closed-loop trajectory satisfy every barrier inequality of every agent for almost every , then for all .
The proof applies the comparison lemma between switches, which keeps every active positive and every distance above the keep-out, where is differentiable.
Proposition 6 (Loss of safety starts with a conflict). Let the first three hypotheses of Proposition 5 hold, let the closed-loop trajectory be absolutely continuous, let every class satisfy , and let every agent apply the solution of its QP whenever that QP is feasible. Suppose that the state first leaves the zero-superlevel set of an active barrier at . Then some constraint , active on with , has , and some agent of has an infeasible QP on a subset of of positive measure, with each of its barrier inequalities feasible on its own at those times.
Before every active barrier is nonnegative, so Lemma 3 makes each inequality feasible, and an infeasible QP can only come from inequalities that exclude each other. Such conflicts are a property of pairwise barriers, independent of the ontology.
Proposition 7 (Unavoidable conflicts). Consider three double integrators with input sets , capabilities and the robot–robot barriers (3) with and keep-out . There is a state with all three barriers positive from which every admissible input of the robots brings two of them closer than .
Two robots approach a third from opposite sides, and each pairwise barrier assumes that the middle robot devotes its full braking to that pair. Hence no controller renders the zero-superlevel set of pairwise braking barriers forward invariant, and a certificate for them holds only on a strict subset of it, which makes precise the remark of Wang et al. [34] that their QP may become infeasible in crowded scenes. A chain of robots extends the construction to capabilities whenever , so every fixed capability reduction leaves such conflicts in large enough teams.
Experiments
Setup. Each m warehouse tile (Appendix § 11) contains two drones that carry parcels across a restricted zone binding only Ground agents, four carriers with payloads, two of them hazardous, and two patrolling rovers. Its four humans ignore the robots, walk at up to m/s and accelerate at up to m/s. Besides the robot–robot (0.3 m), robot–human (0.5 m), aerial–human (1.0 m) and ground–zone (1.0 m) axioms, payloads keep 0.8 m and hazardous payloads 2.0 m from people. Episodes last 40 s with a zero-order hold of 0.05 s, whose inter-sample effect we do not bound (Breeden et al., [5]), m, and is selected on validation episodes. W0 is the design team (Drone, Rover and Carrier with Payload and HazardousPayload). W1 replaces the two drones by a LightDrone and a HeavyDrone and one carrier by a Forklift. W2 types the hazardous payloads as Chemical, W3 lets a LightDrone carry a Chemical and a Drone carry a HazardousPayload, two unlisted combinations, and W4 tiles W0 ().
Methods and metrics. All filters share the barrier (3), the QP (5) and, unless stated otherwise, the activation check, so they differ only in where their constraints and capabilities come from. None applies directly. Flat replaces the ontology by hand-written tables of keep-outs and capabilities for the design classes and (robot, payload) pairs, with defaults for unlisted classes (mean design capability 1.4 m/s, keep-outs 0.5 m and 0.8 m), and Flat-cons uses the most conservative listed values as defaults. No-chain drops property chains, Naive inherits capabilities along without the dynamic-subtype test, Untyped treats every payload as hazardous, and No-act. omits the activation check. Covering W1–W3 requires 10 ontology statements in OntoCBF and 18 table entries in Flat, and all methods use the same go-to-goal PD controller, except MAPPO, on 100 seeds per scenario. A violation is a distance below for a constraint entailed under the true classes, so Flat and No-chain are also judged on the constraints they omit.
Missing constraints
OntoCBF runs unchanged in all scenarios with the same safety axioms, and W2 reproduces W0 seed by seed, as Lemma 8 predicts. In W1 the HeavyDrone and the Forklift run with their own capabilities, for which a certificate is still to be established. Under OntoCBF, 20–70 % of the episodes contain a violation, and an episode has 5.6–47.9 violating steps out of 800 on average, against 100 % and 551–754 steps without a filter (Table 2 and Table 12, intervals in Table 5). Flat performs like OntoCBF on the design team (20 % against 22 %, where OntoCBF keeps one inequality per entailed axiom and Flat one per table entry) and violates safety in at least 89 % of the episodes of W1–W3, mostly through the omitted aerial–human axiom in W1 (76.2 violating constraint-steps per episode against 1.9) and the omitted hazard keep-out in W2 and W3. Flat-cons, which treats every unlisted robot or pair like the weakest and most restricted listed one, violates safety less often than OntoCBF in every scenario, most of all in W3 (28 against 40 %). Its zone default stops the unseen drones, so it achieves 65 and 91 % of the deliveries of OntoCBF in W1 and W3. Its capability default is valid only because every unseen class brakes at least as strongly as the weakest design class (Lemma 3). Untyped has fewer episodes with a violation in every scenario and achieves 66 % of the ordinary deliveries of OntoCBF in W0. Removing the property chain, the robot or payload taxonomy, or the defined class produces violations in at least 79 % of the episodes where the element is needed and reproduces OntoCBF seed by seed elsewhere (Table 7 and Table 8). Without the activation check, 217 pickups in W3 begin inside a keep-out region (Table 7).
| Scenario | Flat | Flat-cons | No-chain | Naive | Untyped | No-act. | OntoCBF | MAPPO | Flat | Naive | OntoCBF |
| W0 (design) | 20 | 20 | 100 | 22 | 16 | 74 | 22 | 12 | 0 | 0 | 0 |
| W1 (unseen robots) | 89 | 17 | 100 | 23 | 17 | 72 | 20 | 11 | 182 | 266 | 0 |
| W2 (unseen payload) | 99 | 20 | 100 | 22 | 16 | 74 | 22 | 12 | 0 | 0 | 0 |
| W3 (unseen combination) | 100 | 28 | 100 | 40 | 37 | 96 | 40 | 26 | 0 | 0 | 0 |
| W4 () | 67 | 67 | 100 | 70 | 69 | 98 | 70 | 65 | 0 | 0 | 0 |
Table 2. Episodes with a violation (left, 100 % for None everywhere) and single-constraint failures summed over episodes (right, zero for all other methods), 100 seeds per cell, all with the PD controller except MAPPO, which runs behind OntoCBF (3 training runs pooled).
| Scenario | None | Flat | Flat-cons | No-chain | Naive | Untyped | No-act. | OntoCBF |
| W0 (design) | 100 [96, 100] | 20 [13, 29] | 20 [13, 29] | 100 [96, 100] | 22 [15, 31] | 16 [10, 24] | 74 [65, 82] | 22 [15, 31] |
| W1 (unseen robots) | 100 [96, 100] | 89 [81, 94] | 17 [11, 26] | 100 [96, 100] | 23 [16, 32] | 17 [11, 26] | 72 [63, 80] | 20 [13, 29] |
| W2 (unseen payload) | 100 [96, 100] | 99 [95, 100] | 20 [13, 29] | 100 [96, 100] | 22 [15, 31] | 16 [10, 24] | 74 [65, 82] | 22 [15, 31] |
| W3 (unseen combination) | 100 [96, 100] | 100 [96, 100] | 28 [20, 37] | 100 [96, 100] | 40 [31, 50] | 37 [28, 47] | 96 [90, 98] | 40 [31, 50] |
| W4 () | 100 [96, 100] | 67 [57, 75] | 67 [57, 75] | 100 [96, 100] | 70 [60, 78] | 69 [59, 77] | 98 [93, 99] | 70 [60, 78] |
Table 5. Episodes with a violation (%) in the warehouse with 95 % Wilson intervals, PD controller, 100 seeds per cell.
| Scenario | Without | Measure | Ablation | OntoCBF |
| W0 (design) | property chain | payload+hazard viol. | 500.0 | 5.2 |
| W1 (unseen robots) | robot taxonomy | aerial–human viol. | 150.7 | 1.9 |
| W1 (unseen robots) | dynamic-subtype test | single-constraint failures (total) | 266 | 0 |
| W2 (unseen payload) | payload taxonomy | hazard–human viol. | 360.1 | 4.4 |
| W3 (unseen combination) | defined class | hazard–human viol. | 835.0 | 28.2 |
| W3 (unseen combination) | activation check | pickups into a violation (total) | 217 | 0 |
| W0 (design) | hazard distinction | deliveries / courier (ordinary) | 1.36 | 2.06 |
Table 7. Ablations that remove one element of OntoCBF, each in the scenario that needs the element. Violations are counted in constraint-steps per episode, and single-constraint failures and pickups into a violation are summed over 100 episodes.
| Scenario | No robot tax. | No payload tax. | No defined class | OntoCBF |
| W0 (design) | = | = | = | 22 |
| W1 (unseen robots) | 100 | = | = | 20 |
| W2 (unseen payload) | = | 99 | = | 22 |
| W3 (unseen combination) | 79 | 100 | 100 | 40 |
| W4 () | = | = | = | 70 |
Table 8. Episodes with a violation (%) under single-element ablations where the element is needed, PD controller, 100 seeds. The sign = marks scenarios in which the ablation acts exactly as OntoCBF, seed by seed, and all three ablations have zero single-constraint failures.
| Scenario | PD | MAPPO | No filter | ||||
| Viol. | Inf. % | Deliv. | Viol. | Inf. % | Deliv. | Viol. ep. % | |
| W0 (design) | 6.1 | 1.20 | 1.76 | 3.1 | 0.73 | 1.73 | 100 |
| W1 (unseen robots) | 5.6 | 1.15 | 1.64 | 3.3 | 0.75 | 1.69 | 100 |
| W2 (unseen payload) | 6.1 | 1.20 | 1.76 | 3.1 | 0.73 | 1.73 | 100 |
| W3 (unseen combination) | 16.0 | 1.62 | 1.43 | 9.0 | 0.95 | 1.43 | 100 |
| W4 () | 47.9 | 2.21 | 1.80 | 39.6 | 1.81 | 1.69 | 100 |
Table 12. PD and MAPPO behind OntoCBF, 100 paired seeds per cell. MAPPO pools 3 runs trained in W0 only, so its cells hold episodes, and the intervals in § 6 resample training runs and then seeds. The columns give violating steps per episode (Viol.), infeasible QPs as a percentage of agent-steps with active constraints (Inf.) and deliveries per courier (Deliv.), and the last column gives the percentage of episodes with a violation for MAPPO without the filter.
Overestimated capabilities
A HeavyDrone heads for a goal through a human who walks straight at it, from 91 initial states inside the safe set of its own capability (Figure 3). Under Naive, which gives it the capability of Drone, single-constraint failures occur in 1 % of the encounters at , 13 % at and 100 % at . From on they lead to violations, in 48 % of the encounters at with distances down to 0.75 m against a keep-out of 1.0 m, as part 3 of Theorem 4 predicts. OntoCBF stays free of failures and violations at every gain. Small gains brake early, so an overestimated capability appears as single-constraint failures before it appears as violations. In W1, Naive causes 266 single-constraint failures at and, at , violates safety in 44 % of the episodes against 21 % for OntoCBF (Table 4). Joint validity, decided by a linear program at each of the 78,837 states visited in W0 (Table 11), holds for the design family at 94.4 % of them. HeavyDrones under the capability of Drone break it at 853 of these states, and with their own capability they form a new family, valid at 91.6 % of the states where the design family is.
Figure 3. Stress test of the capability assignment for HeavyDrone (§ 6.2) in head-on encounters with a human, all starting inside the safe set of the true capability. Under Naive the HeavyDrone inherits the capability of Drone, and OntoCBF assigns it its own capability. The left panel shows the encounters with a single-constraint failure and the right panel those with a violation.
| Naive | OntoCBF | McNemar | violating steps | aerial–human | failures (Naive) | Naive | OntoCBF | |
| 1 | 23 | 20 | 0.63 | +0.8 [−2.6, +4.3] | −0.2 [−1.5, +1.0] | 266 | 1.79 | 1.64 |
| 3 | 27 | 21 | 0.26 | +1.9 [−0.6, +4.3] | +3.1 [+0.4, +6.1] | 907 | 2.00 | 1.87 |
| 10 | 44 | 21 | <0.001 | +2.9 [−0.6, +6.3] | +2.3 [−3.1, +7.3] | 2223 | 2.13 | 2.02 |
Table 4. Naive and OntoCBF in W1 (unseen robots) as the class- gain grows, with the PD controller and 100 paired seeds ( is the main setting). Differences carry paired 95 % bootstrap intervals, and comes from McNemar’s exact test on episodes with a violation. OntoCBF has zero single-constraint failures at every gain.
| Configuration (two drones substituted) | States | Jointly valid | All per-agent QPs feasible | Valid where the design team is |
| Drone (design) | 78837 | 94.43 % | 92.44 % | 74442/74442 |
| LightDrone, inherited | 78837 | 94.57 % | 92.56 % | 74442/74442 |
| HeavyDrone, (Naive) | 78837 | 93.34 % | 91.56 % | 73589/74442 |
| HeavyDrone, own | 64845 | 87.11 % | 84.79 % | 56485/61668 |
Table 11. Joint validity (Definition 2) on the states visited by the design team in W0, with both drones substituted.
Conflicts between constraints
Each of the 466 violation onsets in W0, W1, W3 and W4 follows, within 1 s, a conflict of an agent of the violated constraint, consistent with Proposition 6 in continuous time, so every episode with a violation also has a conflict. Conflicts precede only 1.9–3.3 % of all agent-steps within the same window (Table 10). Of all infeasible QPs, 98 % are conflicts, and the rest contain an inequality whose barrier is already negative. Conflicts arise mostly when humans close in on a drone or a hazardous carrier, and the median deepest intrusion is 10–17 % of . At 56–68 % of the onsets the robot was not moving toward the other entity. The filter takes at most 6.1 ms per step and the full reasoning step 4.6 ms per event on one CPU core (Table 9).
| Scenario | Deliveries/courier None | Deliveries/courier OntoCBF | QP infeas. (%) | Filter (ms/step) | Events/ep. | Reasoning (ms/event) |
| W0 (design) | 2.67 | 1.76 | 1.2 | 1.40 | 268 | 0.91 |
| W1 (unseen robots) | 2.83 | 1.64 | 1.1 | 1.42 | 267 | 0.91 |
| W2 (unseen payload) | 2.67 | 1.76 | 1.2 | 1.40 | 268 | 0.91 |
| W3 (unseen combination) | 2.67 | 1.43 | 1.6 | 1.58 | 262 | 0.93 |
| W4 () | 2.67 | 1.80 | 2.2 | 6.10 | 774 | 4.60 |
Table 9. Deliveries per courier in the warehouse without a filter and with OntoCBF, and computational cost of OntoCBF with the PD controller on one CPU core.
| Scenario | Onsets | Onsets after a conflict within 1 s | Onsets after a conflict within 2 s | Base rate (agent-steps) all | Base rate (agent-steps) close encounter |
| W0 (design) | 45 | 45 | 45 | 2.0 % | 3.9 % |
| W1 (unseen robots) | 31 | 31 | 31 | 1.9 % | 3.7 % |
| W3 (unseen combination) | 108 | 108 | 108 | 2.6 % | 4.6 % |
| W4 () | 282 | 282 | 282 | 3.3 % | 6.4 % |
Table 10. Violation onsets of OntoCBF with PD preceded by a conflict of an agent of the violated constraint, and base rates over agent-steps (window 1 s).
Learning in the loop
Behind the MAPPO policy of § 4.4, trained in W0 only, episodes with a violation fall from 20–70 % to 11–65 % (Table 2). The rate of infeasible QPs per agent-step with active constraints, almost all of them conflicts, drops by 18–41 % in every scenario, with paired 95 % bootstrap intervals below zero over runs and seeds and in each of the 3 runs (Table 13). Mean violating steps per episode change by −8.3 to −2.3, significantly only in W3 (−6.98 [−12.47, −1.70]), deliveries per courier by −0.11 to +0.04, and without the filter the policy violates safety in every episode.
| Scenario | Run | infeasible QPs / ep. | violating steps / ep. | deliveries / courier |
| W0 (design) | 1 | |||
| 2 | ||||
| 3 | ||||
| W1 (unseen robots) | 1 | |||
| 2 | ||||
| 3 | ||||
| W3 (unseen combination) | 1 | |||
| 2 | ||||
| 3 | ||||
| W4 () | 1 | |||
| 2 | ||||
| 3 |
Table 13. MAPPO minus PD behind OntoCBF for each training run (W2 equals W0), with paired 95 % bootstrap intervals over the 100 seeds (10,000 resamples). Violating steps fall significantly in 4 of the 12 run–scenario cells.
Conclusion
OntoCBF makes a domain ontology the specification of a multi-robot safety filter, in which entailment removes missing constraints and the dynamic-subtype test, exact for box-bounded double integrators, removes overestimated capabilities. Conflicts between constraints remain possible with pairwise barriers (Proposition 7), and a policy trained in W0 alone makes them less frequent. The guarantees hold in continuous time for double integrators and humans with acceleration bounded by , and they rely on correct class labels and measured profiles, since each entity is constrained only by the axioms of the classes entailed for it (Appendix B). Verified certificates, fail-safe defaults for unknown labels, axiom priorities that favor people, and nonlinear classes are the next steps.
Generative AI Use
Generative AI tools were used for coding and experimental implementation, literature search, and writing and editing. The authors checked all AI-assisted content and take full responsibility for the paper.
Proofs
The constructions in part 2 of Lemma 3, part 3 of Theorem 4 and Proposition 7 show entry into the keep-out used in the barrier, which for the filter is (§ 4.3).
Barrier derivative and barrier inequalities
Let be a constraint with , and . The payload offset is constant, so for every type of constraint, , and . Differentiating (3) gives
so is equivalent to (4). For two robots, adding and with yields . For a robot and a human with , the inequality yields for every admissible acceleration of the human. For a zone, . Hence, whenever the agents of satisfy their barrier inequalities, holds for all admissible human accelerations. The curvature term in (6) is nonnegative, so is sufficient for braking at rate along to stop the approach at a distance of at least , and tangential motion can only increase that distance.
Single-constraint feasibility
Proof [of Lemma 3] Part 1. Let and . In , the first term is nonpositive and the third term is nonpositive by the Cauchy–Schwarz inequality. If , the second term is nonpositive as well. Otherwise gives , so the second term is at most . Hence . The right-hand side of the barrier inequality of agent therefore satisfies for a robot pair, for a human, and for a zone. The input lies in the box and gives . Inside the filter margin the implementation replaces by , which keeps .
Part 2, a state with . Let , pick , and set and . Then , the curvature term vanishes and , so the right-hand side equals for each type of constraint. Every input in the box violates the inequality, because .
Part 2, an unavoidable violation at a zone. Write and , place the zone at the origin, and fix and . The agent starts at with and , so that . Let be any measurable input with values in , and let . Integrating twice,
At these bounds give with , and . If , then and
so the agent is inside the keep-out region at for every admissible input.
Part 2, an unavoidable violation with a human. Let the human start at the origin at rest and accelerate toward the robot with . In relative coordinates , the first component of lies in with , and the second lies in . The barrier uses . The zone argument with , , , upper acceleration bound , lower bound and lateral bound gives and , so for all sufficiently small , whatever the sign of .
For the zone, the conditions suffice, and with , , and m they allow m/s. At m/s the barrier starts at m/s and the largest reachable distance at is at most m. Smallness of is needed in this argument, since at large speeds a robot may swerve around the keep-out region. For a general compact convex input set containing the origin, the argument of part 1 shows that every barrier inequality is feasible on if and only if , the inradius of about the origin. Inclusion implies this inequality for an inherited capability, and joint validity in Theorem 4 needs the inclusion itself, because a joint witness may use a corner of .
Axiom equivalence and certificate inheritance
Let be the set of concept names that occur in safety axioms or in TBox axioms other than inclusions between two concept names (definitions, disjointness axioms, domain and range restrictions). For the TBox of Section B, consists of Entity, Agent, Object, ControlledAgent, UncontrolledAgent, Aerial, Ground, Human, Payload, HazardousPayload, RestrictedZone and HazardCarrier. Axioms that tie a class to a role, such as , put in , so a class that adds constraints for other individuals in this way has subsumers in that differ from those of its parent.
Lemma 8 (Axiom equivalence) Let be the TBox of Section B, and let be classes with the same subsumers in . Let be an ABox of concept-name and role assertions that contains , and let be obtained from by replacing with . Then is consistent if and only if is consistent, and both entail the same instances of every safety axiom.
Proof. Since , every model of is a model of , so entails everything that entails, and consistency of implies consistency of . For the converse, let be a model of , and let be obtained from by adding to the extension of every concept name with . The ABox consists of assertions of concept names and roles, as in (1), and only enlarges extensions, so satisfies . It also satisfies , because either received or lies in , where it is one of its own subsumers and hence a subsumer of . Every safety axiom and every TBox axiom other than an inclusion between two concept names mentions only individuals, roles and concept names in , which keep their interpretation in . An inclusion between concept names can fail in only if was added to , and then , so received as well or , in which case and . Hence is a model of with the same safety-axiom instances as , so is consistent whenever is, and every instance that entails is entailed by .
The proof never uses for the converse direction, so any two classes with the same subsumers in yield the same constraints. For example, Forklift and Carrier are siblings with the same subsumers in , and Forklift has the larger acceleration limit, so by part 2 of Theorem 4, applied with this weaker hypothesis, the same payload offset and the capability of Carrier, it could reuse the certificate of Carrier. OntoCBF searches for a certificate only along the taxonomy, which is conservative.
Proof [of Theorem 4] Part 1. Applying Lemma 8 to the agents of one at a time shows that, for all near and carries relations, the ABoxes under and under entail the same instances of every safety axiom, so is the same. A barrier (3) depends on the classes of its agents only through , and and the weights depend on the class only through , which is the same under and .
Part 2. Fix and let satisfy Definition 2 for . For , , and for the velocity sets coincide, so . The functions , the index set , the set and the value do not depend on the composition by part 1, so satisfies Definition 2 for . For the QPs, write each barrier inequality of agent as , which reads for double integrators. By part 1, the share of agent is the same under and . If is feasible for agent under , its velocity lies in , so some produces the same velocity and satisfies the same inequalities. The payload offset enters the barriers of nearCarried through , which is why the theorem assumes it unchanged.
Part 3. Under , agent has capability and input bound . By Lemma 3, all its barrier inequalities are feasible at every state with and if , which is the condition , and otherwise each type of inequality is infeasible at some state with . The last claim of part 3 is part 2 of Lemma 3.
Feasibility of every QP versus joint validity. If every agent’s QP is feasible at , the applied inputs satisfy Definition 2 at , because the two shares of a robot pair sum to the barrier condition and each human inequality contains the worst human acceleration. The converse fails for two drones with and on the first axis, drone at 0 with velocity and drone at 0.8 at rest, and a human at 1.914 at rest. The robot pair requires and , and the aerial–human constraint of requires , so the QP of is infeasible, and each of its inequalities is feasible on its own. The inputs and satisfy the unsplit pair condition and both human conditions, so Definition 2 holds. On the states visited in W0 this gap separates the 94.4% of states with joint validity from the 92.4% at which every QP is feasible (Table 11).
Extensions of Theorem 4
Classes with different state spaces. Let , , be continuously differentiable maps between state spaces with
and let apply to the states of the agents in and the identity to all other states.
Proposition 9 (Inheritance through embeddings). Let the hypotheses of Theorem 4 hold with replaced by (7). If is jointly valid for on , then is jointly valid for on .
Proof. Let be in that set and , and take a witness for at . For choose with , which exists by (7), and set for . By the chain rule, , the terms of the uncontrolled entities are unchanged, and , so is a witness for at .
Condition (7) is the infinitesimal form of the consistent abstractions of Pappas et al. [27], and it underlies the transfer of CBFs through reduced-order models (Cohen et al., [9]). The barrier constrains the embedded point, so the keep-out of the physical robot must be stated for that point or enlarged accordingly. The set on which (7) holds must also be kept invariant, which belongs to the certificate of the design team.
A unicycle inheriting from a double integrator. A dynamic unicycle has state and dynamics , , , , with , , and . Its look-ahead point satisfies and
Let . The position components of and coincide, and the inclusion (7) of the acceleration components reads , where is the rotation by . The squares cover the disk of radius as varies, so the inclusion holds for all on the envelope , if
These conditions are also necessary for the inclusion at every heading, and on a set of headings the factor can be replaced by . The barrier then constrains , so the keep-out of the robot itself grows by , and the envelope must be kept invariant, for instance through a speed bound , which implies and . The conditions are demanding, since with , , and they need .
Approximate inclusion.
Corollary 10. Let the hypotheses of the Theorem 4 hold with replaced by for . Suppose that the witnesses for satisfy Definition 2 with margin (right-hand side ), and that for every difference between elements of and . If for all , then is jointly valid for .
Proof. For pick with , and let otherwise. For every and the left-hand side of Definition 2 decreases by at most , so it remains at least .
For the barrier (3) and classes with a common drift, the differences only affect accelerations, which enter through , so .
Activation margin from the sensing radius
Lemma 11. Let a constraint be activated at time by a sensed relation, so that the distance of equals at . For nearCarried, the relation near is sensed at the payload, from which is measured. If the closing speed satisfies and , then .
The proof is immediate from (3) with keep-out . If an implementation senses the relation at the position of the carrier, must exceed the bound by the payload offset . For the HeavyDrone and a human, , , m, m and m/s give m, and at m the barrier is at least 0.13 m/s at activation. A Drone or LightDrone carrying a hazardous payload needs 3.82 m, and a Carrier with a hazardous payload needs 4.61 m, or 5.31 m if the relation is sensed at the carrier. These values assume nominal speeds, and a HeavyDrone at 2.0 m/s would need more than 7 m, as would a HeavyDrone carrying a hazardous payload, a combination outside all our scenarios. Sampling at 0.05 s moves the activation inward by at most about 0.12 m, which the slack absorbs. If near is retracted only when the distance exceeds for some and relative speeds are bounded by , two activations of the same pair are at least apart. Together with a positive dwell time of task events, this guarantees finitely many switches on bounded intervals.
Invariance under switching
Proof [of Proposition 5] Let be the times in at which changes, and set . For a constraint write for the keep-out used by the filter and for its distance. We show by induction over that every active on satisfies and for , where solves with . At , a constraint that was active before satisfies by the induction hypothesis and continuity, a newly active one by Assumption 1, and a constraint active at by hypothesis. Fix such a , and let be the supremum of the times such that and on . On the function is continuously differentiable along the trajectory, is absolutely continuous, and the barrier inequalities of the agents of hold for almost every , so almost everywhere by Section A.1. The comparison lemma gives on , hence by continuity. If , then , which is impossible because on . So , , and by continuity . Since every active barrier is positive, every active distance exceeds , and for all .
The proof uses the applied inputs only through the barrier inequalities along the realized trajectory, which the infeasible-QP counter of the experiments checks only at the sampling times. The barrier is not continuously differentiable on a neighborhood of the closure of its zero-superlevel set, because its gradient grows without bound as , so the standard invariance theorem for CBFs does not apply verbatim and the strict margin of Assumption 1 is needed. From , radial braking at exactly rate keeps and reaches in finite time. The rule that requires the relative acceleration inside the margin makes the inequality discontinuous at , which is harmless because, by the proof, the distance stays above while the barrier stays positive.
Loss of safety starts with a conflict
Proof [of Proposition 6] Let be the infimum of the times at which some has or , which is finite by assumption. Since changes finitely often, it is constant on for some , and some constraint of this set has or along a sequence . Let be the time from which is active without interruption. If , Assumption 1 and continuity give and near , which contradicts the choice of , so . By the definition of , and on . By continuity , because either itself turns negative after , or with , and then . Suppose that every agent of had a feasible QP for almost every . These agents then applied inputs that satisfy their barrier inequalities, including those of , for almost every , and the argument in the proof of Proposition 5, applied to alone on , gives , a contradiction. Hence some agent of has an infeasible QP on a set of positive measure. At every all active barriers are nonnegative because , so by part 1 of Lemma 3 and each barrier inequality of is feasible on its own, and the QP of is infeasible only because its inequalities exclude each other.
Unavoidable conflicts among three robots
Proof [of Proposition 7] Name the robots , and , fix and , and let . Robot starts at the origin at rest, at with velocity , and at with velocity . For the pairs and we have and , so . For the pair , whenever . Let the three robots apply arbitrary measurable inputs with values in , write , and let . Integrating twice gives , , and for each robot, so
and for all robots and . If or for some , the two robots concerned are at distance at most at , which is less than for . Otherwise on , the gaps and at are positive and sum to at most , and the smaller gap is at most . The corresponding pair is then at distance at most , which is less than when . In both cases two robots are closer than at some time in .
The construction holds for every gain and every filter, because it bounds the reachable positions of all three robots directly. At the boundary state with and the same positions, the two barrier conditions read and , which force and , so Definition 2 fails there even for one input of the whole team. With different bounds and capabilities, the same argument with approach speeds in the ratio and applies whenever , which always holds for .
Chains of robots. Let robots with bound and capability start on the first axis at with velocities , so that every adjacent gap closes at speed , and let with . Every adjacent barrier, with , equals , and the other barriers are positive for small . The inputs of the inner robots do not affect , and braking the two outer robots as hard as possible gives, at ,
which is less than because . The crossing and lateral bounds of the three-robot case then give two robots closer than for small . At the state with the same positions, and every adjacent barrier zero, all other barriers are nonnegative, and the adjacent barrier conditions add up to , which no inputs in achieve, so Definition 2 fails there. Hence, for every , a chain of robots contains states with positive barriers from which a collision is unavoidable, and no fixed reduction of the capabilities makes pairwise braking barriers jointly valid on their whole safe set for all team sizes.
Ontology and reasoning
Besides the taxonomy of Table 3, the TBox contains ControlledAgent Agent, UncontrolledAgent Agent, Aerial, Ground ControlledAgent, Payload, RestrictedZone Object, Agent, Object Entity, the disjointness axioms for Agent Object, ControlledAgent UncontrolledAgent, Aerial Ground and Payload RestrictedZone, the symmetric role near on Entity, the role carries from ControlledAgent to Payload, the role nearCarried from ControlledAgent to Entity, the chain carries near nearCarried, and the definition HazardCarrier ControlledAgent carries.HazardousPayload. The safety axioms are near(ControlledAgent, ControlledAgent) with 0.3 m, near(ControlledAgent, Human) with 0.5 m, near(Aerial, Human) with 1.0 m, near(Ground, RestrictedZone) with 1.0 m from the zone center, nearCarried(ControlledAgent, Human) with 0.8 m, and nearCarried(HazardCarrier, Human) with 2.0 m, the last two measured from the payload.
| Class | Parent | Role in the experiments | ||
| Drone | Aerial | design class | 2.0 | 1.5 |
| LightDrone | Drone | test only, Drone, inherits | 3.0 | 1.5 |
| HeavyDrone | Drone | test only, Drone, own capability | 1.0 | 1.5 |
| Rover | Ground | design class | 1.2 | 1.0 |
| Carrier | Ground | design class | 1.0 | 0.8 |
| Forklift | Ground | test only, no design ancestor | 1.2 | 1.0 |
| Human | UncontrolledAgent | uncontrolled | 0.5 (disk) | 0.8 |
| Payload | Object | design class | – | – |
| HazardousPayload | Payload | design class | – | – |
| Chemical | HazardousPayload | test only | – | – |
Table 3. Class profiles with the per-axis acceleration limit in m/s and the nominal speed of the task controller in m/s. Robot–human constraints need , which holds for every robot class.
The OWL file used in all experiments is released with the code.
Completeness of the compiled reasoner. Apart from the definition of HazardCarrier, every TBox axiom above lies in OWL 2 RL ([24]), namely the inclusions and disjointness axioms between concept names, the domain and range restrictions, the symmetric role and the property chain. The definition contributes the inclusions HazardCarrier ControlledAgent and ControlledAgent HazardCarrier, which are RL rules, and the inclusion HazardCarrier , whose existential right-hand side lies outside OWL 2 RL, so we argue directly. Let be the least fixpoint of the ABox under the rules for atomic inclusions, and , symmetry of near, the chain , the two RL inclusions of the definition, and for each disjointness axiom. If , extending by one fresh HazardousPayload carried by every HazardCarrier that carries none gives a model of , because the fresh payload has no near relation and its classes HazardousPayload, Payload, Object and Entity cause no clash. Every rule is sound, so is inconsistent if and only if , and otherwise it entails a class or role assertion about named individuals if and only if the assertion is in . The argument relies on carries having no functionality or cardinality restriction. The compiled reasoner computes , and its agreement with HermiT on 72 random ABoxes with 5 to 9 individuals, some of them inconsistent, checks the implementation.
Inconsistency. If is inconsistent, it has no model and entails every assertion, in particular near and ControlledAgent for every agent . The instance of the robot–robot axiom is then active and requires , which every state violates, so . In a consistent ABox every active constraint relates two distinct entities, because (1) asserts near only between distinct entities and the self-pair nearCarried, derived from a carried payload near its carrier, matches only axioms whose second class is Human, which is disjoint from ControlledAgent. The compiled consistency check tests whether , so it also finds clashes that the taxonomy propagates, such as the one below. In each of 100 W0 episodes, with seeds disjoint from the evaluation seeds, a faulty transponder makes the worker closest to a robot also report the class Rover for 3 s, so that one individual is typed both Human and Rover, which the TBox makes disjoint through ControlledAgent UncontrolledAgent . A filter that trusts the label treats the worker as a robot, with a 0.3 m keep-out and no worst-case human acceleration, and within the fault and the 2 s after it 60 episodes violate safety (1995 violating steps and 5 single-constraint failures). The compiled check detects 100 of the 100 faults at the event, and HermiT agrees on 10/10 sampled events. OntoCBF then keeps the last consistent ABox and brakes the team until the contradiction disappears, after which 4 episodes violate safety (78 steps), at the cost of 9.90 against 10.68 deliveries per episode on the same seeds without a fault.
Labels and defaults. Entailment is monotone, so an entity whose class label is missing or too general receives only the constraints of the classes entailed for it, and an entity typed only as Entity receives none. Class labels and dynamics profiles are therefore safety-relevant inputs, and the profiles must be measured under the loads a robot carries. A fail-safe default, such as typing every sensed entity of unknown class as Human, gives such an entity the largest keep-out distances among uncontrolled entities and the worst-case human acceleration. This default belongs to the layer that builds the ABox, because a description logic draws conclusions only from the axioms and assertions that are present.
Experimental details and additional results
Warehouse layout. Each 12 × 12 m tile has six lanes from a pickup station at m to a drop station at m (Figure 2). The two middle lanes belong to the drones and cross a restricted zone at the center of the tile, which binds only Ground agents, and the four outer lanes belong to the carriers. A second zone lies west of the pickup column. The payload of a drone hangs below it, and the payload of a carrier sits 0.7 m from it along , toward the drop station. After a drop, the next item waits at the pickup station. Two of the four humans per tile are station workers who walk between the pickup and drop areas of their tile, and the other humans and the rovers move between random points of the whole floor. W4 tiles this layout 2 × 2 on a 24 × 24 m floor.

Figure 2. Scenario W3 (a LightDrone carries a Chemical and a Drone a HazardousPayload), same seed under OntoCBF ( s) and under Flat at its first hazard violation ( s). For a drone that carries a hazardous payload, Flat falls back to its default payload keep-out and keeps the payload only 0.8 m from people.
Gain selection. On 100 W0 validation episodes, with seeds disjoint from the evaluation seeds, OntoCBF with the PD controller violates safety in 14, 15, 17, 17, 24 % of the episodes for , with 1.76 to 2.02 deliveries per courier. The selection rule was fixed before evaluation and takes the largest whose violation rate is at most the rate at . Since the rate grows with , the rule selects , the most conservative gain of the grid, which is also the gain at which an overestimated capability is least visible (Figure 3 and Table 4).
Stress test of the capability assignment (§ 6.2). A human walks straight at a robot whose goal lies behind the human, with initial distances of 2 to 6 m, robot speeds of 0.5 to 1.5 m/s and human speeds of 0 to 0.8 m/s. Only initial states in the safe set of the aerial–human barrier with the true capability of the class are kept (91 encounters per setting for HeavyDrone), so with this capability every single constraint stays feasible by Lemma 3. We sweep the gain of , which is 1 in the warehouse (Figure 3). OntoCBF avoids single-constraint failures and violations at all gains, with a minimum distance of 1.12 m, and for Drone, whose capability is the same under both methods, Naive and OntoCBF coincide. At , single constraints fail in 1 % of the encounters under Naive, and the HeavyDrone stays outside the keep-out region in all of them, so in the warehouse an overestimated capability shows up mainly as single-constraint failures.
The dynamic-subtype test in the warehouse at larger gains. Table 4 repeats W1, the only scenario with a class that fails the test, for Naive and OntoCBF at larger , and its row for reproduces the main results seed by seed. OntoCBF violates safety in 20–21 % of the episodes at every gain. Naive violates safety in 23 % of the episodes at (McNemar against OntoCBF on the same seeds), in 27 % at , where its aerial–human violations are significantly more frequent, and in 44 % at (). Its overestimated capability lets the HeavyDrone brake later, which also yields slightly more deliveries.
Single-element ablations. Each ablation changes exactly one element of OntoCBF (Table 7), and Table 8 shows in which scenarios the three ablations below fail.
No robot tax. types every robot outside the design classes as ControlledAgent only, so all its ancestors lie above the design classes and it is assigned its own profile. No payload tax. types every payload class outside the design classes as Payload. No defined class skips the realization of HazardCarrier and writes the hazard keep-out only for the design pair (Carrier, HazardousPayload), closed under both taxonomies.
Specification effort and throughput. W3 alone accounts for 3 of the 10 ontology statements of OntoCBF and 5 of the 18 table entries of Flat that cover W1–W3. Figure 4 places each method by deliveries per courier and episodes with a violation.
Figure 4. Episodes with a violation against deliveries per courier, 100 seeds per point and 3 pooled training runs for MAPPO. Points further down and to the right are better.
| Scenario | Method | hazard–human 2.0 m | payload–human 0.8 m | aerial–human 1.0 m | robot–human 0.5 m | ground–zone 1.0 m | robot–robot 0.3 m | Deliveries/courier ordinary | Deliveries/courier hazardous | Pickups into viol. |
| W0 (design) | None | 423.0 | 252.2 | 344.3 | 250.9 | 71.9 | 63.7 | 3.00 | 2.00 | 348/2200 |
| W0 (design) | Flat | 5.4 | 1.2 | 2.3 | 0.3 | 1.2 | 0.0 | 2.05 | 1.16 | 0/1373 |
| W0 (design) | Flat-cons | 5.4 | 1.2 | 2.3 | 0.3 | 1.2 | 0.0 | 2.05 | 1.16 | 0/1373 |
| W0 (design) | No-chain | 421.5 | 78.6 | 3.2 | 0.3 | 1.1 | 0.0 | 2.12 | 1.65 | 126/1500 |
| W0 (design) | Naive | 4.4 | 0.8 | 2.2 | 0.2 | 1.3 | 0.0 | 2.06 | 1.16 | 0/1373 |
| W0 (design) | Untyped | 5.7 | 0.1 | 0.6 | 0.1 | 1.0 | 0.0 | 1.36 | 1.18 | 0/1136 |
| W0 (design) | No-act. | 20.2 | 1.5 | 2.2 | 0.2 | 1.1 | 0.0 | 2.06 | 1.18 | 99/1391 |
| W0 (design) | OntoCBF | 4.4 | 0.8 | 2.2 | 0.2 | 1.3 | 0.0 | 2.06 | 1.16 | 0/1373 |
| W0 (design) | MAPPO | 0.9 | 0.9 | 1.1 | 0.6 | 1.4 | 0.1 | 1.95 | 1.28 | 0/4042 |
| W1 (unseen robots) | None | 423.0 | 247.2 | 336.2 | 244.9 | 71.9 | 64.9 | 3.25 | 2.00 | 326/2100 |
| W1 (unseen robots) | Flat | 3.3 | 2.4 | 76.2 | 1.4 | 1.4 | 0.0 | 1.95 | 1.17 | 0/1343 |
| W1 (unseen robots) | Flat-cons | 4.8 | 0.3 | 0.8 | 0.5 | 0.6 | 0.0 | 1.00 | 1.18 | 0/996 |
| W1 (unseen robots) | No-chain | 412.9 | 103.9 | 1.9 | 0.6 | 0.9 | 0.0 | 1.92 | 1.66 | 139/1433 |
| W1 (unseen robots) | Naive | 4.1 | 0.5 | 1.8 | 0.8 | 2.2 | 0.0 | 2.10 | 1.17 | 0/1402 |
| W1 (unseen robots) | Untyped | 5.0 | 0.0 | 0.7 | 0.5 | 1.0 | 0.0 | 1.23 | 1.19 | 0/1096 |
| W1 (unseen robots) | No-act. | 20.8 | 1.8 | 1.9 | 0.2 | 0.6 | 0.0 | 1.87 | 1.20 | 102/1331 |
| W1 (unseen robots) | OntoCBF | 3.1 | 0.5 | 1.9 | 0.2 | 1.0 | 0.0 | 1.87 | 1.19 | 0/1315 |
| W1 (unseen robots) | MAPPO | 0.9 | 0.7 | 2.2 | 0.7 | 1.0 | 0.0 | 1.89 | 1.29 | 0/3976 |
| W2 (unseen payload) | None | 423.0 | 252.2 | 344.3 | 250.9 | 71.9 | 63.7 | 3.00 | 2.00 | 348/2200 |
| W2 (unseen payload) | Flat | 359.7 | 2.0 | 2.2 | 0.0 | 0.9 | 0.0 | 2.06 | 1.56 | 115/1451 |
| W2 (unseen payload) | Flat-cons | 5.4 | 1.2 | 2.3 | 0.3 | 1.2 | 0.0 | 2.05 | 1.16 | 0/1373 |
| W2 (unseen payload) | No-chain | 421.5 | 78.6 | 3.2 | 0.3 | 1.1 | 0.0 | 2.12 | 1.65 | 126/1500 |
| W2 (unseen payload) | Naive | 4.4 | 0.8 | 2.2 | 0.2 | 1.3 | 0.0 | 2.06 | 1.16 | 0/1373 |
| W2 (unseen payload) | Untyped | 5.7 | 0.1 | 0.6 | 0.1 | 1.0 | 0.0 | 1.36 | 1.18 | 0/1136 |
| W2 (unseen payload) | No-act. | 20.2 | 1.5 | 2.2 | 0.2 | 1.1 | 0.0 | 2.06 | 1.18 | 99/1391 |
| W2 (unseen payload) | OntoCBF | 4.4 | 0.8 | 2.2 | 0.2 | 1.3 | 0.0 | 2.06 | 1.16 | 0/1373 |
| W2 (unseen payload) | MAPPO | 0.9 | 0.9 | 1.1 | 0.6 | 1.4 | 0.1 | 1.95 | 1.28 | 0/4042 |
| W3 (unseen combination) | None | 1052.7 | 255.4 | 350.1 | 252.3 | 71.9 | 62.2 | 2.00 | 3.00 | 674/2200 |
| W3 (unseen combination) | Flat | 840.6 | 0.5 | 41.8 | 0.2 | 0.9 | 0.0 | 1.62 | 1.71 | 156/1342 |
| W3 (unseen combination) | Flat-cons | 17.8 | 0.5 | 0.4 | 0.3 | 1.3 | 0.0 | 1.65 | 1.12 | 0/1112 |
| W3 (unseen combination) | No-chain | 1237.7 | 78.6 | 3.2 | 0.3 | 1.1 | 0.0 | 2.12 | 1.65 | 126/1500 |
| W4 (N=32) | None | 1561.2 | 874.6 | 111.5 | 84.6 | 239.2 | 191.0 | 3.00 | 2.00 | 1244/8000 |
| W4 (N=32) | Flat | 477.8 | 6.1 | 8.3 | 3.8 | 10.8 | 9.1 | 2.12 | 1.16 | 0/5667 |
| W4 (N=32) | Flat-cons | 47.8 | 97.9 | 9.2 | 3.8 | 2.3 | 0.0 | 2.12 | 1.66 | 0/5667 |
| W4 (N=32) | No-chain | 1588.0 | 6.5 | 5.4 | 3.6 | 10.8 | 0.0 | 2.13 | 1.14 | 0/5662 |
| W4 (N=32) | Naive | 55.5 | 4.0 | 9.2 | 4.3 | 13.7 | 0.3 | 2.48 | 1.16 | 0/4768 |
| W4 (N=32) | Untyped | 122.2 | 10.1 | 9.6 | 4.0 | 13.1 | 0.0 | 2.13 | 1.16 | 0/5666 |
| W4 (N=32) | No-act. | 255.5 | 6.5 | 7.1 | 4.0 | 10.8 | 0.0 | 2.13 | 1.60 | 0/5666 |
| W4 (N=32) | OntoCBF | 44.1 | 5.1 | 7.1 | 4.0 | 7.8 | 0.1 | 1.98 | 1.10 | 0/5980 |
Table 6. Violations by axiom in the warehouse (constraint-steps per episode), deliveries per courier for ordinary and hazardous payloads, and pickups that began inside a keep-out region (out of all pickups). The MAPPO entries pool 3 100 episodes.
Events before a violation. For every violation onset of OntoCBF with PD in W0, W1, W3 and W4 (W2 replays W0), we identify the last infeasible QP of an agent of the violated constraint. An infeasible QP is a conflict if each of its inequalities is feasible on its own in the true input set. As a base rate, we give the share of agent-steps whose agent had a conflict within the window, over all agent-steps and over those in a close encounter, meaning an active constraint within 1 m of its keep-out. Conflicts make up 98 % of all infeasible QPs, and in the others every inequality that is infeasible on its own has an already negative barrier, as Proposition 6 predicts. We call an onset robot-attributable if, at the onset, a robot of the violated constraint has a positive velocity component toward the other entity. § 6.3 reports the complementary share, the onsets at which the robot was not moving toward the other entity.
Joint validity on visited states. On every step of the 100 W0 evaluation episodes under OntoCBF with PD, we take the active family and ask whether one input for the whole team, inside every agent’s true input set, satisfies all barrier conditions at once, with robot pairs unsplit and worst-case human accelerations. This is a linear feasibility problem, solved with HiGHS. We count only states at which every active barrier of the configuration has outside the margin. With LightDrones the inequalities are those of the design team and the input sets are larger, so validity can only grow by Theorem 4, and this configuration tests the implementation. With HeavyDrones under the capability of Drone, the inequalities are the same and the input sets are smaller. HeavyDrones with their own capability form a different family, which is valid at 91.6 % of the visited states at which the design team’s family is valid. Feasibility of every agent’s QP is lower than joint validity throughout, because the responsibility split is more restrictive than a joint input.
Training details
All values below are read from the training configuration of the reported runs. MAPPO runs 200 iterations of 40 episodes with Adam (learning rate 0.001), discount 0.99, GAE parameter 0.95, clip ratio 0.2, 5 epochs per iteration, minibatches of 16384 agent-steps and entropy coefficient 0.001. The reward weights are 1 for progress, 1 for arrival, 0.01 for intervention, 0.5 for infeasible QPs, 0.2 for violations and 0.5 for overspeed. Training episodes use the W0 layout (design robot classes, Payload and HazardousPayload) with random positions along the lanes and random initial carrying states, 400 steps each, and evaluation episodes have 800 steps. The arrival bonus is paid when a goal is reached and switched, so waiting at a pickup deferred by the activation check earns nothing. Validation uses W0 episodes with seeds disjoint from the evaluation seeds. The own features of an agent are its goal offset, its velocity, its payload offset, the keep-out its payload currently requires and its class embedding. Each entry for another entity holds the relative position and velocity of the entity, the distance, the keep-out , the margin between them and the class embedding of the entity, and an agent that carries a payload also sees chain-related entities measured from its payload. The encoder uses attention pooling with 4 heads of total width 64 and a null token, and entry weights fade linearly to zero over the last 1 m before . Class embeddings have dimension 8 and are tied along the taxonomy (, with for classes absent from training), and the actor and critic heads are two-layer tanh MLPs of width 256. Three training seeds are run, so the between-run part of the intervals in § 6 is coarse, and Table 13 gives each run separately. The PD task controller commands , with pointing at the goal at speed , and humans follow waypoints with the same law and . Two reward configurations were trained, the weights above and a larger arrival bonus of 3 with intervention weight 0.003. A rule fixed beforehand admits the configurations whose violating steps on W0 validation episodes are at most those of PD and takes the one with the most deliveries, which selected the configuration above. The residual scale and its cap, used in all results reported here, were introduced in an earlier
version of these experiments with other scenarios, in which the corrections of a policy trained on Drone, scaled up by the larger of LightDrone, cost goals.
References
- [1]Aaron D. Ames, Xiangru Xu, Jessy W. Grizzle, and Paulo Tabuada. Control barrier function based quadratic programs for safety critical systems. IEEE Transactions on Automatic Control, 62(8): 3861–3876, 2017. doi: 10.1109/TAC.2016.2638961.DOI
- [2]Franz Baader, Diego Calvanese, Deborah L. McGuinness, Daniele Nardi, and Peter F. Patel-Schneider, editors. The Description Logic Handbook: Theory, Implementation, and Applications. Cambridge University Press, 2003.DOI
- [3]Nikolaos Bousias, Charalampia Stamouli, Anastasios Tsiamis, and George J. Pappas. On transferring safety certificates across dynamical systems. arXiv:2602.03987, 2026.arxiv.org/abs/2602.03987
- [4]Joseph Breeden and Dimitra Panagou. Compositions of multiple control barrier functions under input constraints. In American Control Conference (ACC), pages 3688–3695, 2023. doi: 10.23919/ACC55779.2023.10156625.
- [5]Joseph Breeden, Kunal Garg, and Dimitra Panagou. Control barrier functions in sampled-data systems. IEEE Control Systems Letters, 6:367–372, 2022. doi: 10.1109/LCSYS.2021.3076127.DOI
- [6]Lukas Brunke, Yanni Zhang, Ralf Römer, Jack Naimer, Nikola Staykov, Siqi Zhou, and Angela P. Schoellig. Semantically safe robot manipulation: From semantic scene understanding to motion safeguards. IEEE Robotics and Automation Letters, 10(5):4810–4817, 2025. doi: 10.1109/LRA.2025.3553046.DOI
- [7]Zhiyuan Cai, Huanhui Cao, Wenjie Lu, Lin Zhang, and Hao Xiong. Safe multi-agent reinforcement learning through decentralized multiple control barrier functions. arXiv:2103.12553, 2021.arxiv.org/abs/2103.12553
- [8]Satchit Chatterji and Erman Acar. Analyzing probabilistic logic shields for multi-agent reinforcement learning. In European Conference on Artificial Intelligence (ECAI), volume 413 of Frontiers in Artificial Intelligence and Applications, pages 2538–2545. IOS Press, 2025. doi: 10.3233/FAIA251103.
- [9]Max H. Cohen, Tomas G. Molnar, and Aaron D. Ames. Safety-critical control for autonomous systems: Control barrier functions via reduced-order models. Annual Reviews in Control, 57: 100947, 2024. doi: 10.1016/j.arcontrol.2024.100947.DOI
- [10]Mandeep Dhanda, Stephanie Hall, Haoruo Zhao, Elies Dekoninck, and Vimal Dhokia. Ontology-based framework for modelling safety in human-robot collaboration systems. International Journal of Computer Integrated Manufacturing, pages 1–22, 2026. doi: 10.1080/0951192X.2026.2711052.DOI
- [11]Ingy ElSayed-Aly, Suda Bharadwaj, Christopher Amato, Rüdiger Ehlers, Ufuk Topcu, and Lu Feng. Safe multi-agent reinforcement learning via shielding. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 483–491, 2021. doi: 10.65109/KRJR7599.
- [12]Maham Fatima and Boris Tolg. A triage ontology-guided knowledge graph for constraining reinforcement learning decisions in mass casualty incidents. In Third Workshop on Knowledge Graphs and Neurosymbolic AI (KG-NeSy 2026), ISWC 2026 Workshops Joint Proceedings, volume 4313 of CEUR Workshop Proceedings, 2026.
- [13]Birte Glimm, Ian Horrocks, Boris Motik, Giorgos Stoilos, and Zhe Wang. HermiT: An OWL 2 reasoner. Journal of Automated Reasoning, 53(3):245–269, 2014. doi: 10.1007/s10817-014-9305-1.DOI
- [14]Shangding Gu, Jakub Grudzien Kuba, Yuanpei Chen, Yali Du, Long Yang, Alois Knoll, and Yaodong Yang. Safe multi-agent reinforcement learning for multi-robot control. Artificial Intelligence, 319:103905, 2023. doi: 10.1016/j.artint.2023.103905.DOI
- [15]IEEE. IEEE standard ontologies for robotics and automation. IEEE Std 1872-2015, IEEE Standards Association, New York, NY, USA, 2015.
- [16]Axton Isaly, Masoumeh Ghanbarpour, Ricardo G. Sanfelice, and Warren E. Dixon. On the feasibility and continuity of feedback controllers defined by multiple control barrier functions. IEEE Transactions on Automatic Control, 69(11):7326–7339, 2024. doi: 10.1109/TAC.2024.3383069.DOI
- [18]ISO. Industrial trucks – safety requirements and verification – part 4: Driverless industrial trucks and their systems. ISO 3691-4:2023, International Organization for Standardization, Geneva, Switzerland, 2023.
- [19]Joonkyung Kim, Janghyun Cho, Jeehyun Lee, Changjoo Nam, and Yiwei Lyu. Capability-conditioned graph control barrier functions for zero-shot safe navigation of heterogeneous mobile robot teams. arXiv:2610.41214, 2026a.arxiv.org/abs/2610.41214
- [20]Joonkyung Kim, Yanze Zhang, Wenhao Luo, and Yiwei Lyu. Capability-aware heterogeneous control barrier functions for decentralized multi-robot safe navigation. arXiv:2604.13245, 2026b.arxiv.org/abs/2604.13245
- [21]Lars Lindemann and Dimos V. Dimarogonas. Control barrier functions for signal temporal logic tasks. IEEE Control Systems Letters, 3(1):96–101, 2019. doi: 10.1109/LC SYS.2018.2853182.DOI
- [22]Barbara H. Liskov and Jeannette M. Wing. A behavioral notion of subtyping. ACM Transactions on Programming Languages and Systems, 16(6):1811–1841, 1994. doi: 10.1145/197320.197383.DOI
- [23]Tamas G. Molnar and Aaron D. Ames. Composing control barrier functions for complex safety specifications. IEEE Control Systems Letters, 7:3615–3620, 2023. doi: 10.1109/LCSYS.2023.3339719.DOI
- [24]Boris Motik, Bernardo Cuenca Grau, Ian Horrocks, Zhe Wu, Achille Fokoue, and Carsten Lutz. OWL 2 web ontology language profiles (second edition). W3C Recommendation, 2012.
- [25]Shima Sadat Mousavi, Xiao Tan, and Aaron D. Ames. From vertices to convex hulls: Certifying set-wise compatibility for CBF constraints. IEEE Control Systems Letters, 9:3011–3016, 2025. doi: 10.1109/LCSYS.2025.3648436.DOI
- [26]Alireza Nadali, Ashutosh Trivedi, and Majid Zamani. Transfer of safety controllers through learning deep inverse dynamics model. IFAC-PapersOnLine, 58(11):129–134, 2024. doi: 10.1016/j.ifacol.2024.07.436.DOI
- [27]George J. Pappas, Gerardo Lafferriere, and Shankar Sastry. Hierarchically consistent control systems. IEEE Transactions on Automatic Control, 45(6):1144–1160, 2000. doi: 10.1109/9.863598.DOI
- [28]Zachary Ravichandran, David Snyder, Alexander Robey, Hamed Hassani, Vijay Kumar, and George J. Pappas. Contextual safety reasoning and grounding for open-world robots. arXiv:2602.19983, 2026.arxiv.org/abs/2602.19983
- [29]Gunther Reissig, Alexander Weber, and Matthias Rungger. Feedback refinement relations for the synthesis of symbolic controllers. IEEE Transactions on Automatic Control, 62(4):1781–1796, 2017. doi: 10.1109/TAC.2016.2593947.DOI
- [30]Chandan Kumar Sah and Jishnu Keshavan. Exact feasibility certification and optimal responsibility allocation for multi-robot CBF safety filters. arXiv:2609.14935, 2026.
- [31]Xiao Tan and Dimos V. Dimarogonas. Compatibility checking of multiple control barrier functions. In IEEE Conference on Decision and Control (CDC), pages 939–944, 2022. doi: 10.1109/CDC51059.2022.9993001.DOI
- [32]Chris van der Ploeg, Michiel Braat, Beatrice Masini, N. J. Brouwer, and J. P. Paardekooper. Connecting the dots: Context-driven motion planning using symbolic reasoning. In IEEE Intelligent Vehicles Symposium (IV), pages 1–8, 2023. doi: 10.1109/IV55152.2023.10186794.DOI
- [33]Li Wang, Aaron D. Ames, and Magnus Egerstedt. Safety barrier certificates for heterogeneous multi-robot systems. In American Control Conference (ACC), pages 5213–5218, 2016. doi: 10.1109/ACC.2016.7526486.DOI
- [34]Li Wang, Aaron D. Ames, and Magnus Egerstedt. Safety barrier certificates for collisions-free multirobot systems. IEEE Transactions on Robotics, 33(3):661–674, 2017. doi: 10.1109/TRO.2017.2659727.DOI
- [35]Feiyu Wu, Xu Zheng, Yue Qu, Zhuocheng Wang, Zicheng Feng, and Hui Li. Grounding generative planners in verifiable logic: A hybrid architecture for trustworthy embodied AI. In International Conference on Learning Representations (ICLR), 2026.arxiv.org/abs/2602.08373
- [36]Wenli Xiao, Yiwei Lyu, and John M. Dolan. Model-based dynamic shielding for safe and efficient multi-agent reinforcement learning. In International Conference on Autonomous Agents and Multiagent Systems (AAMAS), pages 1587–1596, 2023.arxiv.org/abs/2304.06281
- [37]Zhibo Yang, Junwei Liu, Zhuohan Li, and Wei Zhang. hGCBF: Heterogeneous graph control barrier functions for safe multi-agent systems. IEEE Access, 14:43463–43471, 2026. doi: 10.1109/ACCESS.2026.3670495.

