Researchers posted an arXiv preprint proposing OntoCBF, a method that builds safety constraints for mixed robot teams from a knowledge base of robot classes.
The authors say safety filters for teams of different robot types often fail, which matters where robots with different payloads and braking limits share space, such as the simulated warehouse used in the paper. The preprint, titled "Ontology-Compiled Control Barrier Functions for Heterogeneous Robot Teams," is listed under robotics on arXiv1.
According to the abstract, existing filters fail because of missing constraints, overestimated braking capabilities and conflicts between constraints that can each be satisfied alone. A control barrier function is a mathematical condition that keeps a system inside a set of safe states. A quadratic program is the optimization problem solved at each step to pick the control command closest to the intended one that still meets the safety condition.
OntoCBF compiles an ontology, a structured knowledge base of robot classes, relations, safety rules and a dynamics profile per class, into a filter with one quadratic program per robot. The authors say reasoning over the ontology supplies a team's constraints online, so existing safety rules can cover new robot classes, payload classes and combinations.
The authors say they prove that a subclass may reuse its parent's braking capability whenever it can produce every velocity of the parent. They describe this as a property separate from subsumption, the formal rule that a subclass inherits its parent's properties. The abstract adds that once three robots interact, the safe set built from pairwise barriers contains states from which a collision is unavoidable. According to the authors, such states persist in larger teams under any fixed capability reduction. These are the authors' proof claims in a preprint that has not been peer-reviewed.
Two simulation results are reported, and the abstract does not present them as a like-for-like comparison. In a simulated warehouse where humans ignore the robots, a filter written for the design classes violated safety in 89-100% of episodes1 involving classes outside that design set. Separately, across all scenarios, including one with 32 robots, OntoCBF violated safety in 20-70% of episodes, each time after a conflict. The median deepest intrusion was 10-17% of the keep-out distance, the separation robots are meant to keep. The abstract does not give the number of episodes.
The authors also report that a multi-agent PPO policy, a reinforcement-learning controller, trained with the filter in the loop lowered the rate of infeasible quadratic programs by 18-41%. The abstract says nearly all of those cases were conflicts. It names no baseline for this figure and does not say whether the reduction is relative or absolute.
The abstract reports results only from a simulated warehouse. The full paper, any hardware testing and replication were not reviewed for this report. The retrieved page does not include author names or a submission date. Its submission history was blank, and its arXiv-issued DOI was listed as pending registration. The listing identifies this version as v1.