A Temporal Barrier Framework for Collision Avoidance in Multi-Agent Autonomous Aerial Vehicles

What the math gives to ML

The paper offers a constructive temporal control-barrier mechanism: replace geometric separation with adversarial time-to-collision (aTTC), the minimum time in which an admissible adversary could drive two agents into a collision set. This produces an anticipatory safety variable whose zero crossing is an interpretable intervention boundary, and a differentiable neural surrogate enables real-time use inside a quadratic-program safety filter. The strongest transfer is to learned multi-agent policies, world models, and neural controllers: train or deploy a policy normally, then project its action through a differentiable temporal barrier QP using aTTC estimated from the learned dynamics. Unlike distance penalties, the method predicts a measurable safety transition at aTTC equal to the chosen time margin and can be tested through intervention frequency, QP feasibility, and collision-rate scaling.