Contact-Persistent Full Actuation for Aerial Physical Interaction

What the math gives to ML

This paper provides a constructive geometric certificate for whether a constrained actuator can continue stabilizing a system after part of its wrench budget has been consumed by a persistent task. The transferable asset is the feasible-polytope interior margin: instead of checking only rank or unconstrained controllability, a policy can be trained or filtered to keep its requested action away from actuator-feasibility boundaries. The most direct neural-network use is a safety layer for an aerial-control or safe-RL policy that projects task commands into the feasible wrench polytope and penalizes low residual authority, with the margin computed differentiably from a half-space representation or a small allocation linear program.