Learning When to Act: Communication-Efficient Reinforcement Learning via Run-Time Assurance

Safe reinforcement learning (RL) typically asks what an agent should do. We ask when it needs to act, and show that a single policy can jointly learn control inputs and communication-efficient timing decisions under a pointwise Lyapunov safety shield. We scope the framework to stabilization around a known equilibrium, where CARE-based LQR backups, Lyapunov certificates, and classical Lyapunov- STC are well defined, enabling a clean comparison against the analytical baseline. A run-time assurance (RTA) layer overrides the policy pointwise via a one-step-ahead Lyapunov prediction and a precomputed LQR backup, providing a strictly stronger guarantee than constrained MDP methods that enforce safety only in expectation. On an inverted pendulum, cart–pole, and planar quadrotor, the learned policy achieves 1.91×, 1.45×, and 3.51× higher mean inter-sample interval (MSI) than a classical Lyapunov-triggered baseline; a fixed LQR controller at the same average rate is unstable on all three plants, showing that adaptive timing, not a lower average rate, is what makes sparsity safe. A CARE-derived Lyapunov reward transfers across environments without redesign, with a single weight wc controlling the stability–communication tradeoff; ablations confirm the RTA shield is essential, with its removal reducing MSI by 1.27–1.84× and degrading state norms. A preference-conditioned extension recovers the full tradeoff frontier from a single model at 2 11 of training compute, and SAC experiments confirm the results are algorithm-agnostic across discrete and continuous domains. A 12-state 3D quadrotor case study extends the framework to higher-dimensional systems where classical STC design is analytically intractable: a SAC agent reaches MSI = 0.302 s (94% of τmax) at 0% RTA, while classical Lyapunov-STC remains pinned at τmin and a fixed-rate LQR controller at the same average interval crashes within two control updates. Robustness to ±30% plant-mass variation and additive disturbances confirms graceful degradation, with the RTA absorbing what the learned policy cannot.