{"paper":{"title":"Learning When to Act: Communication-Efficient Reinforcement Learning via Run-Time Assurance","license":"http://creativecommons.org/licenses/by/4.0/","headline":"A single RL policy learns both control actions and sparse timing decisions while a Lyapunov run-time assurance shield enforces stability via LQR overrides.","cross_cats":["cs.RO"],"primary_cat":"cs.LG","authors_text":"Adam Haroon, Cody Fleming, Erick J. Rodr\\'iguez-Seda, Tristan Schuler","submitted_at":"2026-05-11T23:55:15Z","abstract_excerpt":"Safe reinforcement learning (RL) typically asks $\\textit{what}$ an agent should do. We ask $\\textit{when}$ 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 focus on stabilization around a known equilibrium, where CARE-based LQR backups, Lyapunov certificates, and classical Lyapunov-STC are well defined, enabling clean comparison against analytical baselines. A run-time assurance (RTA) layer overrides the policy via a one-step-ahead Lyapunov prediction and a precomputed LQR b"},"claims":{"count":4,"items":[{"kind":"strongest_claim","text":"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 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, makes sparsity safe.","source":"verdict.strongest_claim","status":"machine_extracted","claim_id":"C1","attestation":"unclaimed"},{"kind":"weakest_assumption","text":"The approach assumes a known equilibrium point where CARE-based LQR backups and Lyapunov certificates are well-defined, allowing the RTA to provide strictly stronger safety guarantees than expectation-based methods.","source":"verdict.weakest_assumption","status":"machine_extracted","claim_id":"C2","attestation":"unclaimed"},{"kind":"one_line_summary","text":"Learned policies with runtime Lyapunov shields achieve substantially higher communication intervals than baselines while maintaining stability on inverted pendulum, cart-pole, and quadrotor systems.","source":"verdict.one_line_summary","status":"machine_extracted","claim_id":"C3","attestation":"unclaimed"},{"kind":"headline","text":"A single RL policy learns both control actions and sparse timing decisions while a Lyapunov run-time assurance shield enforces stability via LQR overrides.","source":"verdict.pith_extraction.headline","status":"machine_extracted","claim_id":"C4","attestation":"unclaimed"}],"snapshot_sha256":"1c56a1bce8c46f358765398c6867bf6a22cdbc346d6e410efcf732e5fc9d31ae"},"source":{"id":"2605.12561","kind":"arxiv","version":1},"verdict":{"id":"372e4dac-007a-41bf-bbdf-85f2f9e24c73","model_set":{"reader":"grok-4.3"},"created_at":"2026-05-14T20:39:20.796514Z","strongest_claim":"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 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, makes sparsity safe.","one_line_summary":"Learned policies with runtime Lyapunov shields achieve substantially higher communication intervals than baselines while maintaining stability on inverted pendulum, cart-pole, and quadrotor systems.","pipeline_version":"pith-pipeline@v0.9.0","weakest_assumption":"The approach assumes a known equilibrium point where CARE-based LQR backups and Lyapunov certificates are well-defined, allowing the RTA to provide strictly stronger safety guarantees than expectation-based methods.","pith_extraction_headline":"A single RL policy learns both control actions and sparse timing decisions while a Lyapunov run-time assurance shield enforces stability via LQR overrides."},"references":{"count":31,"sample":[{"doi":"","year":2019,"title":"A. Abels, D. Roijers, T. Lenaerts, A. Nowé, and D. Steckelmacher. Dynamic weights in multi-objective deep reinforcement learning. InInternational Conference on Machine Learning, pages 11–20. PMLR, 201","work_id":"a099ba9b-0c0a-474e-83c6-97c4ca689983","ref_index":1,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2017,"title":"J. Achiam, D. Held, A. Tamar, and P. Abbeel. Constrained policy optimization. InInternational Conference on Machine Learning, pages 22–31. PMLR, 2017","work_id":"7fdc9762-60da-432e-bc31-8b4b4df1d8be","ref_index":2,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2025,"title":"S. Aggarwal, D. Maity, and T. Ba¸ sar. Interq: A dqn framework for optimal intermittent control. IEEE Control Systems Letters, 2025","work_id":"baa7daaf-d36b-494e-b330-45fa25673810","ref_index":3,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2018,"title":"M. Alshiekh, R. Bloem, R. Ehlers, B. Könighofer, S. Niekum, and U. Topcu. Safe reinforcement learning via shielding. InProceedings of the AAAI Conference on Artificial Intelligence, volume 32, 2018","work_id":"b22ca70f-4c39-4e2c-b237-a18a9ff40fe2","ref_index":4,"cited_arxiv_id":"","is_internal_anchor":false},{"doi":"","year":2021,"title":"Altman.Constrained Markov Decision Processes","work_id":"201db9f5-e923-4c71-8b35-465badee3afc","ref_index":5,"cited_arxiv_id":"","is_internal_anchor":false}],"resolved_work":31,"snapshot_sha256":"178017d36ff07c71762999f698794874f1070cb5f9d883d861cf4b8ba0811ad1","internal_anchors":2},"formal_canon":{"evidence_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"author_claims":{"count":0,"strong_count":0,"snapshot_sha256":"258153158e38e3291e3d48162225fcdb2d5a3ed65a07baac614ab91432fd4f57"},"builder_version":"pith-number-builder-2026-05-17-v1"}