- Rating
- 2
- Confidence
- 3
- Informativeness
- 1
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
The paper proposes a sampling-based pipeline for estimating the region of attraction (RoA) of a fixed closed-loop system, applied to small neural controllers on CartPole (4D) and quadrotor attitude (6D). A quadratic Lyapunov function from the linearization gives a small seed set, which is enlarged into a "trusted core" by checking that boundary samples reach the seed within a horizon. Convergent states found by uniform or pullback sampling are covered with Euclidean balls or ellipsoids, plus flow-based patches for the leftovers. A patch is kept if a Clopper–Pearson bound on its rollout failure rate is below 1%. The final union is validated on 8,192 fresh samples from the added region (zero failures, bound at 99.9% confidence). The represented region is 7.8× (CartPole) and 6.0× (quadrotor) the core's volume within the DoI, and random sampling with balls gives the largest region on both systems.
Critical strengths
-
The setting is practical. Analyzing a given controller after training, without retraining it or assuming a particular controller structure, is often what is needed in practice, and the reach-to-core idea is simple and works with any simulator.
-
The final statistical check is done properly: the validation samples are independent of seed learning and patch construction, the reported bound is correct (), and the authors are upfront that this is not a formal certificate. The 2×2 ablation showing that seed coverage does not predict represented volume is a useful observation.
Critical weaknesses
-
There are no baselines and no ground truth, so the results can't be judged. The only reference point is the authors' own quadratic core. We don't know whether 16.45% of the CartPole DoI is most of the true RoA or a small part of it; since the controller was trained on this DoI, I would expect much of it to converge. This is cheap to check: uniform rollouts over the DoI (200,000 were already run for random exploration) give a Monte Carlo estimate of the RoA volume, and the method's region should be reported as a fraction of it. Comparisons under the same final validation are also needed, e.g., the largest sublevel set that passes the same reach-to-core test, a classifier (kNN or a small MLP) trained on the rollout labels, and a post-hoc neural Lyapunov/Zubov method for a fixed system (e.g., Liu et al. 2025, which the paper cites). The learning contribution is also thin for ICLR: the pipeline is a set of sampling and covering heuristics, and the representation is a union of roughly 1,400 balls that does not generalize beyond the sampled seeds.
-
The reliability claim is weaker than it reads. (a) Everything depends on lying inside the RoA, yet the core is checked only on its boundary (4,000 + 8,192 boundary rollouts reaching , which in turn is checked only by sampled boundary decrease). Interior points of the core are never tested, and the final validation samples only the added region. For the quadrotor, , so the core is about 20× the radius of the part with any Lyapunov evidence. (b) The "99.9% at 99.9% confidence" figure is simply the zero-failure Clopper–Pearson bound for . It is identical for every configuration and only says that the failure volume under uniform sampling of the added region is below about 0.08%; thin failure sets inside accepted patches would go unnoticed. (c) Fig. 2(e) appears to show accepted quadrotor patches extending to rad, beyond where the Euler kinematics in Eq. (21) are singular and far outside the DoI (), so part of the "validated" region is where the model is not well defined.
-
each configuration is run once, so rankings such as random–ball vs. random–ellipsoid (16.45% vs. 13.48%) may not be robust, and on the quadrotor the better geometry flips with the seed strategy (pullback: ellipsoid 68.5% vs. ball 27.6%). The flow layer, presented as a main component, adds 0.0009% of the CartPole DoI (about 9 of the volume samples) and 0.2% on the quadrotor. The appendix algorithms are too high-level to reproduce: quantile levels, neighborhood sizes, per-patch sample counts, coverage targets, and patch budgets are not given. Fig. 2 is hard to read, and its caption calls the plots slices while Sec. 4.1 calls them projections.
Questions for authors
- What fraction of the DoI is in the RoA according to brute-force Monte Carlo, and what fraction of that does cover? How does the method compare with the simple baselines above under the same validation?
- How do you justify that every interior point of converges? Did you test any interior samples of the core?
- Do accepted quadrotor patches contain states with ? If so, how are they handled in simulation, and do the volume and validation numbers change when they are excluded?
- How much do the occupancy numbers and the ranking of the four configurations vary across random seeds?
- What is the total cost of the full pipeline (number of rollouts and wall-clock time), not just seed discovery?
Suggestion for future work (not required here): a study of how the number of patches and rollouts grows with state dimension would show whether ball covering remains viable beyond 6D.