Certified Set Convergence for Piecewise Affine Systems

arXiv:2604.00286 · eess.SY, cs.SY · Submitted 2026-03-31 · Read on arXiv

Listen

Radio episode about this paper

Transcript

Introduction to the show: ident: Robotics Radio. Generated commentary on the latest robotics and control papers.

Rosa: We're looking at the paper: "Certified Set Convergence for Piecewise Affine Systems".

Dev: Here's the gist of the paper. Safety-critical control of piecewise affine (PWA) systems under bounded additive disturbances requires guarantees not for individual states but for entire state sets simultaneously: a single control action must steer every state in the set toward a target, even as sets crossing mode boundaries split and evolve under distinct affine dynamics.

Rosa: Certifying such set convergence via neural Lyapunov functions couples the Lipschitz constants of the value function and the policy, yet certified bounds for expressive networks exceed true values by orders of magnitude, creating a certification barrier.

Dev: This paper resolves this through a three-stage pipeline that decouples verification from the policy.

Rosa: System Modeling and Representation Piecewise affine (PWA) systems model hybrid dynamics where continuous state evolution switches across polyhedral regions.

Paper summary: Rosa: We’ve looked at the paper "Certified Set Convergence for Piecewise Affine Systems" and how it tackles the problem of guaranteeing entire state sets converge in piecewise affine systems, even when they split up across mode boundaries.

Dev: The authors present a three-stage pipeline where they train a deep Hamilton-Jacobi value function as a Lyapunov candidate, then distill a set-based controller from that learned actor using regret minimization, and finally provide theoretical guarantees at three certification levels.

Taro: What this really means for the world is that we can get formal safety guarantees for complex control systems under disturbances by separating the verification process from the policy training process.

Rosa: It’s about using deep Hamilton-Jacobi backward reachability to create a verified Lyapunov function, and then using a permutation invariant Deep Sets controller to produce one common action for all parts of the set.

Dev: The key is that Theorem one certifies trajectory specific convergence by propagating zonotopes through the learned value network, while Theorem two closes the terminal guarantee using spectrally constrained local networks <ref:2604.00286#pg1>.

Taro: The main limitation they point out is that Theorem one only certifies trajectory specific convergence, and they also rely on the assumption of CQLF for their value function approach <ref:2604.00286#pg1>.

Rosa: So, this paper offers a framework where we can have high-level safety assurances for complex hybrid dynamics by combining deep learning techniques with formal verification methods in a structured way.

Conclusion: Rosa: So, we're wrapping up on "Certified Set Convergence for Piecewise Affine Systems," and basically, this paper is about getting formal proof that groups of states in complex hybrid systems will all end up in a safe spot under some kind of disturbance.

Dev: Yeah, the authors are tackling those piecewise affine systems where the dynamics jump between different rules, and they’re showing how to guarantee convergence for the entire set at once, not just individual states.

Taro: What's interesting is that they do this by using a deep Hamilton-Jacobi value function as a guide—essentially making the system learn what the safe path looks like through its optimal return calculation.

Rosa: And then they take that learned guidance and turn it into a concrete control action for the whole set, which they call a set-based controller, using some Deep Sets architecture to handle those variable shapes of state groups.

Dev: The big deal here is their three-stage pipeline: training the value function, distilling the controller, and then giving you these three levels of certification—trajectory specific convergence, and a terminal guarantee for when things settle down near a target.

Taro: It means we can move beyond just hoping an AI finds *a* safe path; we get mathematical certificates that prove *every* state in that group will follow the desired behavior, even when the world throws some noise at it.

Rosa: The authors acknowledge there are still some hurdles, like relying on a specific assumption about their value function approach and the fact that the first theorem only guarantees convergence along one specific path.

Dev: Exactly; they’re not claiming this solves every single problem in hybrid systems right now, but it’s a huge step in separating the learning of *what* to do from the rigorous verification of *if* it will actually work safely under real-world conditions.

Yanliang Huang, *, Peng Xie, *, Zhen Zhang, *, Wenyuan Wu, *, Zhuoqi Zeng2

School of Computation, Information and Technology, Technical University of Munich · School of Engineering, Hainan Bielefeld University of Applied Sciences

eess.SY, cs.SY

Submitted: 2026-03-31

Updated: 2026-11-09

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 99/100

The gist: Safety-critical control of piecewise affine (PWA) systems under bounded additive disturbances requires guarantees not for individual states but for entire state sets simultaneously…

Key concepts

Piecewise Affine (PWA) Systems
These systems model hybrid dynamics where the continuous state evolution changes based on which polyhedral region the state currently occupies. They are common in robotics and control applications, representing systems that switch between different linear or affine dynamics at specific boundaries.
Zonotopes
Zonotopes are a mathematical tool used to represent sets of possible states in PWA systems. They define these sets using a center point and a generator matrix. This representation allows for the analysis of reachable sets and control objectives over continuous time.
Deep HJ Value Function
This is a neural network trained to approximate the optimal travel-cost formulation, which encodes reachability. It acts as a Lyapunov candidate, where its negative sublevel set proves that the system's trajectory under the optimal policy will stay within safe bounds relative to a target state.
Set-Based Controller Distillation
This stage takes a standard neural network actor and distills it into a controller that operates directly on zonotopes. It uses 'Deep Sets' architecture to process variable-size zonotope fragments, allowing the resulting policy to generate a single control action suitable for steering entire sets of states.

Terminology

Summary

Safety-critical control of piecewise affine (PWA) systems under bounded additive disturbances requires guarantees not for individual states but for entire state sets simultaneously: a single control action must steer every state in the set toward a target, even as sets crossing mode boundaries split and evolve under distinct affine dynamics. p. 0 Certifying such set convergence via neural Lyapunov functions couples the Lipschitz constants of the value function and the policy, yet certified bounds for expressive networks exceed true values by orders of magnitude, creating a certification barrier. p. 0 This paper resolves this through a three-stage pipeline that decouples verification from the policy. p. 0

System Modeling and Representation

Piecewise affine (PWA) systems model hybrid dynamics where continuous state evolution switches across polyhedral regions. The system dynamics are defined by xk+1 = Ai xk + Bu,i uk + wk, xk ∈ Xi. p. 1 Zonotopes are defined as Z = c + Gα: α ∈ [−1, 1]p, where c ∈ R n is the center and G ∈ R n×p is the generator matrix. p. 1 The control objective is to find a set-based controller πset such that for a given initial zonotope Z0, the reachable set converges to a compact certification target T within finite time K. p. 1

The Three-Stage Pipeline

The pipeline comprises three stages: Section III-A trains a deep HJ value function V∗ as a Lyapunov candidate, Section IIIB distills a set-based controller from the pointwise actor, and Section III-C provides the theoretical guarantees at three certification levels.

Deep HJ as Lyapunov Guidance

The travel-cost formulation encodes reachability through a calibrated reward r(x) = (-(rt - x - xt 2), x - xt squared < rt, 0, otherwise. Since every reward term satisfies r(x) ≤ 0, the optimal discounted return V∗(x) = P∞ k=0 γ k r(x∗k), where the state trajectory under the optimal policy starting from x∗ 0 = x is non-positive for all x. p. 2 The negative sublevel set proves to recover the backward reachable set [12]. p. 2 The Lyapunov candidate is then defined as V∗(x) = Qϕ(x, πθ(x)). p. 2 The certification target is T = cT = min x∈∂Tref V∗(x). p. 2

Set-Based Policy Distillation

To account for set geometry, the trained actor serves as the oracle π∗ for distilling a set-based controller πset that maps zonotopes to a single control action. The set-actor uses a Deep Sets [16] architecture to process variable-size zonotope fragments as unordered token sets. p. 0 The training loss is L = 1/HBz X H k=1 X Bz b=1 meanop30%ReLUregk,b. p. 2

Theoretical Guarantees and Certification

Theorem 1 certifies trajectory-specific set convergence via zonotope propagation through the learned value function. This theorem provides a verified certificate that decouples the verification from the policy entirely. p. 0 Theorem 2 closes the terminal guarantee using spectrally constrained local networks to bound the gap within a near-target core. p. 0 The diagnostic condition (Propositions 1–2) quantifies the expressivity–certifiability trade-off. p. 0

Experimental Results and Trade-offs

The set-actor achieves 100% strict containment on all four benchmarks and is the only tested method to do so. The diagnostic condition (16) quantifies the expressivity–certifiability trade-off. p. 0 Empirical analysis shows that global spectralnorm Lipschitz bounds severely over-estimate the true local behavior, and a spectrally constrained certificate network resolves this gap by achieving a feasible margin. p. 4 The dimension barrier is resolved by restricting the local certificate training domain C to a compact neighbourhood of Tref. p. 6

Conclusion

This paper presents a framework for certified set convergence in neural-controlled PWA systems combining deep HJ reachability with a set-based controller trained via regretbased policy distillation. The three-level certification hierarchy operates at different scales, with Theorem 1 certifying trajectory-specific convergence and Theorem 2 closing the terminal guarantee. p. 7 The main limitations are the CQLF assumption and the trajectory-specific nature of Theorem 1. p. 7

REFERENCES

[1] D. Liberzon, Switching in Systems and Control. p. 7 Birkhauser Boston, ¨ 2003 <ref:2604. p. 700286#pg7>

[5] Y.-C. Chang, N. Roohi, and S. p. 7 p. 7 Gao, “Neural Lyapunov control,” in Advances in Neural Information Processing Systems (NeurIPS), vol. p. 7 p. 7 32, 2019 <ref:2604. p. 0 p. 0 p. 000286#pg8>

[16] M. Zaheer, S. Kottur, S. Ravanbakhsh, B. p. 7 Poczos, R. Salakhutdinov, ´ and A. p. 7 J. Smola, “Deep sets,” in Advances in Neural Information Processing Systems (NeurIPS), 2017

[17] S. Bansal and C. p. 7 J. Tomlin, “DeepReach: A deep learning approach to high-dimensional reachability,” in IEEE International Conference on Robotics and Automation (ICRA), 2021

[18] J. F. Fisac, M. Chen, C. J. Tomlin, and S. p. 7 Shankar Sastry, “Reachavoid problems with time-varying dynamics, targets and constraints,” in International Conference on Hybrid Systems: Computation and Control, 2015

[20] T. P. Lillicrap, J. p. 7 J. Hunt, A. Pritzel, N. p. 7 Heess, T. Erez, Y. Tassa, D. Silver, and D. p. 7 Wierstra, “Continuous control with deep reinforcement learning,” in International Conference on Learning Representations (ICLR), 2016

[23] M. Althoff, “An introduction to cora 2015,” in Proc. p. 7 p. 7 p. 7 of the workshop on applied verification for continuous and hybrid systems, 2015

[24] P. Xie, J. Betz, D. M. Raimondo, and A. p. 7 Alanwar, “Data-driven reachability analysis for piecewise affine systems,” in 2025 IEEE 64th Conference on Decision and Control (CDC). p. 7 IEEE, 2025

[3] B. Schurmann and M. p. 0 p. 0 p. 0 Althoff, “Optimal control of sets of solutions ¨ to formally guarantee constraints of disturbed linear systems,” in 2017 American control conference (ACC). p. 7 p. 7 p. 7 IEEE, 2017 <ref:2604. p. 7 p. 7 p. 700286#pg6>

[4] ——, “Optimizing sets of solutions for controlling constrained nonlinear systems,” IEEE Transactions on Automatic Control, vol. 66, no. 3, pp. 981–994, 2020 <ref:2604. p. 7 p. 7 p. 700286#pg6>

[8] G. Singh, T. Gehr, M. Mirman, M. Puschel, and M. p. 7 Vechev, “Fast and ¨ effective robustness certification,” in Advances in Neural Information Processing Systems (NeurIPS), vol. p. 7 31, 2018 <ref:2604. p. 000286#pg6>

[9] N. Kochdumper, C. p. 7 p. 7 p. 7 Schilling, M. p. 7 p. 7 p. 7 Althoff, and S. p. 0 p. 0 Bak, “Open- and closed-loop neural network verification using polynomial zonotopes,” in NASA Formal Methods (NFM), 2023

[11] T. J. Bird, H. C. Pangborn, N. p. 7 Jain, and J. p. 7 P. Koeln, “Hybrid zonotopes: A new set representation for reachability analysis of mixed logical dynamical systems,” Automatica, vol. p. 7 154, p. 111107, 2023

[9] N. Kochdumper, C. Schilling, M. Althoff, and S. Bak, “Open- and closed-loop neural network verification using polynomial zonotopes,” in NASA Formal Methods (NFM), 2023

[15] N. Jovanovic, M. p. 7 p. 7 Balunović, M. Baader, and M. p. 7 p. 7 Vechev, “On the ´ paradox of certified training,” Transactions on Machine Learning Research (TMLR), 2022

[19] C. Dawson, S. Gao, and C. p. 7 p. 7 Fan, “Safe control with learned certificates: A survey of neural Lyapunov, barrier, and contraction methods for robotics and control,” IEEE Transactions on Robotics, vol. p. 7 p. 7 39, no. 3

[21] Y. Bengio, J. Louradour, R. p. 7 p. 7 Collobert, and J. p. 7 Weston, “Curriculum learning,” in International Conference on Machine Learning (ICML), 2009

[23] M. Althoff, “An introduction to cora 2015,” in Proc. of the workshop on applied verification for continuous and hybrid systems, 2015

[3] B. Schurmann and M. Althoff, “Optimal control of sets of solutions ¨ to formally guarantee constraints of disturbed linear systems,” in 2017 American control conference (ACC). IEEE, 2017 <ref:2604.00286#pg6>

[4] ——, “Optimizing sets of solutions for controlling constrained nonlinear systems,” IEEE Transactions on Automatic Control, vol. 66, no. 3, pp. 981–994, 2020 <ref:2604.00286#pg6>

[12] P. Solanki, I. p. 0 p. 0 El-Hajj, J. p. 7 J van Beers, E. p. 7-J van Kampen, and C. p. 7 C de Visser, “Formalizing the relationship between HamiltonJacobi reachability and reinforcement learning,” arXiv preprint arXiv:2601. p. 708050, 2026

[13] ——, “Certifying Hamilton-Jacobi reachability learned via reinforcement learning,” arXiv preprint arXiv:2602.16475, 2026

[14] M. Wendl, L. Koller, T. Ladner, and M. p. 7 p. 7 Althoff, “Training verifiably robust agents using set-based reinforcement learning,” arXiv preprint arXiv:2408. p. 7 p. 709112, 2024

[5] Y.-C. Chang, N. Roohi, and S. Gao, “Neural Lyapunov control,” in Advances in Neural Information Processing Systems (NeurIPS), vol. 32, 2019 <ref:2604.00286#pg8>

[6] H. Dai, B. Landry, L. Yang, M. Pavone, and R. p. 7 Tedrake, “Lyapunovstable neural-network control,” in Robotics: Science and Systems (RSS), 2021 <ref:2604. p. 700286#pg8>

[7] S. Chen, M. Fazlyab, M. p. 7 Morari, G. J Pappas, and V. p. 7 M Preciado, “Learning Lyapunov functions for piecewise affine systems with neural network controllers,” arXiv preprint arXiv:2008. p. 706546, 2020

[11] T. J Bird, H. C Pangborn, N Jain, and J P Koeln, “Hybrid zonotopes: A new set representation for reachability analysis of mixed logical dynamical systems,” Automatica, vol. p. 7 154, p. 111107, 2023

[9] N. Kochdumper, C. Schilling, M. Althoff, and S Bak, “Open- and closed-loop neural network verification using polynomial zonotopes,” in NASA Formal Methods (NFM), 2023

[15] N. Jovanovic, M. Balunović c, M. Baader, and M. Vechev, “On the ´ paradox of certified training,” Transactions on Machine Learning Research (TMLR), 2022

[19] C. Dawson, S. Gao, and C. Fan, “Safe control with learned certificates: A survey of neural Lyapunov, barrier, and contraction methods for robotics and control,” IEEE Transactions on Robotics, vol. 39, no. 3

[21] Y. Bengio, J. Louradour, R. Collobert, and J Weston, “Curriculum learning,” in International Conference on Machine Learning (ICML), 2009

[23] M. Althoff, “An introduction to cora 2015,” in Proc. of the workshop on applied verification for continuous and hybrid systems, 2015

[3] B. Schurmann and M. Althoff, “Optimal control of sets of solutions ¨ to formally guarantee constraints of disturbed linear systems,” in 2017 American control conference (ACC). IEEE, 2017 <ref:2604.00286#pg6>

[4] ——, “Optimizing sets of solutions for controlling constrained nonlinear systems,” IEEE Transactions on Automatic Control, vol. 66, no. 3, pp. 981–994, 2020 <ref:2604.00286#pg6>

[12] P. Solanki, I. El-Hajj, J J van Beers, E. p. 7-J van Kampen, and C C de Visser, “Formalizing the relationship between HamiltonJacobi reachability and reinforcement learning,” arXiv preprint arXiv:2601. p. 708050, 2026

[13] ——, “Certifying Hamilton-Jacobi reachability learned via reinforcement learning,” arXiv preprint arXiv:2602.16475, 2026

[14] M. Wendl, L. Koller, T. Ladner, and M. Althoff, “Training verifiably robust agents using set-based reinforcement learning,” arXiv preprint arXiv:2408.09112, 2024

[5] Y.-C. Chang, N Roohi, and S Gao, “Neural Lyapunov control,” in Advances in Neural Information Processing Systems (NeurIPS), vol. p. 7 32, 2019 <ref:2604.00286#pg8>

[6] H Dai, B Landry, L Yang, M Pavone, and R Tedrake, “Lyapunovstable neural-network control,” in Robotics: Science and Systems (RSS), 2021 <ref:2604.00286#pg8>

[7] S Chen, M Fazlyab, M Morari, G J Pappas, and V M Preciado, “Learning Lyapunov functions for piecewise affine systems with neural network controllers,” arXiv preprint arXiv:2008.

Green — the paper says it, and the page is printed. Red — no passage carries it, so treat that sentence as unverified. The same check runs on your own text: check a summary against its source, $5.

Improvements for AI systems

  1. Set-based control pipeline: Implement a controller that processes variable-size zonotope fragments as unordered token sets using a Deep Sets architecture to produce a single control action for all state sets simultaneously.

  2. Trajectory certification: Achieve trajectory certified set convergence by propagating zonotopes through the learned value function via zonotope neural network propagation [8], [14], yielding a verified certificate that decouples the verification from the policy entirely.

  3. Terminal guarantee: Establish a terminal guarantee using a spectrally constrained local certificate network to close the gap, ensuring that for any zonotope fragment entering a core region C, its diameter is bounded by 2c lT - L cert V etaC/(1−sigma¯)γ Lcert V LB Lcert π(17).

Sources

Related papers