acceptodds
Under review as a conference paper at ICLR 2027

Learning Continuous Neural Representation of Stochastic Hybrid Systems

Abstract

A stochastic hybrid system (SHS) is governed by a stochastic differential equation (SDE) describing the continuous dynamics and a Markov reset kernel triggered on the guard surface. Its probability evolution can be described by a hybrid Fokker–Planck (HFP) equation with a partial differential term corresponding to the SDE and an integral term arising from the reset kernel. This work shows that such an SHS can be approximated by an SDE in a higher-dimensional latent space where the sample paths are continuous. The key to this result is to encode different branches of the reset kernel using auxiliary variables, transforming the resets into deterministic ones that enable topological gluing. By the embedding theorem, the glued manifold can then be embedded into a higher-dimensional Euclidean space. We show that the probability evolution on the embedded image no longer requires explicit reset terms in the HFP equation. Building on this theorem, we design a loss that matches the evolving state distributions, enabling a single latent SDE to recover the probability evolution of the SHS without mode labeling, trajectory segmentation, or event-based simulations.

Primary area:
learning on time series and dynamical systems
Keywords:
Learning dynamical system, stochastic differential equation, hybrid systems, Fokker–Planck equation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_mGAMOpenReview ↗
Rating
2(−1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Hybrid systems are those which have continuous dynamics in the bulk, and a reset kernel that injects particles back into system, should they exit the system through its boundary. Stochastic hybrid systems are those in which the bulk dynamics and/or the reset kernel is stochastic. When the reset kernel is a one-to-one mapping, one can embed the dynamics into a higher dimensional ambient space in which they appear piecewise continuous. This is useful, since modelling the continuous dynamics alongside the discrete events in the base space is challenging (I think the authors could elaborate on this).

In this work, the authors construct a method for applying this analysis to stochastic hybrid systems with a stochastic reset kernel. They do so by extending the state space with an auxiliary variable (something akin to a momentum), whose dynamics are tuned to approximate the action of the reset kernel on the base manifold, through the interaction of the auxiliary variable with the base variable.

The present an algorithm for implementing their theory and illustrate the effectiveness of their result on two test cases. Finally, they compare their method against others in the field, and find it to be the most accurate.

Critical strengths

Their proofs are very well attributed, and appear to be quite sound.

The numerical tests are quite convincing. Their approximated dynamics correctly captures the evolution of particle ensembles over timescales ~10x times their training horizon.

Critical weaknesses

The authors could do a better job explaining the importance of this work to non-experts. Why is working with a hybrid system so difficult? Why does its treatment benefit from having a piecewise continuous analog.

The authors could also benefit from adding more exposition into the manuscript, at large. It seems as though the theory is a proof of existence? It does not predict the value of F, G, E, or D, but it provides insight into what size of latent space permits their existence? and then the numerical results show that the trained network of the appropriate form indeed reproduces the SHS? It might be nice to outline this for the reader.

I found Section 4.3 and "Cost Function Design" the most dense to parse through.

Questions for authors

I am slightly confused about how to interpret the dynamics in the lifted space. When is vv able to modify the dynamics of xx? Is it always, with the discrete jumps only appearing due the projection applied by the decoder x=D(z)x = D(z)? Or is (x−,v−)=(x−+v−,v−)(x^−, v^−) = (x^− + v^−, v^−) still being applied at the boundary in some manner that leaves the map piecewise continuous but not continuous?

Do you have to know S a priori to apply this analysis? Or can you take any dataset whose dynamics appears to have discrete reset events and apply it? I imagine you can always infer S as the set of all x in your data where a discrete event occurs?

Are there some kernels that make this make this approximation more difficult than others? I feel like that was an under discussed topic in the current version of the manuscript. But please correct me if you disagree! Please do not just add a discussion of this to pander to me, if you disagree.

3 comments

Critical Weaknesses

1. Why is working with a hybrid system so difficult?

Learning a hybrid system requires capturing continuous dynamics, event boundaries, and reset behavior simultaneously. Small prediction errors near a guard can change whether or when a reset occurs, producing substantially different subsequent trajectories. Stochastic resets additionally require learning a distribution of possible outcomes.

2. Why does a continuous representation help?

We aim to avoid explicit discontinuities during rollout, as errors near reset events can be amplified and compromise rollout stability. Evolving a continuous latent SDE avoids explicit event detection and reset operations during integration. The latent sample paths are continuous; “piecewise smooth” refers to the geometric representation and coefficients. We will clarify this motivation.

3. Is the theory primarily an existence result?

Yes, the geometric embedding result primarily establishes existence under the stated assumptions. Section 4.3 additionally provides a construction of augmented dynamics targeting the original system’s state-density evolution.

4. Does the theory determine (F,G,E,D)(F,G,E,D), or mainly the required latent dimension?

The theory provides a sufficient/general embedding dimension, dz≥4nd_z\geq 4n, rather than a minimum dimension or a unique representation. Given an embedding EE, the transformed coefficients F,GF,G follow from Itô’s formula, and the ideal decoder selects a representative of each equivalence class. The practical algorithm learns their neural approximations from trajectories.

5. Do the numerical results demonstrate that the trained model reproduces the SHS?

The experiments empirically evaluate how closely the learned model reproduces the SHS’s state and path distributions, including beyond the training horizon. We will distinguish this empirical validation from the theoretical representation results.

Recovering the graph structure of the hybrid system remains an interesting future work, as we need to extract the discrete structure from the decoder.

6. Section 4.3 and “Cost Function Design” are difficult to follow.

Sec. 4.3 provides an intuitive construction to show how the variable vv can encode different branches of resets to reproduce stochastic resets at the guard.

The cost function design section basically tells us to match the path distribution. We will make the motivation of the two sections clearer.

Questions for Authors

1. When does vv affect xx? Are resets still applied, and are the lifted paths continuous?

In the construction of Section 4.3, xx follows its original SDE between events. The auxiliary variable vv determines the displacement at the guard through

(x+,v+)=(x−+v−,v−).(x^{+},v^{+})=(x^{-}+v^{-},v^{-}).

Thus, resets remain explicit in the augmented (x,v)(x,v) coordinates. After gluing,

z+=E(x−+v−,v−)=E(x−,v−)=z−,z^{+}=E(x^{-}+v^{-},v^{-})=E(x^{-},v^{-})=z^{-},

so the embedded trajectory is continuous. The ideal decoder recovers the observed jumps through its discontinuity at the glued seam; the neural decoder approximates this behavior at sampled observation times.

2. Must SS be known? Can the method use unlabelled trajectories, or infer SS from observed jumps?

The analytical construction uses SS, but the learning algorithm requires only state trajectories, without a known guard, reset kernel, or event labels. It can therefore be applied directly to unlabelled trajectories within our SHS setting. Observed jumps can help locate the guard, but finite samples do not generally identify the entire SS exactly. Such identification is unnecessary for our training procedure.

3. Are some reset kernels harder to approximate?

At the level of branch representation, we do not see a kernel-specific topological obstruction. In our formulation, any reset kernel can be represented through a conditional distribution over v=x+−x−v=x^+-x^-, while the reset R(x,v)=(x+v,v)R(x,v)=(x+v,v) remains deterministic. The kernel determines the distribution over reset branches, without restricting the branch representation itself. Our experiments demonstrate successful learning with both Gaussian-mixture and uniform reset distributions.

The difficulty may instead arise in learning the augmented dynamics required to redistribute probability mass and match the prescribed vv-distribution at the guard. For sharply concentrated distributions, the redistribution requirements may demand rapidly varying drift coefficients, making the dynamics hard to learn.

Reviewer_mGAMOpenReview ↗

Revised score

Thank you for the responses! They have clarified my questions about the role of vv and the guard. I find the ideas of the paper very novel, but following the discussion with Reviewer DFnu (see my comment above on Lemma 1), I have remaining concerns about the theory. I think it might need some extra fine-tuning before publication. Therefore, I currently sit between weak accept and weak reject.

Thank you for the question. They are very helpful for us in improving the clarity of this work.

We have replied to the question after your discussion with Reviewer DFnu.

The TL;DR is that the reset in A-SHS does not need R(S×V)⊆(∂M×V)R(S \times V) \subseteq (\partial M \times V). We require R(S×V)⊆∂(M×V)R(S \times V) \subseteq \partial( M \times V), as our argument in Lemma 1 is for the product space M×VM \times V, not MM.

Reviewer_DFnuOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces an augmented-state stochastic hybrid system (SHS) which learns a model of both the state and deterministic reset. This allows the joint stochastic system plus Markov reset kernel to be represented by one stochastic system. It is shown that this construction can be used to model timeseries data obtained from toy dynamical systems with stochastic reset kernels at the boundary surface.

## Presentation The general argument of the paper of replacing an SHS plus reset kernel with one augmented stochastic model is easy to follow. However, I found there to be essentially no motivation to the class of problems an SHS can be applied to or sufficient motivation for why one augmented model is required over previous methods.

The introduction mentions robotics and autonomous driving as examples of dynamical systems. It is not clear how these systems are well modelled by stochastic dynamics with Markov reset kernels on the surface boundary. Additionally, it is not clear what class of problems a joint stochastic model can be applied to or why this model would be more performant than previous work.

## Informativeness The demonstration that stochastic dynamics plus Markov reset kernel can be modelled with a single latent SDE over an augmented space is interesting. However, the usefulness of this construction is unclear since the class of problems this method can be applied to is not described. Additionally, the objective function contains five distinct loss terms including a state distribution loss, an unconditional path distribution loss and a conditional path distribution loss. It is therefore difficult to untangle the key training contributions even with the aid of the ablation in Table 2.

Soundness

My main concerns are the applicability of the proposed method rather than its soundness. The numerical experiments appear to suggest that the implementation is sound given the performance on toy problems.

However, there seem to be missing properties of the reset map which are required to establish the augmented SHS. Lemma 1 uses the result of Simic et al 2005, however it is not clear that the relevant assumptions have been demonstrated for the reset (x,v)→(x+v,v)(x, v) \rightarrow (x + v, v). Specifically, A4 of Simic et al. 2005 requires that "The image of every reset lies on the boundary of the corresponding domain." which is not demonstrated.

Critical strengths

The construction of the augmented SHS to represent the stochastic dynamics plus boundary kernel is theoretically interesting. This is demonstrated to perform well on toy stochastic dynamics problems on a bounded domain. If this performance extends to a non-toy time series problem then it would provide a contribution beyond baseline SHS models.

Critical weaknesses

The main weakness is in the methods applicability. It is not clear what problems this method can be applied to other than the toy problems constructed for the experiments. Further, while interesting, it is not clear why this augmented SHS model is desired over previous methods that do not model the dynamics over this augmented state.

The interpretation of the experimental results is lacking. Many previous methods are claimed to diverge on simple toy problems which is not investigated sufficiently. The ablation of the architecture and objective function in Table 2 is difficult to follow since it appears to show that various pieces of the algorithm can be removed and performance improvements over the baseline methods are still obtained.

Questions for authors
  • In Lemma 1, can you demonstrate that the reset map satisfies the assumptions required by Simic et al. 2005?
  • Can you provide an established timeseries application of the SHS model and numerically validate the proposed method on this application in comparison to the previous methods shown Table 1?
  • Can you explain the divergence behaviour of the baselines in Table 1? Specifically, Teng et al 2026b demonstrate strong performance on exactly these problems while the method is claimed to diverge in the present work.
  • Table 2 (and associated expanded ablation tables in the appendix) show that the presented method outperforms previous baselines with various architectural and objective components removed. What is the core/minimal algorithm of the proposed method that improves over baselines?
8 comments

Thank you for the constructive comments.

1. Applicability and motivation (Weakness 1; Q2)

Our main contribution is the nontrivial extension of the deterministic topological framework to stochastic hybrid systems through augmented states. We use synthetic benchmarks to control the topological structure and validate this construction. The continuous latent representation avoids explicit event detection and resets during rollout, where prediction errors near a guard can otherwise be amplified.

To address the request for an established application, we add experiments on simultaneous frictional impacts, a challenging problem in robotic manipulation and locomotion (Halm and Posa, 2024). Small differences in contact ordering and impulse allocation can produce distinct outcomes: a falling object may rotate clockwise or counterclockwise, while an object striking a wall may stop or continue moving. A stochastic reset kernel provides a natural way to model these multiple possible post-impact states.

We evaluate phone drop and box–wall impact against all methods in Table 1. Entries report mean conditional (unconditional) path-distribution losses; lower is better.

CaseOursCHyLLLatent SDELatent SDE⋆SDE MatchingNSF
Phone drop0.561 (0.758)0.843 (9.46)0.635 (4.08)0.570 (0.885)0.658 (5.23)1.09 (25.8)
Box–wall impact0.220 (0.488)0.322 (1.96)0.245 (0.756)0.227 (0.516)0.331 (4.37)Diverged

Our method achieves the lowest mean conditional and unconditional losses in both cases, supporting its applicability to learning trajectory distributions arising from robotic impacts.

2. Assumptions in Lemma 1 (Q1)

Let Ω−=M×V\Omega_-=M\times V and take the receiving domain as Ω+=R(Ω−)\Omega_+=R(\Omega_-). Since R(x,v)=(x+v,v)R(x,v)=(x+v,v) is a global diffeomorphism with inverse R−1(x′,v′)=(x′−v′,v′)R^{-1}(x',v')=(x'-v',v'),

R(S×V)⊆R(∂Ω−)=∂Ω+.R(S\times V)\subseteq R(\partial\Omega_-)=\partial\Omega_+.

This establishes the boundary clause of A4 for this domain choice. The relevant boundary is that of the receiving domain R(Ω−)R(\Omega_-), rather than necessarily that of the unchanged product domain. We will explicitly specify these domains and verify compatibility with the dynamics and remaining gluing assumptions.

3. Baseline divergence (Weakness 2; Q3)

Teng et al. (2026b) studies deterministic hybrid dynamics, whereas our benchmarks introduce stochastic resets. Although the geometric templates are shared, random reset branches change the prediction problem: the same initial state can generate different future trajectories. CHyLL produces one trajectory per initial state and therefore cannot represent a non-degenerate conditional path distribution.

The observed numerical breakdown is consistent with a representational mismatch: a general stochastic reset kernel does not provide the deterministic, invertible reset map required by the earlier gluing construction in the original state space. Our auxiliary displacement v=x+−x−v=x^+-x^- records the reset branch, making the augmented reset deterministic and invertible and enabling the required gluing under the stated assumptions.

This limitation explains why the deterministic construction cannot be directly transferred, but does not imply inevitable numerical divergence. CHyLL has finite reported losses in the new impact experiments, while our method achieves lower conditional and unconditional losses.

4. Core components (Weakness 2; Q4)

The main contributors are latent distribution matching LzL_z and trajectory-distribution supervision. The former aligns encoded and rolled-out latent distributions, while the latter captures temporal dependence and conditioning on the initial state beyond individual-time marginals.

Removing LzL_z increases conditional/unconditional losses by an average of 19.5%/246.6%; removing both path losses increases them by 23.3%/54.7%, worsening both metrics on all 11 benchmarks. Removing either path loss individually has a substantially smaller average effect, suggesting partial redundancy. The observed-state marginal loss LxL_x and stochastic encoder provide smaller average benefits.

A tested reduced objective retains LzL_z, the conditional path loss, and LgL_g, while removing LxL_x and the unconditional path loss. With the default architecture retained, its average conditional/unconditional loss changes are only −0.7%/+3.8%. These results support a simpler objective centered on latent distribution matching and conditional trajectory supervision. Because encoder and loss simplifications were evaluated separately, we will clarify this supported simplification without claiming a uniquely minimal architecture.

  1. Applicability and motivation

The Halm and Posa (2024) systems admit a set of impact outcomes, however the description in the updated paper seems to claim that you have adapted these systems to sample contact priorities randomly. That is, the original systems are set-valued without specifying probabilities over outcomes yet the adapted experiments explicitly introduce stochasticity. The physical evolution of the system may not be captured by this change to the simulator.

Further, this system is again a simulated example. As such, this experiment does not demonstrate the applicability of the proposed model to a real-world timeseries dataset.

  1. Assumptions in Lemma 1

This reasoning establishes the boundary clause of A4, as you suggest. However, it does not establish the remaining hypotheses required by Simic et al. The application of Simic's theorem in Lemma 1 remains unjustified.

  1. Baselines

Thank you for the clarification. I understand the difference between the two approaches is the stochastic reset. As you state, this presents an argument for why the deterministic method cannot be directly transfered but does not explain why the methods diverge numerically. This remains un-accounted for.

  1. Core components

The reduced objective is much simpler and retains effectively the same performance. This objective should be described clearly as the contribution with additional terms introduced if required. The current description remains difficult to follow with all terms introduced first and then later removed to show little change in the performance.

Overall, I appreciate the clarifications provided however I do not believe the questions have been addressed. As such, I maintain my current rating.

Thank you for the follow-up and we would like to futher clarify these points:

1. Applicability and motivation

  • Set-valued impacts and sampling. We agree that the model of Halm and Posa (2024) is set-valued. It specifies the admissible support of the post-impact distribution while leaving the density over that support unspecified. Their Algorithm 1 explicitly samples bounds on normal impulse increments to generate admissible outcomes. Because simultaneous rigid-body contact is underdetermined, generating a trajectory requires a rule for selecting among these outcomes. Our simulation follows this requirement through its contact-priority sampling rule.

  • Synthetic data. We agree that the new experiment uses a synthetic dataset. We used simulation because we could not find a suitable real-world time-series dataset for testing the proposed theoretical construction.

2. Assumptions in Lemma 1

We will address this point in the next response.

3. Baseline divergence

  • ODE uniqueness. (Teng et al., 2026 ICML) uses a deterministic latent ODE. By the Picard–Lindelöf theorem, a locally Lipschitz vector field is a sufficent condition to generate a unique trajectory from each latent initial state. Reproducing distinct reset branches from the same latent state would therefore require a loss of uniqueness. As in our previous reponse, the topological obstruction of the one-to-many map (reset kernel) will lead to vector field that is not L-continuous.

  • Example. Consider x˙=2∣x∣\dot{x}=2\sqrt{|x|} with x(0)=0x(0)=0. Besides x(t)≡0x(t)\equiv 0, every xτ(t)=(max⁡{t−τ,0})2x_\tau(t)=(\max\{t-\tau,0\})^2 with τ≥0\tau\geq 0 is a solution. It stays at zero until τ\tau and then departs. For x>0x>0, the vector-field derivative is f′(x)=1/xf'(x)=1/\sqrt{x}, which becomes unbounded as x↓0x\downarrow 0.

  • Connection to numerical divergence. When fitting trajectories that separate sharply after impact, CHyLL may learn large vector-field derivatives near the guard to separate nearby latent states. These derivatives can amplify integration errors during rollout. This is a plausible mechanism for the observed numerical failures.

4. Core components

Equation (14) has four distribution-matching terms in two pairs. Additionally, we have latent dynamics and state space dynamiics. As our system naturally induces these terms, we introduce them at first and then discuss if they reallty work.

  • Per-step losses. It is common to compare predicted and observed states at each observation time. LxL_x and LzL_z extend this to SDE, whose evolution is described by the Fokker–Planck equations in Sections 3.1 and 4.2.

  • Path losses. The conditional loss matches trajectories from a given initial state, while the unconditional loss matches trajectories across initial states. These terms address a feature of SDE rollouts: paths from the same initial state can differ, and their temporal dependence is not determined by the per-step marginals alone.

We will present these two pairs together in Section 5 and identify latent distribution matching and path supervision as the main contributors. Our ablations indicate that strong performance requires LzL_z and at least one path loss (conditional or unconditional), while LxL_x is optional.

Reviewer_yoyGOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The central idea of the paper is to represent stochastic dynamics with discontinuous jumps as continuous stochastic dynamics in a higher-dimensional latent space. The authors consider stochastic hybrid systems (SHSs), where the continuous evolution is governed by a stochastic differential equation (SDE), while discrete transitions are triggered when the state reaches a guard surface and are governed by a Markov reset kernel. In other words, the system evolves continuously until it reaches a guard surface, at which point its state undergoes a stochastic reset before continuing its evolution.

The work extends the geometric framework developed in [1, 2] for deterministic hybrid systems (HSs). The main challenge is that stochastic resets cannot be directly handled by the existing embedding theory, which assumes deterministic reset maps. To address this, the authors introduce an auxiliary variable vv representing the reset displacement, transforming the stochastic reset into a deterministic mapping in an augmented state space (A-SHS). This allows them to apply topological gluing and the Whitney embedding theorem to obtain a higher-dimensional representation with continuous sample paths. They further show that the corresponding Fokker–Planck equation no longer requires explicit source and sink terms associated with resets.

Based on this theoretical construction, the authors propose a neural framework consisting of a stochastic encoder, a latent SDE, and a deterministic decoder. The model is trained using distribution-matching objectives for states and trajectories, without requiring explicit identification of guard surfaces, trajectory segmentation, or mode labels.

Experimental validation is performed on 11 synthetic benchmark settings involving bouncing-ball systems, torus and Klein-bottle dynamics, and different stochastic reset distributions. The proposed method achieves lower conditional and unconditional trajectory-distribution errors than the considered baselines, including latent SDE approaches and the deterministic framework [2]. Ablation studies additionally investigate the effects of latent dimensionality, stochastic encoding, and individual loss components.

References

[1] Teng, S., Liu, H., Song, J., and Sreenath, K. "CHyLL: Learning Continuous Neural Representations of Hybrid Systems." Transactions on Machine Learning Research, 2026.

[2] Teng, S., Liu, H., and Sreenath, K. "Embedding Hybrid Systems into Continuous Latent Vector Fields." Proceedings of the 43rd International Conference on Machine Learning, 2026.

Critical strengths

1. Conceptually simple and well-motivated approach. The paper presents an elegant idea of augmenting the state space to transform stochastic resets into deterministic ones, making it possible to represent discontinuous stochastic hybrid systems through continuous latent dynamics. This is a natural extension of the geometric framework developed in [1, 2] and provides a clear theoretical motivation for the proposed method.

2. Strong experimental results. The proposed method consistently outperforms the considered baselines on 11 synthetic benchmark settings, particularly in long-horizon prediction of stochastic trajectories. The experiments also include ablation studies that investigate the importance of latent dimensionality, stochastic encoding, and individual loss components. These results provide empirical support for the effectiveness of the proposed framework.

References

[1] Teng, S., Liu, H., Song, J., and Sreenath, K. "CHyLL: Learning Continuous Neural Representations of Hybrid Systems." Transactions on Machine Learning Research, 2026.

[2] Teng, S., Liu, H., and Sreenath, K. "Embedding Hybrid Systems into Continuous Latent Vector Fields." Proceedings of the 43rd International Conference on Machine Learning, 2026.

Critical weaknesses

1. Experimental validation of the general stochastic setting. Although the theoretical framework considers stochasticity in both continuous dynamics and reset mechanisms, all 11 learning benchmarks use deterministic continuous dynamics with stochastic resets. It would be valuable to include experiments where both components are stochastic to better demonstrate the generality of the proposed approach.

2. Practical considerations regarding training. The method requires simulating latent SDE trajectories during training, which may introduce computational overhead, particularly for longer time horizons. Additionally, the objective in Eq. (14) involves several loss terms with separate weighting coefficients. While the authors provide ablations of individual loss components, further discussion of computational efficiency and sensitivity to the loss weights would help clarify the practical applicability of the method.

Minor Remarks

  • Line 62: Typo: "data vy a latent SDE" should be "data by a latent SDE."
  • Theorem 1 and Section 4.2: The concept of a quotient manifold is introduced in Theorem 1, whereas the explanatory footnote on the quotient map appears later in Section 4.2. Moving this explanation closer to Theorem 1 could improve readability.
Questions for authors

Overall, I liked the paper and found its main contribution interesting and well-motivated. The following questions are relatively minor and are mainly intended to clarify some practical aspects of the proposed method rather than challenge its central contribution.

1. Sensitivity to loss weighting. The training objective in Eq. (14) contains several loss terms with separate weighting coefficients. How sensitive is the method to the choice of these coefficients? Is careful tuning required for each system, or do the same weights work reliably across different benchmarks?

2. Computational efficiency. The proposed method requires simulating latent SDE trajectories during training. How computationally expensive is this procedure compared with the considered baselines? It would be helpful to understand how the computational cost scales with trajectory length and latent dimensionality.

3. Validation with stochastic continuous dynamics. All 11 learning benchmarks use deterministic continuous dynamics with stochastic resets. Have the authors explored systems where both the continuous dynamics and reset mechanisms are stochastic? Are there any additional challenges expected in this more general setting?

4. Real-world applicability. The authors motivate stochastic hybrid systems through applications in robotics and autonomous driving, but the experiments focus on synthetic benchmarks. Have the authors considered evaluating the method on real-world trajectory data? If so, what practical challenges might arise?

3 comments

Thank you for the positive assessment and constructive suggestions!

1. Loss weighting and training stability (Weakness 2; Question 1)

Across all experiments, we fixed the weights of the xx-trajectory distribution loss and the xx- and zz-distribution losses to 1.01.0, without system-specific loss-weight tuning. We encountered no tuning difficulties and consistently observed stable training. Our theoretical construction motivates both the continuous latent representation and distribution matching in the observed and latent spaces. Empirically, this formulation enables stable latent-space rollouts during training and accurate distribution matching.

2. Computational efficiency (Weakness 2; Question 2)

Training uses windows of 32 observations with ten conditional samples, requiring 31 Euler–Maruyama steps per rollout. Longer-horizon evaluations therefore do not require equally long training windows.

For a fixed architecture, rollout cost grows linearly with the number of integration steps and sampled trajectories. Increasing latent dimensionality increases the drift and diffusion network evaluation costs; for Gθ∈Rdz×dwG_\theta\in\mathbb{R}^{d_z\times d_w}, the diffusion–noise multiplication additionally costs O(dzdw)O(d_zd_w) per step. Encoder/decoder evaluations and distributional-loss calculations contribute further overhead.

Under its stated assumptions, our theorem provides a sufficient latent dimension of dz=4nd_z=4n, so the latent dimension scales linearly with the original system dimension. The computational cost also depends on the network architecture and diffusion parameterization: when both dzd_z and dwd_w scale proportionally with nn, a dense diffusion–noise multiplication costs O(n2)O(n^2) per step.

3. Stochastic continuous dynamics and stochastic resets (Weakness 1; Question 3)

To directly address this concern, we conducted additional learning experiments by extending the 1D bouncing ball to stochastic continuous-time dynamics:

dv_t=-9.81\ dt+0.05dW_t,$$ while retaining the original GMM and uniform reset kernels. These experiments therefore include stochasticity in both the continuous dynamics and the resets. All six methods were trained from scratch as in the paper. Here we show the conditional (unconditional) path loss: | Case | Ours | CHyLL | Latent SDE | Latent SDE⋆ | SDE Matching | NSF | |---|---:|---:|---:|---:|---:|---:| | b-ball SDE (GMM) | **0.309 (0.415)** | 0.504 (1.05) | 0.371 (1.14) | 0.361 (0.997) | 1.96 (22.9) | 0.614 (4.47) | | b-ball SDE (Uniform) | **0.357 (0.587)** | 0.534 (0.926) | 0.531 (2.72) | 0.498 (2.29) | 1.48 (16.2) | 0.573 (3.47) | Here, Latent SDE⋆ denotes Latent SDE augmented with the observed-state distribution loss and the unconditional trajectory-distribution loss. Under this matched training budget, our method achieves the lowest mean loss for both metrics in both settings. These results provide additional empirical support for learning systems with stochastic continuous dynamics and stochastic resets. The original benchmarks **isolate** randomness in the resets to highlight its effect. The additional experiments show that our framework also accommodates continuous process noise. The central geometric challenge (**topological obstruction of stochastic reset**) addressed by our work remains the representation of stochastic reset branches through continuous latent dynamics. ### 4. Real-world applicability (Question 4) Our primary contribution is to extend the topological framework for deterministic hybrid systems to the stochastic setting, a nontrivial step requiring an augmented-state construction. We therefore focus on synthetic benchmarks that allow us to control the topological structure and validate this theoretical development. To demonstrate relevance to robotics, we add an experiment on simultaneous frictional impacts of rigid bodies, using the Phone-drop and Box–wall models. These examples require the learned dynamics to capture multimodal trajectory distributions arising from multiple contacts occurring at nearly the same time. Our method achieves the lowest mean conditional and unconditional path distribution losses in both cases. | Case | Ours | CHyLL | Latent SDE | Latent SDE⋆ | SDE Matching | NSF | |---|---:|---:|---:|---:|---:|---:| | Phone drop | **0.561 (0.758)** | 0.843 (9.46) | 0.635 (4.08) | 0.570 (0.885) | 0.658 (5.23) | 1.09 (25.8) | | Box–wall impact | **0.220 (0.488)** | 0.322 (1.96) | 0.245 (0.756) | 0.227 (0.516) | 0.331 (4.37) | Diverged† | Real-data validation remains future work, as the dataset benchmarking this task is lacking.

Thank you to the authors for their detailed and constructive response. The clarifications and additional experiments have satisfactorily addressed my concerns, particularly regarding training stability, computational efficiency, and the ability to handle stochasticity in both continuous dynamics and resets.

I find the theoretical contribution elegant and well-developed, and the new experimental results further strengthen the paper. While validation on larger-scale and real-world systems remains an important direction for future work, I do not consider its absence a significant limitation at this stage.

I believe the paper makes a valuable contribution that deserves acceptance. Making these ideas available to the broader research community could encourage further development, practical implementations, and new applications. I therefore maintain my positive recommendation for acceptance and raise the score.

We thank the reviewer for the suggestions and positive evaluation!

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.