MOSS: Mode-Oriented Source Shaping for Robot Flow Matching Policies

arXiv:2610.22193 · cs.RO · Submitted 2026-10-15 · 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: I'm Rosa.

Dev: And I'm Dev, here with Taro.

Rosa: Today's paper: "MOSS: Mode-Oriented Source Shaping for Robot Flow Matching Policies".

Dev: The gist: MOSS proposes ModeOriented Source Shaping,

Rosa: First, who's behind it and why it matters.

Title and authors: Rosa: We're moving into the summary section now, where they explain the core mechanism again, focusing on that idea of deterministic flow transport inducing mode-dependent structure in the source space <ref:2610.22193#pg6>.

Dev: They point out that under ideal deterministic ODE transport, distinct behavioral modes have distinct source-space preimages <ref:2610.22193#pg6>. But practical approximation errors can blur those boundaries due to imbalanced data, limited model capacity, and few inference steps <ref:2610.22193#pg6>.

Taro: So they're saying that even if the ideal structure exists in theory, it doesn't always show up clearly when we use finite data or limited compute?

Rosa: Exactly. This realization leads them to deliberately imposing a simple, semantically aligned source geometry to make mode separation explicit and controllable <ref:2610.22193#pg6>.

Dev: They observed empirically that standard flow matching on multimodal tasks already has a structured source space: initial noises leading to different modes tend to occupy distinct regions of the Gaussian source, even though they get entangled near their interface <ref:2610.22193#pg6>.

Taro: So they are saying the structure isn't completely lost, it's just hidden underneath the standard isotropic Gaussian noise <ref:2610.22193#pg6>.

Rosa: That’s what Theorem one shows, which proves that up to a null boundary, the Gaussian source is already partitioned into semantic noise zones, each carrying exactly the probability mass of its behavioral mode <ref:2610.22193#pg6>.

Dev: Theorem two goes deeper because it shows that ideal flow matching induces not only distinct semantic noise zones but also locally stable ones <ref:2610.22193#pg6>.

Taro: What does locally stable mean in this context, Rosa? Is it about the regions being less likely to change during the process?

Rosa: Under standard regularity conditions on the flow velocity field, Theorem two proves that behaviors lying farther from the semantic boundary correspond to a larger same-mode neighborhood in the source space <ref:2610.22193#pg6>.

Dev: So, if you're far enough away from a boundary, you have more room for that specific mode to exist in the source distribution <ref:2610.22193#pg6>.

Taro: That connects back to the idea of persistence I mentioned earlier—the strength qeff controlling the separation <ref:2610.22193#pg6>.

Rosa: It does, because that effective strength qeff prevents that separation from decaying before the behavior is completed <ref:2610.22193#pg6>.

Dev: So, the paper is arguing that we can impose a simple geometry because the underlying deterministic transport actually creates this structure in source space <ref:2610.22193#pg6>.

Taro: It's like they are using the known physics of deterministic transport to justify putting these semantic zones onto the noise <ref:2610.22193#pg6>.

Rosa: That’s a good way to put it. And this leads directly into how MOSS imposes that structure explicitly rather than just hoping it emerges <ref:2610.22193#pg6>.

The paper's summary: Dev: Now let's talk about the specific improvements MOSS suggests over previous methods and what those mean for practical use, Rosa.

Rosa: The main improvement is that MOSS turns mode selection into an explicit choice of where generation starts <ref:2610.22193#pg6>. This is much more robust against issues like imbalanced mode distributions or limited model capacity that cause standard methods to suffer from mode preference or even mode collapse <ref:2610.22193#pg6>.

Taro: So instead of the policy having to guess which mode it's in, the initial noise is already set up to point toward a specific behavior?

Dev: Exactly. It allows for steering behavior directly through the initial distribution, which is more robust against those problems <ref:2610.22193#pg6>.

Rosa: They also show that adaptive source geometry improves guidance rate over fixed Gaussian separation by twelve to fourteen percentage points in average guidance rate at the same training budget <ref:2610.22193#pg6>.

Dev: That's a tangible improvement because it means you get better results without having to train for much longer <ref:2610.22193#pg6>.

Taro: And they manage to leverage shared trajectory data from other modes while preserving clear mode boundaries, which leads to strong minority-mode sample efficiency and high guidance rate with relatively few target-mode sample exposures <ref:2610.22193#pg6>.

Rosa: So it uses the majority data structures to learn more stable common motion patterns while keeping the mode boundaries sharp <ref:2610.22193#pg6>.

Dev: That's a clever way to balance learning common structure with preserving distinct behavior <ref:2610.22193#pg6>.

Taro: I wonder about the robustness against inference-time errors, because I think that's where things usually break down in real-world deployment <ref:2610.22193#pg6>.

Rosa: MOSS ensures that mode-specific sources still initialize transport from regions already established during training <ref:2610.22193#pg6>.

Dev: That could potentially reduce trajectory-quality degradation caused by inference-time guidance, and it extends naturally to currently emerging one-step denoising algorithms that don't support guidance <ref:2610.22193#pg6>.

The paper's improvements: Rosa: We're wrapping up with the conclusion of this discussion about MOSS: Mode-Oriented Source Shaping for Robot Flow Matching Policies. They summarize that this structure makes the choice of where generation starts explicit and adjust the separation based on local multimodality <ref:2610.22193#pg6>.

Dev: This results in strong mode controllability and task success rate across simulated and real-world manipulation tasks, even under highly imbalanced mode distributions <ref:2610.22193#pg6>.

Taro: I think the key is that they made source-space geometry a very effective design dimension for controllable generative robot policies <ref:2610.22193#pg6>.

Rosa: It shows that we can achieve strong mode controllability and task success rate, and it extends naturally to multiple behavioral modes <ref:2610.22193#pg6>.

Dev: So, the paper is highlighting how source-space geometry is a design dimension for these kinds of generative robot policies <ref:2610.22193#pg6>.

Taro: I just think that this approach provides a way to handle mode selection explicitly rather than relying on implicit random sampling <ref:2610.22193#pg6>.

Rosa: It’s a solid piece of work, showing how to achieve mode controllability and task success rate through explicit source shaping with MOSS <ref:2610.22193#pg6>.

Conclusion: Rosa: So MOSS is essentially taking mode selection and turning it into an explicit choice of where generation starts by shaping the source distribution itself <ref:2610.22193#pg6>.

Dev: That means we're not just picking a mode at the end; we're conditioning the initial noise so that generation begins in a specific area of the source space, which is much more robust against imbalanced data <ref:2610.22193#pg6>.

Taro: I mean, that explicit choice really tackles those problems where standard flow matching methods just get confused by noisy or unstable neighbors <ref:2610.22193#pg6>.

Rosa: Right, and they do this by using a modality-strength estimator to adaptively control the geometry of these mode-specific sources <ref:2610.22193#pg6>.

Dev: It's interesting how it contracts components toward shared Gaussian when behaviors are similar but separates them sharply right at those behavioral branching points <ref:2610.22193#pg6>.

Taro: And that local strength estimation, using the difference between within-mode and between-mode variations, helps suppress spurious signals from noisy neighbors <ref:2610.22193#pg6>.

Rosa: They normalize this strength into a calibrated value q in the range of zero to one and even apply temporal persistence during rollout to get an effective strength qeff <ref:2610.22193#pg6>.

Dev: So, that means when you deploy it, you just sample from the corresponding component based on that qeff value for mode k <ref:2610.22193#pg6>. That sounds like a very clean way to control multimodal generation.

Taro: The evaluation results are pretty compelling too, showing up to one hundred percent target-mode steering even when the modes are heavily imbalanced <ref:2610.22193#pg6>.

Rosa: They also show that on real-world tasks, MOSS outperforms mode-conditioned flow policy by forty-three point three percent in average guidance rate <ref:2610.22193#pg6>.

Dev: That level of improvement over existing methods is significant, especially since it happens without changing the training budget much <ref:2610.22193#pg6>.

Taro: It’s cool that it even reduces the Dmu and total squared transport during offline analysis, which suggests the source alignment actually makes sense geometrically <ref:2610.22193#pg6>.

Rosa: The limitation they mention is that right now it assumes a fixed set of discrete behavioral modes specified before training, so it doesn't directly support open-ended mode control <ref:2610.22193#pg6>.

Dev: And they also flag that while the source formulation is compatible with larger generative policies, their performance in those massive settings still needs evaluation <ref:2610.22193#pg6>.

Taro: I just think MOSS is a good demonstration of how to make the underlying structure of flow matching explicit through source space geometry <ref:2610.22193#pg6>.

Rosa: Exactly, and it shows that this structure can be controlled adaptively based on local multimodality <ref:2610.22193#pg6>.

Dev: Anyway, MOSS is a solid piece of work for getting explicit control in flow matching policies. Next up, we'll look at how other agentic systems are handling mission-state governance.

Zirun Ma, Jindou Jia, Tuo An, Jiaqi Bai, Jiahua Dong, Jianfei Yang

Nanyang Technological University · Tsinghua University

cs.RO

Submitted: 2026-10-15

Updated: 2026-10-15

Project page: https://mmmmazr.github.io/moss

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

The gist: The gist: MOSS proposes ModeOriented Source Shaping, a method that turns mode selection into an explicit choice of where generation starts by dynamically shaping the source distribution to encode

Key concepts

Flow Matching: Flow Matching is a generative model technique that learns a continuous transport path from a simple starting distribution (source) to the complex data distribution of actions. In robotics, it helps generate robot movements by learning how to smoothly transition from noise to desired control signals.
Source Space Structure: The paper investigates how deterministic flow transport creates specific structures in the initial noise space. Ideally, different behavioral modes should map to distinct regions in this source space. Practical training often blurs these boundaries due to data limitations and model errors, which MOSS aims to overcome.
Mode-Oriented Source Shaping (MOSS): MOSS is the core method that explicitly designs the source distribution. It creates mode-specific Gaussian sources where each mode's identity is encoded directly into the initial noise. A modality strength estimator then adaptively adjusts how these sources are separated or contracted based on local behavior.
Modality Strength Estimator: This estimator predicts how multimodal a specific part of a robot's trajectory is locally. It compares variations within and between modes to gauge the degree of multimodality, helping MOSS decide whether to contract modes together (if they behave similarly) or separate them (if they branch into different behaviors).

Terminology

Flow Matching — 1× in this paper · explained in 17 episodes

Summary

The gist: MOSS proposes ModeOriented Source Shaping, a method that turns mode selection into an explicit choice of where generation starts by dynamically shaping the source distribution to encode mode identity in initial noise, allowing for controllable multimodal generation in robot flow matching policies.

How it works

MOSS constructs semantically aligned mode-specific Gaussian sources that explicitly encode mode identity into the initial noise <ref:2610.22193#pg6>. A modality-strength estimator predicts the local degree of multimodality, which adaptively controls the geometry of these mode-specific sources <ref:2610.22193#pg6>. When different modes in the training demonstrations behave alike, MOSS contracts the components toward a shared Gaussian and separates them near behavioral branching points <ref:2610.22193#pg6>. During deployment, the desired behavior is selected at inference simply by sampling from the corresponding component <ref:2610.22193#pg6>.

Modality Strength Estimation

The degree of behavioral multimodality can vary substantially along a robot trajectory <ref:2610.22193#pg6>. The modality strength estimator is trained to predict this local degree from previous observation and action history <ref:2610.22193#pg6>. It computes a supervision signal where the between-mode variation VarB,i and within-mode variation VarW,i are compared <ref:2610.22193#pg6>. A large within-mode variation suppresses spurious multimodality signals caused by noisy or unstable neighbors <ref:2610.22193#pg6>. The resulting strength is normalized using task-wise percentile calibration to yield a calibrated strength q in [0, 1] <ref:2610.22193#pg6>. Temporal persistence is applied during rollout to obtain an effective strength qeff, preventing separation from decaying before the behavior is completed <ref:2610.22193#pg6>.

Adaptive Mode-Oriented Source Shaping

Given the estimated modality strength q, MOSS adaptively separates the source distributions associated with different behavioral modes <ref:2610.22193#pg6>. It defines the mode-discriminative subspace using a multimodality supervision signal ri to identify training states where future behaviors diverge <ref:2610.22193#pg6>. For a requested mode k, MOSS defines its source center as ck(q) = α(q)a˜k, and samples the initial condition from p(x0 k, qeff) <ref:2610.22193#pg6>. The modality strength q controls both the displacement of the source centers and their overlap, where α(q) and ρ(q) increase monotonically with q <ref:2610.22193#pg6>.

Evaluation and Results

MOSS achieves up to 100% target-mode steering while maintaining high task success rate (SR) across simulated and real-world manipulation tasks <ref:2610.22193#pg6>. Even under a highly imbalanced mode distribution (10: 90), the method can reliably recover the minority mode, achieving over 90% guidance rate (GR) <ref:2610.22193#pg6>. On real-world tasks, MOSS achieves an average GR/SR of 84.2%/85.3%, outperforming mode-conditioned FM by 43.3 and 12.0 percentage points in average GR and SR, respectively <ref:2610.22193#pg6>. The adaptive source shaping further improves guidance rates over fixed Gaussian separation by 12-14 percentage points in average GR at the same training budget <ref:2610.22193#pg6>.

Robustness and Efficiency

MOSS exhibits strong minority-mode sample efficiency, reaching high GR with relatively few target-mode sample exposures <ref:2610.22193#pg6>. The advantage is particularly pronounced on Push Cube, where MOSS approaches its final performance substantially earlier than both baselines <ref:2610.22193#pg6>. Across the three tasks, adaptive MOSS reaches consistently high guidance rates for both requested modes and remains stable at later checkpoints <ref:2610.22193#pg6>. Offline analysis shows that adaptive shaping reduces Dµ and total squared transport in all three tasks <ref:2610.22193#pg6>.

Limitations

MOSS has two main limitations; first, the current formulation assumes a fixed set of discrete behavioral modes specified before training, and therefore does not directly support open-ended mode control or continuous behavioral variation <ref:2610.22193#pg6>. Second, although the source-space formulation is in principle compatible with larger generative robot policies, its performance in such large-scale settings remains to be evaluated <ref:2610.22193#pg6>.

Conclusion

MOSS makes this structure explicit by associating semantic behavioral modes with mode-specific source distributions and adaptively adjusting their separation according to local multimodality, without conditioning the shared velocity field on the requested mode <ref:2610.22193#pg6>. Across simulated and real-world manipulation tasks, MOSS achieves strong mode controllability and task success rate, extends naturally to multiple behavioral modes, and improves minority-mode learning under highly imbalanced demonstrations <ref:2610.22193#pg6>. Together, these results highlight source-space geometry as an effective design dimension for controllable generative robot policies <ref:2610.22193#pg6>.

How it works

The paper reveals that deterministic flow transport induces mode-dependent structure in the source space <ref:2610.22193#pg6>. Under ideal deterministic ODE transport, distinct behavioral modes have distinct source-space preimages, while practical approximation errors can blur their boundaries due to imbalanced data, limited model capacity, and few inference steps <ref:2610.22193#pg6>. This leads to the realization that the existence of an ideal mode partition does not imply that sufficiently clear, stable, and controllable source regions emerge under finite data, limited model capacity, and finite-step inference <ref:2610.22193#pg6>. MOSS deliberately imposes a simple, semantically aligned source geometry to make mode separation explicit and controllable <ref:2610.22193#pg6>.

Multimodal Property of Flow Matching

Standard FM learns a continuous transport from a simple source to a complex action distribution conditioned on an observation o ∈ O <ref:2610.22193#pg6>. Empirically, we observe that FM policies trained on multimodal tasks already exhibit a structured source space: initial noises leading to different behavioral modes tend to occupy distinct regions of the Gaussian source, albeit noticeably entangled near their interface <ref:2610.22193#pg6>. Theorem 1 shows that up to a null boundary, the Gaussian source is already partitioned into semantic noise zones, each carrying exactly the probability mass of its behavioral mode <ref:2610.22193#pg6>.

Noise Zoning and Its Geometric Regularity

Theorem 2 shows that ideal FM induces not only distinct semantic noise zones, but also locally stable ones <ref:2610.22193#pg6>. Under standard regularity conditions on the FM velocity field, the induced ODE flow is a bi-Lipschitz homeomorphism <ref:2610.22193#pg6>. Theorem 2 proves that behaviors lying farther from the semantic boundary correspond to a larger same-mode neighborhood in the source space <ref:2610.22193#pg6>.

Practical FM Blurs the Noise Zoning

In practice, finite demonstrations, imperfect optimization, and finite-step ODE integration make the realized endpoint map Φeo: R d → R d deviate from Φo <ref:2610.22193#pg6>. Corollary 4 shows that source points with γo(z) = O(ηo(z)) form a boundary-sensitive region in which mode confusion is likely to occur <ref:2610.22193#pg6>. This exposes a fundamental limitation of leaving noise zoning to be learned implicitly from a standard Gaussian source <ref:2610.22193#pg6>.

Source Alignment and Transport Cost

Adaptive shaping reduces Dµ and total squared transport in all three tasks <ref:2610.22193#pg6>. The difference is most visible in D∥, particularly on Push Cube and Stack Cube <ref:2610.22193#pg6>. This is expected because MOSS modifies the source primarily along the mode-discriminative direction, while the much larger orthogonal transport component is largely shared by both variants <ref:2610.22193#pg6>.

Conclusion

Source-space geometry as an effective design dimension for controllable generative robot policies has been demonstrated through MOSS <ref:2610.22193#pg6>. The paper concludes that MOSS makes the structure explicit by associating semantic behavioral modes with mode-specific source distributions and adaptively adjusting their separation according to local multimodality <ref:2610.22193#pg6>.

Improvements for AI systems

Improve mode selection reliability by explicitly controlling it through source space manipulation rather than relying on implicit random sampling; MOSS turns mode selection into an explicit choice of where generation starts. This allows for steering behavior directly through the initial distribution, which is more robust against imbalanced mode distributions, limited model capacity, and limited inference steps that cause standard methods to suffer from mode preference or even mode collapse.
Enhance minority-mode learning efficiency by leveraging shared trajectory data while preserving distinct boundaries; MOSS can effectively leverage shared trajectory data from other modes, allowing the policy to learn more stable common motion structures while preserving clear mode boundaries, leading to strong minority-mode sample efficiency and high GR with relatively few target-mode sample exposures.
Increase task success rate and training stability by dynamically adjusting source geometry based on local multimodality; MOSS uses a modality strength estimator to adaptively control the geometry of the mode-specific sources, contracting them where behaviors are similar while separating them near behavioral branching points, which improves performance over fixed separation methods, as shown by MOSS achieving surpasses mode-conditioned flow policy by 43.3 points of guidance rate on real robots.
Improve robustness against inference-time errors by initializing transport from already established regions; unlike methods that apply strong guidance at inference, MOSS ensures that mode-specific sources still initialize transport from regions already established during training, which can potentially reduce trajectory-quality degradation caused by inference-time guidance and extends naturally to one-step denoising algorithms.
Improve data efficiency under highly imbalanced demonstrations by reusing majority mode structures; MOSS benefits from the majority data, as shown in its ability to observe the same number of target-mode examples as other methods while exploiting motion shared with the majority mode, suggesting that the shared velocity field can reuse motion structure learned from the majority mode rather than learning the minority behavior entirely from scratch.