Test-Time Improvement of VLA Policies via Selected Diffusion Noise for Spurious-Robust Action Smoothing
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: "Test-Time Improvement of VLA Policies via Selected Diffusion Noise for Spurious-Robust Action Smoothing".
Dev: Here's the gist of the paper. Diffusion-based Vision-Language-Action (VLA) policies enable strong generalization in robotic manipulation, but remain sensitive to spurious visual correlations and noisy action generation, leading to brittle behavior under perturbations <ref:2606.14084#pg6>.
Rosa: The paper introduces SELECTED DIFFUSION NOISE (SDN), a simple, training-free test-time method that improves both robustness and success rate by leveraging the diffusion noise space as a controllable degree of freedom <ref:2606.14084#pg3>.
Dev: SDN dynamically samples noise vectors that are maximally separated from a reference set to mitigate reliance on spurious cues, while selecting candidates that yield more coherent action trajectories <ref:2606.14084#pg4>.
Rosa: Problem Statement Current VLA policies often remain fragile in out-of-distribution settings because they tend to overfit to spurious correlations and non-essential behaviors in human demonstrations <ref:2606.14084#pg2>.
Title and authors: Rosa: So we're looking at this paper called "Test-Time Improvement of VLA Policies via Selected Diffusion Noise for Spurious-Robust Action Smoothing." It's about fixing the brittle behavior we see in these vision language action policies when they encounter things that aren't in their training data.
Dev: Right, Rosa, it tackles that exact problem—the sensitivity to spurious visual correlations and noisy action generation that makes these robotic systems so fragile under small changes. The core idea here is introducing something called Selected Diffusion Noise, or SDN, which is a simple method you can use right at test time without needing any extra training.
Taro: I'm curious about the "test-time" part. When you're dealing with diffusion models, that usually implies some sort of generation process happening during inference, so how does adding this noise selection change that?
Rosa: Well, SDN treats the initial diffusion noise itself as a controllable degree of freedom. Instead of just picking from a standard set of noisy actions, SDN dynamically samples those noise vectors that are maximally separated from what the model usually relies on to mitigate reliance on those spurious cues.
Dev: That separation is key—it’s about avoiding the unreliable parts of the noise space while simultaneously selecting candidates that result in more coherent action trajectories. It's a dual approach, focusing both on filtering out bad visual inputs and ensuring physical smoothness in the resulting motion.
Taro: So, if we think about what's failing right now—where these policies hallucinate things like assuming an object is already grasped or selecting jerky movements just because it looks plausible semantically—how does SDN specifically address that failure mode?
Rosa: It does it in two stages. First, there's the Grounding Filter. This stage uses a contrastive evaluation to identify "grounded" initial noise seeds by comparing them against a set of failed behaviors where the target object is masked, which pushes those bad candidates away while pulling the good ones toward what actually matches the observation and instruction.
Dev: That grounding score, R ground equals d (i) B minus d (i) G, is essentially measuring how far an action chunk is from what's hallucinatory versus what makes sense for the task, and you keep the top candidates after that stage.
Taro: And then there's the second stage. After you narrow it down to those grounded candidates, you run a kinematic stability refinement using something called the JerkRMS metric to select the final action chunk that minimizes jerk instability. That’s trying to make sure the movement is physically smooth, not just logically correct.
Rosa: Exactly. The paper shows that this sequential filtering process—grounding first, then smoothness—leads to success rate gains of eight percent in simulation and ten percent in real-world settings, while producing smoother and more stable actions overall.
Title and authors: Dev: I'm interested in the efficiency aspect here because these methods often add complexity. SDN is presented as a plug-and-play enhancement for architectures like pi0 and GR00T N1 point 6 without needing any auxiliary networks or parameter updates, which means it operates entirely self-contained and training-free at test time <ref:2606.14084#pg1>.
Taro: That’s a big deal if you think about deployment. Being able to improve the policy right when you run it, without having to fine-tune a multi-billion parameter model or risk catastrophic forgetting from retraining, that makes sense for real robotics where you don't have infinite compute power.
Rosa: It really changes how we think about improving these policies—it moves the improvement mechanism from a slow, expensive training process to something immediate and autonomous during inference. It shows that the diffusion noise space can genuinely be treated as a steerable sampling space instead of just a black box generator.
Dev: And this approach is contrasted against other methods like Classifier-Free Guidance or Policy Contrastive Decoding because SDN avoids the expensive per-step image inpainting needed by those, and it actually runs about one point five three times faster than Policy Contrastive Decoding while achieving a higher success rate.
Taro: But I wonder what the limitations are here when you take this outside of a perfectly controlled simulation environment. The paper mentions that SDN might suffer from corruption specificity because object-centric visual shortcuts might fail to capture hallucinations caused by things like dynamic lighting or texture shifts in the real world.
Rosa: That's a fair point, Taro. And they also noted that the method currently relies on a fixed number of noise candidates at inference time, which limits how much it can adapt if the task complexity or uncertainty changes rapidly during operation. Future work needs to look into adaptive candidate allocation and richer multimodal corruption modeling to improve this robustness further.
Dev: So, while SDN shows strong gains across different diffusion-based backbones like pi0 and GR00T, the authors acknowledge that relying too heavily on visual shortcuts can lead to catastrophic performance collapse if the target object is masked in certain ways. That confirms their suspicion that these models are still heavily reliant on those ungrounded visual cues.
Taro: It sounds like they've identified the core tension: you need strong grounding, but you also need physical feasibility, and SDN tries to solve both sequentially by using the noise space as a filter for both.
Rosa: That’s the essence of it—it's not just about picking a better output; it’s about structuring how we select that output by enforcing both visual consistency and kinematic coherence. It provides a cleaner pathway for grounding VLA actions in real-time sensorimotor streams, which is exactly what we need to move toward reliable deployment.
Title and authors: Dev: So, to wrap up this discussion on "Test-Time Improvement of VLA Policies via Selected Diffusion Noise for Spurious-Robust Action Smoothing," the paper shows that leveraging the diffusion noise space as a controllable degree of freedom through its dual filtering process—grounding followed by smoothness—can provide consistent success rate gains across simulation and real-world settings.
Taro: I just want to emphasize that this mechanism demonstrates how we can actively steer the policy's output distribution during inference, rather than just passively sampling from it. That active steering is what gives it the advantage over standard best-of-N sampling methods <ref:2606.14084#pg2>.
Rosa: It really does offer a cleaner and more efficient pathway for improving these policies in real time, validated across different diffusion backbones like pi0 and GR00T on benchmarks such as Google Robot and even some real-world tasks with the Aloha robot.
Dev: Indeed, it's a plug-and-play enhancement that doesn't require retraining or auxiliary networks, making immediate test-time improvement feasible for many applications in robotics.
Taro: It’s interesting how they managed to show significant improvements like tripling the success rate on the "Open Drawer" task, going from four point zero percent up to fourteen point zero percent on the Google Robot platform.
Rosa: That jump is substantial, and it really shows that this approach can be a general mechanism for enhancing VLA policies at test time across diverse setups. It gives us a concrete tool to enhance robustness without needing massive retraining cycles.
Dev: So we've seen how SDN uses contrastive grounding and kinematic filtering to mitigate hallucinations and jerkiness simultaneously, providing measurable improvements in both success rate and action quality when compared to baselines like Classifier-Free Guidance or PCD.
Taro: And while the gains are impressive, we have to keep in mind that the method stops working well when visual shortcuts fail due to dynamic lighting or texture shifts, which is a real-world challenge we have to solve before this becomes universally reliable.
Rosa: So, "Test-Time Improvement of VLA Policies via Selected Diffusion Noise for Spurious-Robust Action Smoothing" gives us a concrete framework—SDN—to dynamically select noise vectors that are maximally separated from unreliable action sets while ensuring the resulting actions are kinematically smooth and grounded in the task.
Dev: It’s a solid improvement over existing test-time methods because it integrates both visual grounding and physical feasibility into a single, sequential filtering process, which is what makes it stand out.
Taro: I just think that for anyone building these systems, understanding how the noise space can be steered like this is crucial for making them reliable when they're actually operating in the wild.
Rosa: It’s definitely a tool that warrants attention because it shows that diffusion noise selection can play an effective and general role in enhancing VLA policies right when they need it most.
The paper's summary: Rosa: So, we're looking at this paper called "Test-Time Improvement of VLA Policies via Selected Diffusion Noise for Spurious-Robust Action Smoothing." Basically, they found a way to make these vision language action policies much more stable when they're running outside the lab.
Dev: Right, it addresses that issue where these policies get brittle because they rely on visual cues that aren't actually important, leading to jerky or wrong movements under slight changes in the environment.
Rosa: They introduce something called Selected Diffusion Noise, or SDN. It treats the initial noise in the diffusion process as a controllable part of the system, instead of just letting it be random.
Dev: The idea is that instead of picking actions randomly from a pool, SDN dynamically samples noise vectors that are intentionally far away from what we know causes errors. That’s about avoiding those unreliable visual signals while also selecting candidates that lead to more coherent movement.
Rosa: They do this with a two-stage filter. First, there's the Grounding Filter where they use a contrastive method to separate actions that look like "hallucinations" from actions that actually match the task instruction and observation.
Dev: The score they use is basically comparing how much an action looks like it violates the object’s presence versus how much it follows what we see in the input data. They keep the top candidates after this stage.
Rosa: Then, in the second stage, they take those few good candidates and run a kinematic check using something called JerkRMS to make sure the resulting action is physically smooth and doesn't involve sudden jolts.
Dev: So, what does this mean for us on the ground? They report success rate gains of about eight percent in simulation and ten percent in real-world tests, and they also produce actions that are much smoother overall.
Rosa: It shows that this noise selection technique can be a general way to improve these policies at test time across different architectures, like pi0 or GR00T.
Dev: And what's interesting is the efficiency. SDN works as a plug-and-play enhancement for existing models without needing any new training or extra networks; it’s training-free and happens right when you need it.
Rosa: That means we can improve performance immediately in a real robotic setting without spending hours on retraining or worrying about catastrophic forgetting.
Dev: The authors point out that while the gains are good, SDN has limitations. They say it might struggle with corruption specificity because if there are weird things happening—like sudden changes in lighting or texture—those visual shortcuts might fail to catch the hallucinations.
Rosa: And they also flag that right now, it relies on a fixed number of noise candidates at test time, which means its ability to adapt when the task gets much more complex is limited.
Dev: So, while it’s a solid method for immediate improvement and getting better success rates, we still need to figure out how to make that candidate selection more adaptive for truly unpredictable real-world situations.
Rosa: It really highlights that steering the diffusion process itself can be a powerful mechanism for grounding these policies in real-time sensorimotor streams.
Dev: Next up, we're going to look at some other papers on how decentralized coordination handles airspace overload and how we can build safer navigation in crowded environments.
The paper's improvements: Taro: So, we've seen how SDN works—it filters noise based on visual grounding and then smooths the movement using jerk metrics. Now, what are the other improvements they suggest for this approach?
Rosa: They focus on making this method more self-diagnostic. The idea is to have the system check if a noise seed is actually causing hallucinations by comparing its output when it sees the original view versus when it sees an object masked out.
Dev: That’s like giving the AI a built-in sanity check. It asks, "Does this action candidate still rely on a visual shortcut that isn't there?" If the answer is yes, it steers away from that noise vector.
Rosa: That self-diagnostic steering is really clever because it prevents the model from getting stuck on spurious background cues, which we saw was a weakness before.
Dev: It unifies two different objectives into one selection strategy: you’re simultaneously maximizing separation from the unreliable action sets and minimizing that smoothness metric to eliminate jerky motions.
Taro: That dual-objective optimization is what makes it more robust than just picking the highest score or looking at smoothness alone, which is a pretty strong point.
Rosa: And as an enhancement, they make it plug-and-play across different vision language action backbones like pi0 and GR00T without needing any extra training or auxiliary networks.
Dev: That training-free nature is huge for deployment because you don't need to spend time fine-tuning a multi-billion parameter model just to get an inference boost.
Rosa: It’s about making test-time improvement immediate and accessible, which makes it much more practical for real robotic environments.
Taro: But they also admit a major limitation: the method relies on a fixed number of noise candidates during inference, which means its ability to adapt gracefully when the task gets really complex is limited.
Dev: Exactly. So while SDN gives us those big success rate bumps, we still have to figure out how to make that candidate selection more dynamic based on how uncertain the system is about the current state.
Rosa: That's exactly where future work needs to go—making that candidate allocation adaptive and using richer models of visual corruption so it doesn't fail when lighting or textures change unexpectedly.
Dev: So we’ve covered how SDN improves performance, what it does for deployment, and the specific limitations they identified. Next up, we’re looking at how decentralized systems handle overload in air traffic control.
Conclusion: Rosa: So, to wrap up, this paper on "Test-Time Improvement of VLA Policies via Selected Diffusion Noise for Spurious-Robust Action Smoothing" shows that we can dynamically steer the diffusion process to pick actions that are both visually grounded and physically smooth during operation.
Dev: It basically gives us a training-free way to improve robustness by filtering noise based on visual consistency and kinematic feasibility at inference time.
Taro: I think what this does for autonomy is showing how we can actively refine the policy's output distribution instead of just relying on a static sample, which is pretty important for real-world failure recovery.
Rosa: It validates the idea that treating diffusion noise as a controllable degree of freedom can be an effective mechanism to enhance these policies across different vision language action backbones.
Dev: The implication is that we have a cleaner pathway for grounding VLA actions in real-time sensorimotor streams without needing massive retraining cycles.
Taro: It moves the focus from just making the model *look* correct to actually ensuring it's doing something physically feasible and grounded in reality.
Rosa: We saw significant success rate gains, like tripling the success on a specific task, which means this isn't just theoretical; it’s hitting real-world benchmarks.
Dev: But we do have to remember their caveat: if there are complex visual shifts, like sudden changes in lighting or textures, that might break the grounding mechanism because it relies on those visual shortcuts.
Taro: So, while SDN is a big step forward for robustness, the authors flag that we still need better ways to handle those unpredictable real-world corruptions gracefully.
Rosa: Right. It’s a powerful tool for immediate test-time improvement, but it shows us exactly where the next generation of policy refinement needs to focus.
Dev: We’ll keep an eye on how this noise selection strategy integrates with other methods we're looking at, especially those dealing with decentralized coordination in complex systems.
Taro: Next time we talk about these challenges, we’ll look at how decentralized air traffic management systems try to handle overload and maintain safe separation.
Duc Minh Nguyen, Bao-Ngoc Dao, Tung M. Luu, Binh Gia Nguyen, Vinh Tong, Anji Liu, Vu N. Duong, Dung D. Le, Daniel Sonntag, Trung Le©Ngan Le¡ Jan Peters An Thai Le Minh Nhat Vu Mathias Niepert Khoa D. Doan Duy M. H. Nguyen Vien Anh Ngo
Center for AI Research, VinUniversity · VinRobotics · KAIST · University of Stuttgart · National University of Singapore · DFKI · University of Oldenburg · ©Monash University · ©TU Darmstadt
cs.RO
Submitted: 2026-06-12
Updated: 2026-11-04
Code: https://github.com/allenzren/open-pi-zero4https:
License: http://creativecommons.org/licenses/by/4.0/
Importance score: 99/100
The gist: Diffusion-based Vision-Language-Action (VLA) policies enable strong generalization in robotic manipulation, but remain sensitive to spurious visual correlations and noisy action generation…
Key concepts
- Diffusion Noise Space
- This refers to the set of random vectors used by diffusion models during generation. SDN treats this space not as a black box but as a controllable input. By intelligently sampling and selecting noise vectors from this space, the system can steer the policy toward more reliable actions, directly addressing issues caused by noisy or spurious visual correlations in VLA policies.
- Grounding Filter (Stage 1)
- This initial stage filters noise candidates based on visual relevance. It uses a contrastive density-ratio problem to score noise seeds that align with the task instruction while being maximally distant from 'confused' versions of the input. This ensures the selected noise vectors correspond to visually grounded and relevant parts of the observation, suppressing reliance on irrelevant background details.
- Kinematic Stability Refinement (Stage 2)
- The second stage evaluates the remaining candidates for physical feasibility. It uses a metric called JerkRMS to measure kinematic smoothness, which quantifies how 'jerky' or unstable an action trajectory is. By selecting the candidate that minimizes this jerk instability, SDN ensures that the final chosen action is not only visually grounded but also physically coherent and smooth for robotic execution.
Terminology
Summary
Diffusion-based Vision-Language-Action (VLA) policies enable strong generalization in robotic manipulation, but remain sensitive to spurious visual correlations and noisy action generation, leading to brittle behavior under perturbations <ref:2606. p. 0 Doan1
Duy M. Nguyen† 4,5,7
, Vien Anh Ngo† 1,2
1Center for AI Research, VinUniversity 2VinRobotics 3KAIST 4University of Stuttgart
5
IMPRS-IS 6National University of Singapore 7DFKI 8University of Oldenburg,
9Monash University 10University of Arkansas 9TU Darmstadt
Abstract
Diffusion-based Vision-Language-Action (VLA) policies enable strong generalization in robotic manipulation, but remain sensitive to spurious visual correlations
and noisy action generation, leading to brittle behavior under perturbations. Nguyen† 4,5,7
, Vien Anh Ngo† 1,2
1Center for AI Research, VinUniversity 2VinRobotics 3KAIST 4University of Stuttgart
5
IMPRS-IS 6National University of Singapore 7DFKI 8University of Oldenburg,
9Monash University 10University of Arkansas 9TU Darmstadt
Abstract
Diffusion-based Vision-Language-Action (VLA) policies enable strong generalization in robotic manipulation, but remain sensitive to spurious visual correlations
and noisy action generation, leading to brittle behavior under perturbations. We
introduce SELECTED DIFFUSION NOISE (SDN), a simple, training-free test-time
method that improves both robustness and success rate by leveraging the diffusion
noise space as a controllable degree of freedom. We
introduce SELECTED DIFFUSION NOISE (SDN), a simple, training-free test-time
method that improves both robustness and success rate by leveraging the diffusion
noise space as a controllable degree of freedom. SDN dynamically samples noise
vectors that are maximally separated from a reference set to mitigate reliance on
spurious cues, while selecting candidates that yield more coherent action trajectories. We
introduce SELECTED DIFFUSION NOISE (SDN), a simple, training-free test-time
method that improves both robustness and success rate by leveraging the diffusion
noise space as a controllable degree of freedom. SDN dynamically samples noise
vectors that are maximally separated from a reference set to mitigate reliance on
spurious cues, while selecting candidates that yield more coherent action trajectories.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Problem Statement
Current VLA policies often remain fragile in out-of-distribution settings because they tend to overfit to spurious correlations and non-essential behaviors in human demonstrations <ref:2606.14084#pg2>. Furthermore, pre-trained models cannot be easily retrained to handle local distribution shifts due to the prohibitive computational cost of fine-tuning multi-billion-parameter architectures and the risk of catastrophic forgetting <ref:2606. p. 0 Furthermore, in real-world environments, a pre-trained model cannot be easily
retrained to handle local distribution shifts, such as novel lighting conditions or specific tool wear, due
to the prohibitive computational cost of fine-tuning multi-billion-parameter architectures [19, 14, 35]
and the risk of catastrophic forgetting [23, 52, 8]. These issues highlight a critical need for test-time
self-improvement [44, 13]: a mechanism that enables VLA models to autonomously refine their
action distribution during inference. Furthermore, in real-world environments, a pre-trained model cannot be easily
retrained to handle local distribution shifts, such as novel lighting conditions or specific tool wear, due
to the prohibitive computational cost of fine-tuning multi-billion-parameter architectures [19, 14, 35]
and the risk of catastrophic forgetting [23, 52, 8]. These issues highlight a critical need for test-time
self-improvement [44, 13]: a mechanism that enables VLA models to autonomously refine their
action distribution during inference. These issues highlight a critical need for test-time
self-improvement [44, 13]: a mechanism that enables VLA models to autonomously refine their
action distribution during inference. Toward that goal, current research moves beyond treating the diffusion process inside VLA as a simple
black-box generator, instead viewing it as a dynamically steerable sampling space and focusing on
three main directions.Where the paper says it
Where the paper says it
Where the paper says it
How the training signal is built
The SDN framework treats the initial diffusion noise as a controllable degree of freedom to enhance diffusion-based VLA policies by treating the initial diffusion noise as a controllable degree of freedom <ref:2606.14084#pg3>. The method employs a dual-stage filter that evaluates each candidate’s visual grounding and kinematic instability <ref:2606. p. 1 We bridge these gaps with the Selected Diffusion Noise framework, a simple, training-free one that
enhances diffusion-based VLA policies by treating the initial diffusion noise as a controllable degree
of freedom. Unlike standard “best-of-N” sampling, which typically picks trajectories based on raw
probability or scores from an external VLM [49], SDN employs a dual-stage filter that evaluates each
candidate’s visual grounding and kinematic instability [21].Where the paper says it
** Stage 1: Grounding Filter:**
The first stage aims to identify “grounded” initial noise seeds via a contrastive evaluation, where the positive set (G) consists of candidates from the original observation and instruction, and the negative set (B) is generated under a “confused” version of the input where the target object is zero-masked <ref:2606.14084#pg4>. Grounding is formulated as a contrastive density-ratio problem between a task-aligned distribution pG and a failure distribution pB, with the score defined as log density ratio R(Ai) = log pG(Ai) − log pB(Ai) <ref:2606. p. 5 We formulate grounding as a contrastive density-ratio problem between a task-aligned
distribution pG and a failure distribution pB. Given a candidate action chunk Ai
, we define its score
as the log density ratio R(Ai) = log pG(Ai) − log pB(Ai). For efficiency and
robustness, we use a monotonic surrogate that preserves the induced ranking:
R
(i)
ground = d
(i)
B − d
(i)
G
. (7)
We rank candidates by Rground and retain the top-M for refinement.Where the paper says it
Where the paper says it
** Stage 2: Kinematic Stability Refinement:**
The second stage evaluates the remaining M candidates for kinematic smoothness using the JerkRMS metric to ensure physical feasibility <ref:2606.14084#pg4>. The smoothness metric S(Ai) is computed as a third-order finite difference capturing discrete jerk, and the final action is selected as A∗ = arg min Ai∈GM S(Ai) <ref:2606. p. 5 , a
(i)
L
],
we compute:
S(Ai) =
vuut
1
L − 3
L
X−3
t=1
∆a
(i)
t
2
2
, where ∆a
(i)
t = a
(i)
t+3 − 3a
(i)
t+2 + 3a
(i)
t+1 − a
(i)
t
. (8)
This third-order finite difference captures discrete jerk, and S(Ai) corresponds to its RMS magnitude
over time.Where the paper says it
Where the gains come from
SDN achieves success rate gains of +8% in simulation and +10% in real-world settings, while producing smoother and more stable actions <ref:2606.14084#pg4>. The results highlight that diffusion noise selection can play as an effective and general mechanism for enhancing VLA policies at test time <ref:2606. p. 0 SDN consistently improves success rates by +8% in simulation
and +10% in real-world settings, while producing smoother and more stable actions. Our results highlight that diffusion noise selection can play as an effective and
general mechanism for enhancing VLA policies at test time. 2 Simulation Experiments
Table 1 presents the performance of SDN on the SIMPLER benchmark using Gr00t N1. On the
Google Robot platform, SDN consistently surpasses the base model, achieving significant gains,
most notably tripling the success rate on the "Open Drawer" task (from 4. 2 Simulation Experiments
Table 1 presents the performance of SDN on the SIMPLER benchmark using Gr00t N1. On the
Google Robot platform, SDN consistently surpasses the base model, achieving significant gains,
most notably tripling the success rate on the "Open Drawer" task (from 4.Where the paper says it
Open Drawer
task (from 4.0% to 14.0%) on the Google Robot platform <ref:2606. p. 6Where the paper says it
Where the paper says it
Efficiency and Robustness
SDN is a plug-and-play enhancement for architectures such as π0 and GR00T without requiring auxiliary networks or parameter updates <ref:2606.14084#pg4>. It operates entirely self-contained and training-free, enabling immediate test-time improvement without secondary evaluators <ref:2606. p. 2 These requirements often limit the immediate,
zero-shot deployment of frozen foundation models in the wild, necessitating self-contained test-time
improvements. In contrast, SDN is entirely self-contained and training-free, enabling immediate
test-time improvement without secondary evaluators. To address core visual grounding failures, Policy Contrastive Decoding (PCD) [50] dynamically
steers probability distributions by contrasting original and object-masked views; however, this often
requires expensive per-step image inpainting. Concurrently, methods like White Noise Guidance
(WNG) [3] and Action Coherence Guidance (ACG) [36] disrupt internal temporal consistency to
steer models away from jerky trajectories. 33%) while operating
1. 53× faster than PCD. (6)
Substituting into the density-ratio objective gives R(Ai) ∝ log d
(i)
B − log d
(i)
G
. For efficiency and
robustness, we use a monotonic surrogate that preserves the induced ranking:
R
(i)
ground = d
(i)
B − d
(i)
G
.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Limitations
SDN might suffer from corruption specificity, as object-centric visual shortcuts may fail to capture hallucinations induced by dynamic lighting, texture shifts, or broader environmental distractions <ref:2606.14084#pg4>. Additionally, SDN currently relies on a fixed number of noise candidates at inference time, limiting adaptivity to varying task complexity and uncertainty <ref:2606. p. 9 , object-centric visual shortcuts may fail to capture hallucinations
induced by dynamic lighting, texture shifts, or broader environmental distractions. In addition, SDN
currently relies on a fixed number of noise candidates at inference time, limiting adaptivity to varying
task complexity and uncertainty. In addition, SDN
currently relies on a fixed number of noise candidates at inference time, limiting adaptivity to varying
task complexity and uncertainty. We believe future work on adaptive candidate allocation and richer
multimodal corruption modeling could further improve the robustness, efficiency, and generalization
of diffusion-based robotic policies.Where the paper says it
Where the paper says it
visual shortcuts
<ref:2606.14084#pg4>.
Discussion and Conclusion
The mechanism demonstrates consistent gains in success rate across settings, both in simulation and in real-world setups, and transfer across diverse backbones <ref:2606.14084#pg4>. The sequential integration of both stages ultimately contributes to the strong overall performance of SDN <ref:2606. p. 7 When combining both stages, we achieve the
best performance (48. ), revealing strong synergy between semantic grounding and motion
stability. 53× faster than PCD. Consequently, SDN provides a cleaner and more efficient pathway for
grounding VLA actions in real-time sensorimotor streams. • Dual-Objective Optimization: We propose a novel noise-selection strategy that unifies visual
grounding with kinematic stability, simultaneously maximizing separation from unreliable
action sets and minimizing a smoothness metric to eliminate high-jerk, oscillatory motions. • Plug-and-Play Enhancement: We provide extensive validation across diverse diffusion-based
VLA backbones (π0, GR00T-N1. • Plug-and-Play Enhancement: We provide extensive validation across diverse diffusion-based
VLA backbones (π0, GR00T-N1. 6) and benchmarks such as Google Robot, Widow-X
X [24], as well as two real-world robot tasks with Aloha robot, demonstrating consistent
success-rate gains (8 − 10%) and improved physical reliability at test-time without requiring
retraining or external evaluators. SDN consistently improves success rates by +8% in simulation
and +10% in real-world settings, while producing smoother and more stable actions. Our results highlight that diffusion noise selection can play as an effective and
general mechanism for enhancing VLA policies at test time.Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
Where the paper says it
--- Page 1 ---
Self-Improving VLA Policies: Selected Diffusion Noise for Spurious-Robust Action Smoothing <ref:2606.14084#pg2>
--- Page 2 ---
Diffusion-based Vision-Language-Action (VLA) policies enable strong generalization in robotic manipulation, but remain sensitive to spurious visual correlations and noisy action generation, leading to brittle behavior under perturbations <ref:2606.14084#pg3>. SDN dynamically samples noise vectors that are maximally separated from a reference set to mitigate reliance on spurious cues, while selecting candidates that yield more coherent action trajectories <ref:2606.14084#pg4>.
--- Page 3 ---
Toward that goal, current research moves beyond treating the diffusion process inside VLA as a simple black-box generator, instead viewing it as a dynamically steerable sampling space and focusing on three main directions <ref:2606.14084#pg2>.
--- Page 4 ---
First, we observe that current VLA models often suffer from a reliance on raw visual input that leads to “hallucinations” and spurious correlations <ref:2606.14084#pg3>. Second, these policies frequently prioritize semantic likelihood over physical feasibility, meaning they might select a trajectory that technically reaches the goal but involves sudden, high-jerk rotations or “jerky” motions <ref:2606. p. 1 For example, in a “place the carrot
onto the plate” task, the robot may incorrectly assume that the carrot has already been grasped and
continue executing the placement behavior even after a failed grasp attempt, causing the arm to move
toward the plate while holding nothing, as illustrated in Figure 3. Second, these policies frequently
prioritize semantic likelihood over physical feasibility, meaning they might select a trajectory that
technically reaches the goal but involves sudden, high-jerk rotations or “jerky” motions [6, 15].Where the paper says it
--- Page 5 ---
We bridge these gaps with the Selected Diffusion Noise framework, a simple, training-free one that enhances diffusion-based VLA policies by treating the initial diffusion noise as a controllable degree of freedom <ref:2606.14084#pg3>. Unlike standard “best-of-N” sampling, which typically picks trajectories based on raw probability or scores from an external VLM [49], SDN employs a dual-stage filter that evaluates each candidate’s visual grounding and kinematic instability <ref:2606. p. 1 We bridge these gaps with the Selected Diffusion Noise framework, a simple, training-free one that
enhances diffusion-based VLA policies by treating the initial diffusion noise as a controllable degree
of freedom. Unlike standard “best-of-N” sampling, which typically picks trajectories based on raw
probability or scores from an external VLM [49], SDN employs a dual-stage filter that evaluates each
candidate’s visual grounding and kinematic instability [21].Where the paper says it
--- Page 6 ---
Stage 1 (Contrastive Grounding): We compute a grounding score R(i) ground = d(i)B −d(i)G using non-parametric k-NN density estimation, selecting action chunks that maximize distance from “hallucinated” behaviors (push from B) while maintaining task consensus (pull to G).
--- Page 7 ---
Stage 2 (Smoothness Filtering): From the top-M grounded candidates, we perform a final kinematic optimization to select the action chunk S(A(q)i) that minimizes jerk instability to select the most coherent action <ref:2606.14084#pg4>.
--- Page 8 ---
We model the policy output distribution as a mixture of the expert data distribution p∗(·) and a failure distribution pfail(·): πθ(A o, l) = 1 −ϵ(o, l)p∗(A o, l) + ϵ(o, l) pfail(A o, l). <ref:2606.14084#pg4>.
--- Page 9 ---
To intuitively understand these failures, we model the policy output distribution as a mixture of the expert data distribution p∗(·) and a failure distribution pfail(·): πθ(A o, l) = 1 −ϵ(o, l)p∗(A o, l) + ϵ(o, l) pfail(A o, l). <ref:2606.14084#pg4>.
--- Page 10 ---
From Local Guidance to Global Selection. A common strategy to mitigate failure modes in generative models is guidance, which biases sampling toward more desirable outputs and often improves sample quality [16, 41]. p. 4 We
render zero-masked regions as transparent for better
visualization
From Local Guidance to Global Selection. A common strategy to mitigate failure modes in generative models is guidance, which biases sampling toward more
desirable outputs and often improves sample quality [16, 41].Where the paper says it
--- Page 12 ---
We define the corresponding local scale estimates: d(i)G = 1/kXj∈omegaG i∥Ai − Aj∥2, d(i)B = 1/kB Xj∈omegaB i∥Ai − Aj∥2. <ref:2606.14084#pg4>.
--- Page 13 ---
We compute: R(i)ground = d(i)B − d(i)G. <ref:2606.14084#pg4>.
--- Page 15 ---
We then perform the Stage-2 refinement by evaluating the remaining M candidates using the kinematic smoothness metric JerkRMS [10].
--- Page 19 ---
Theorem A.2. Let R(A) denote the population grounding score: R(A) = pB(A) − 1/D − pG(A) <ref:2606. p. 3 To improve adherence to the instructions, guidance can be applied during ODE integration by modifying the velocity vector field. Let
V
cond
θ
(Aτ
, τ, e) and V
uncond
θ
(Aτ
, τ, e
uncond) denote the conditional and unconditional vector
fields, where the latter is often obtained by masking the instruction (i. To improve adherence to the instructions, guidance can be applied during ODE integration by modifying the velocity vector field. Let
V
cond
θ
(Aτ
, τ, e) and V
uncond
θ
(Aτ
, τ, e
uncond) denote the conditional and unconditional vector
fields, where the latter is often obtained by masking the instruction (i.Where the paper says it
Where the paper says it
--- Page 18 ---
Hallucination Suppression via Best-of-N SDN Selection Theorem A.2. Let R(A) denote the population grounding score: R(A) = pB(A) − 1/D − pG(A) <ref:2606.14084#pg4>.
--- Page 19 ---
P(A∗ ∼ pfail) ≤ exp(−cN), for some constant c > 0 <ref:2606. p. 0 , classifier-free [16] or
†: Project leads. arXiv:2606.Where the paper says it
--- Page 17 ---
SDN consistently improves upon the base model, with significant gains such as tripling the success rate on the Open Drawer
task (from 4.0% to 14.0%) on the Google Robot platform <ref:2606.14084#pg4>.
--- Page 18 ---
SDN achieves an average success rate of 48.33%, a substantial 18. p. 4 A
natural extension of CFG would take the
form
V
guided
θ = V
cond
θ + w
V
cond
θ − V
fail
θ
. (4)
In practice, failure-aware guidance is challenging because failure trajectories are rarely labeled,
learning an additional failure model introduces substantial overhead [18], and trajectory quality
remains sensitive to guidance scale during sampling [43, 45]. 33%, a
substantial 18. 33% absolute improvement over the vanilla model and significantly outperforming
baselines like CFG, ACG, and PCD.Where the paper says it
Where the paper says it
--- Page 19 ---
SDN - Smoothness 50.00 ↑20.00 43.33 ↑13.33 46.67 ↑16.67 <ref:2606. p. 0 , classifier-free [16] or
†: Project leads. arXiv:2606.Where the paper says it
--- Page 21 ---
SDN - Full 53.33 ↑29.99 50.00 ↑28.0 <ref:2606.14084#pg4>.
--- Page 17 ---
SDN - Grounding 50.00 ↑28.7% <ref:2606. p. 0 , classifier-free [16] or
†: Project leads. arXiv:2606. , classifier-free [16] or
†: Project leads. arXiv:2606.Where the paper says it
Where the paper says it
--- Page 17 ---
SDN - Full 53.33 ↑29.99 <ref:2606. p. 0 , classifier-free [16] or
†: Project leads. arXiv:2606. , classifier-free [16] or
†: Project leads. arXiv:2606.Where the paper says it
Where the paper says it
--- Page 18 ---
SDN - Grounding 50.00 ↑28.7% <ref:2606.14084#pg4>.
--- Page 17 ---
SDN - Smoothness 53.33 ↑29.99 <ref:2606.14084#pg4>.
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
-
Self-Diagnostic Latent Steering: The system can identify and filter
hallucinating
noise seeds by contrasting policy outputs from original and object-masked views, steering the frozen model away from spurious visual correlations. This mechanism suppresses reliance on spurious background cues by checkingwhether action candidates remain sensitive to the target object’s visibility.
-
Dual-Objective Optimization: The system can perform a novel noise-selection strategy that unifies visual grounding with kinematic stability, simultaneously
maximizing separation from unreliable action sets and minimizing a smoothness metric to eliminate high-jerk, oscillatory motions.
This ensures both logical task grounding and physical feasibility for the selected trajectory. -
Plug-and-Play Enhancement: The system provides an enhancement across diverse diffusion-based VLA backbones like pi0 and GR00T N1.6 without requiring retraining or auxiliary networks. This makes it a
plug-and-play
solution for immediate test-time improvement in real robotic environments.
Abstract
Diffusion-based Vision-Language-Action (VLA) policies have demonstrated strong generalization in robot manipulation, but their deployment can remain unreliable due to spurious visual reliance and stochastic action generation, leading to grounding failures and kinematically unstable behaviors. We introduce Selected Diffusion Noise (SDN), a training-free method that treats diffusion noise as a controllable degree of freedom for improving test-time reliability. SDN samples multiple candidate action chunks and applies two-stage selection: it first filters candidates based on task-relevant visual grounding using a counterfactual negative set, then selects the kinematically most stable chunk based on JerkRMS. Experiments on three VLA policies across both simulation and real-world environments show improvements of 8-10% and 18.33%, respectively, demonstrating SDN's effectiveness for test-time improvement.
Sources
- GR00T N1: An Open Foundation Model for Generalist Humanoid Robots
- $\pi_0$: A Vision-Language-Action Flow Model for General Robot Control
- SteerVLA: Steering Vision-Language-Action Models in Long-Tail Driving Scenarios
- Self-Supervised Policy Adaptation during Deployment
- Classifier-Free Diffusion Guidance
- Latent Policy Steering through One-Step Flow Policies
- Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success
- OpenVLA: An Open-Source Vision-Language-Action Model
- RoboReward: General-Purpose Vision-Language Reward Models for Robotics
- Evaluating Real-World Robot Manipulation Policies in Simulation
- Global Prior Meets Local Consistency: Dual-Memory Augmented Vision-Language-Action Model for Efficient Robotic Manipulation
- Flow Matching Guide and Code
- OK-Robot: What Really Matters in Integrating Open-Knowledge Models for Robotics
- VLS: Steering Pretrained Robot Policies via Vision-Language Models
- LaST$_{0}$: Latent Spatio-Temporal Chain-of-Thought for Robotic Vision-Language-Action Model
- SAM 2: Segment Anything in Images and Videos
- DINO-X: A Unified Vision Model for Open-World Object Detection and Understanding
- Goal-Conditioned Imitation Learning using Score-based Diffusion Policies
- Stay on topic with Classifier-Free Guidance
- Steering Your Diffusion Policy with Latent Space Reinforcement Learning
Related papers
- Rethinking Actions: Continuity-Aware Priors for World-Model Planning
- MOSAIC: Modular Obstacle and Spatial-Aware Framework for Manipulation In Cluttered Scenes
- SAGAS: Semantic-Aware Graph-Assisted Stitching for Offline Temporal Logic Planning
- DexSeed: Expanding Human Hand-Object Interactions into Robot Demonstrations for Dexterous Policy Learning
- LaST0.5: A Generalist World Action Model with Interaction-Centric Latent Reasoning
- CIALLO: Learning Shared Intent Representations for Cross-Embodiment Manipulation