acceptodds
Under review as a conference paper at ICLR 2027

Beyond Positive Bias: Customizing and Correcting Generative Robot Policies via Negative Guidance

Abstract

Diffusion policies can represent multimodal robot manipulation behaviors, but deployment may introduce constraints that were not present during policy training. Extending negative guidance to robot control is nontrivial: guided action chunks are repeatedly executed in a closed loop, where excessive repulsion can compound over time, reduce task completion, or collapse the remaining valid behaviors. We extend Dynamic Negative Guidance to diffusion robot policies and develop a multi-target formulation for simultaneously avoiding multiple undesirable outcome modes. We instantiate the framework through two complementary applications. First, language-conditioned negative guidance provides a semantic interface for specifying single or multiple outcomes to avoid. Second, when undesirable behavior is identified through deployment feedback rather than language, an auxiliary diffusion policy trained on a small set of failure rollouts provides the negative guidance signal while the base-policy parameters remain fixed. We evaluate the framework across object picking, target placement, planar pushing, and precise charger insertion, including a real-robot manipulation experiment. As an additional generalization study, we further evaluate language-conditioned negative guidance with flow-matching policies on LIBERO. Results show that dynamic negative guidance suppresses specified undesirable outcomes while better preserving task completion and alternative valid behaviors than static negative guidance, and that failure-policy guidance improves task success across the evaluated deployment-time environment changes.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robot safety, alignment, and safe learning-based systems, Robot manipulation, Diffusion Policies

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_zXXeOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates how generative robot policies can adapt to constraints during deployment without requiring modifications to the pre-trained base policy. Building on the "Dynamic Negative Guidance" (DNG) method, the authors extend the negative guidance mechanism to closed-loop diffusion policies and derive a joint-complement formulation to avoid multiple mutually exclusive undesirable outcomes. The framework is implemented in two scenarios: language-conditioned avoidance, which converts learned positive conditioning into negative guidance signals; and failure-driven avoidance, which utilizes an auxiliary diffusion policy—trained on deployment-time failure cases—to provide negative guidance. Additionally, the authors explore the adaptive extension of this method to flow-matching policies.

Experiments cover object grasping, target placement, Push-T tasks, charger insertion, real-world manipulation, and LIBERO-Goal tasks based on the π0 policy. The results demonstrate that Dynamic Negative Guidance achieves a favorable trade-off between avoiding undesirable outcomes, completing tasks, and preserving behavioral diversity; furthermore, the failure-driven guidance mechanism effectively enhances system performance across various deployment environments.

Critical strengths
  1. Addresses practical issues within a unified framework. This paper explores the practical challenge of adapting pre-trained robot policies to new deployment constraints without retraining the base policy. The unified negative guidance framework supports both language-guided obstacle avoidance and failure-feedback-based correction. Evaluations consider both task completion and the preservation of other effective behavioral patterns.

  2. Shows great promise across various manipulation tasks. Experiments cover tasks such as object grasping, obstacle avoidance, and charger insertion. The results based on failure feedback are particularly encouraging, showing significant improvements in success rates—especially for the Push-T and charger insertion tasks. Experiments with real robots and extensions to the π0 model further demonstrate the method's potential applicability.

  3. Clear overall presentation. The paper is well-written and logically organized. The research motivation, the proposed framework, and the experimental results are all clearly and accessibly presented.

Critical weaknesses

1. Limited methodological innovation and insufficient validation of the proposed extensions.

The core methodology is primarily built upon Dynamic Negative Guidance (DNG, ICLR 2025), a fact clearly stated in Section 3.1 (Page 3, Equations 2–3). The multi-objective extension proposed in Section 3.2 (Pages 3–4, Equations 4–7) is derived directly from the assumption that the outcomes are mutually exclusive. However, the paper lacks ablation studies comparing the proposed "joint-complement" formulation against a simple combination of independent, single-objective DNG corrections; this makes it difficult to assess the actual contribution of the extension.

Similarly, the failure-driven approach in Section 3.3 (Page 5, Equation 12) essentially replaces "language-conditioned negative scoring" with "scoring learned from failure trajectories." While the performance gains shown in Tables 5–6 are encouraging, the experiments fail to adequately distinguish the specific benefits of this design beyond the mere application of existing negative guidance techniques to an auxiliary failure model.

Furthermore, in the language-conditioned experiments (Section 4.1 and Appendix 7.3), the policy has already learned positive instructions for specific goals. A simple alternative would be to select an admissible goal at the start of an episode and execute its corresponding positive instruction. The paper lacks a comparison against this baseline.

Overall, while the method's application to robotic manipulation tasks is significant, the paper has not convincingly demonstrated that its proposed extensions offer an independent methodological contribution compared to existing guidance mechanisms and simple alternatives.

2. Inadequate control in baseline comparisons; performance gains depend on specific metrics.

The experimental results fail to clarify whether the reported performance improvements stem from the proposed dynamic weighting mechanism or from differences in effective guidance strength. Specifically, Appendices 7.2.2–7.2.3 (p. 14, Equations 31–32) indicate that some language-conditioned experiments employed a fixed CFG coefficient (value of 3), whereas the DNG method utilized a coefficient dependent on posterior probabilities. Under the reported initialization settings (p0=0.25p_0=0.25 and λ0=3\lambda_0=3), the initial effective DNG coefficient was merely 1 (compared to 3 for static CFG), and this coefficient typically decreased during the denoising process. A similar discrepancy exists in the π0\pi_0 experiments (see Appendix 7.1, p. 13). Consequently, the higher failure rates and mode collapse observed with static CFG may stem, in part, from its stronger effective guidance, rather than from the inherent superiority of the proposed adaptive mechanism.

Although Appendix Table 8 (p. 20) includes an ablation study on the scale coefficient, the results demonstrate that reducing the CFG scale coefficient from 3 to 1 improved the AUS from 0.800 to 0.830, thereby narrowing the gap with the proposed method's best score (0.851). Therefore, a more comprehensive comparison involving properly tuned static guidance and simple decaying guidance strategies is warranted.

Furthermore, the magnitude of performance improvement depends, to some extent, on the chosen evaluation metrics. In Table 1 (p. 6), static CFG achieved a higher "constraint-satisfying task success rate" across all four object-grasping scenarios; the proposed method attained a higher AUS because it tends to maintain uniformity among the remaining valid results. For instance, in the task involving avoiding a cube, CFG achieved an effective success rate of 90.0%, whereas the proposed method achieved 86.0%. Appendix Table 7 (p. 17) further shows that when the uniformity weight is reduced to 0.1, CFG outperforms the proposed method in terms of the average AUS metric for object-grasping tasks.

These results suggest that the advantages claimed in the paper should be interpreted with caution—viewing them as a trade-off between constraint satisfaction and behavioral diversity rather than an across-the-board improvement in task completion quality. To robustly support the paper's core empirical conclusions, stronger baselines and a clearer distinction between evaluation metrics are required.

Questions for authors

Q1. Contribution of the multi-target formulation.

In Section 3.2 (Eqs. 4–7), the joint-complement formulation follows from the assumption of mutually exclusive outcomes. Could the authors clarify what practical advantage the shared denominator provides over simply combining independent single-target DNG corrections? Can we add an abslation.

Q2. Fairness of the guidance baseline comparisons.

According to Appendix 7.2.2–7.2.3, the initial effective DNG coefficient is 1 when the nominal scale is 3 and the prior is 0.25, whereas static CFG uses a fixed coefficient of 3. Could the authors provide comparisons with better-tuned static CFG, including smaller coefficients, and simple decaying guidance schedules? In particular, do the improvements in Tables 1, 4, 5, and 6 remain consistent when the effective guidance strengths are more closely matched? This would be important for determining whether the gains come from the proposed adaptive mechanism.

Q3. Necessity of negative guidance in the language-conditioned setting.

In Section 4.1 and Appendix 7.3, the base policy is already trained with target-specific positive instructions. For the closed-set object-picking and charger-insertion tasks, could the authors compare against selecting an allowed target at the beginning of each episode and executing its corresponding positive instruction? Such a baseline could potentially achieve constraint satisfaction while preserving diversity across episodes.

5 comments
AuthorsOpenReview ↗

W1 / Q1. Methodological contribution and joint-complement ablation

We thank the reviewer for the constructive assessment. We acknowledge that single-target Dynamic Negative Guidance (DNG) was introduced for image generation. Our contribution is not to reinvent that mechanism, but to formulate and evaluate negative behavioral steering for physically executed, closed-loop robot policies, together with multi-target, failure-driven, and flow-matching extensions.

Why the robotics setting matters. In image generation, negative guidance suppresses undesirable visual content. In robotics, generated action chunks are executed: they alter subsequent observations, task feasibility, and future policy decisions. A behavior that succeeded during training (e.g., insertion into a receptacle) may become forbidden at deployment. Simply suppressing that outcome is insufficient unless the robot still completes the task through an allowed alternative. In Table 2, when receptacles (1,2,3) are forbidden, rejection sampling reduces forbidden selections to 12.5% but incurs 46.0% failures (41.5% constraint-valid success); DNG achieves 9.5% forbidden selections, 19.5% failures, and 71.0% constraint-valid success. This demonstrates the distinct avoidance–completion challenge of embodied steering. We do not claim a cross-chunk posterior-consistency guarantee.

Why joint-complement guidance differs from independent composition. For mutually exclusive negative outcomes with posterior probabilities pip_i, the desired joint complement is proportional to pt(xt)(1−∑ipi)p_t(x_t)(1-\sum_i p_i), whereas independent single-target composition corresponds to pt(xt)∏i(1−pi)p_t(x_t)\prod_i(1-p_i). Their score corrections therefore use, respectively,

gijoint=λ0pi1−∑jpj,giind=λ0pi1−pi.g_i^{\mathrm{joint}}=\lambda_0\frac{p_i}{1-\sum_jp_j},\qquad g_i^{\mathrm{ind}}=\lambda_0\frac{p_i}{1-p_i}.

The shared denominator explicitly accounts for the total forbidden probability mass without enumerating allowed targets. We agree that this is a natural extension of DNG, rather than a new probability identity.

New ablation. We implemented independent single-target DNG and evaluated both formulations on charger insertion. At the same nominal scale 3, over 200 episodes per condition, joint DNG improves constraint-valid success from 51% to 54% for forbidden targets (1,2,4), and from 46% to 48% for (1,4,5). Since nominal scales do not match initial effective strengths, we additionally performed a matched-coefficient sweep on 40 common seeds (101–140). At initial coefficient 2.0 per branch (joint scale 4; independent scale 8), joint DNG obtains 42.5% versus 35.0% valid success, and 12.5% versus 22.5% forbidden selections. The ranking changes at other coefficients; thus, we claim a principled joint-complement formulation, not universal superiority over independently composed guidance.

Failure-driven and flow-matching extensions. Deployment failures reveal what to avoid, but need not provide a successful replacement trajectory. Our auxiliary failure diffusion policy supplies that negative distribution while keeping the pretrained base policy frozen; at a single fixed scale λ0=2\lambda_0=2, this improves performance across all six obstacle and unavailable-receptacle conditions in Tables 5–6. For flow-matching π0\pi_0, we derive a complement-velocity relation and a state-dependent surrogate instead of reusing DNG's stochastic posterior update. On LIBERO-Goal, A2 avoidance reduces forbidden outcomes 10 to 3, and A4+A2+A5 avoidance reduces 25 to 6, with the baseline's 18/100 failures unchanged in both cases.

We will sharpen these methodological distinctions and limitations in the revision.

AuthorsOpenReview ↗

W2 / Q2. Matched-strength CFG, decay controls, and distribution fidelity

We agree that comparing the same nominal scale is insufficient: for single-target DNG, gt=λ0pt/(1−pt)g_t=\lambda_0p_t/(1-p_t), so λ0=3\lambda_0=3 with p0=0.25p_0=0.25 begins at effective coefficient 1, not 3. We therefore added smaller static-CFG coefficients, matched-initial-strength controls, and deterministic decays.

Banana avoidance (200 episodes per setting). The complete original Table 8 scale sweep, supplemented with new CFG controls, gives:

MethodScale/scheduleForbidden banana ↓Valid success ↑AUS ↑
DNG17.0%83.0%.811
DNG31.5%89.5%.851
DNG52.5%88.5%.844
DNG101.5%93.0%.829
CFG0.5 (new)9.0%83.5%.793
CFG13.0%91.5%.830
CFG1.5 (new)1.5%93.0%.829
CFG2 (new)1.5%91.5%.812
CFG31.0%93.0%.800
CFG51.5%93.5%.783
CFG101.5%91.0%.734
CFGlinear 1→0.5 (new)4.5%86.0%.790

The key advantage in picking is preservation of alternative valid behaviors, not maximal raw success. DNG-3 selects cube/apple/cup at 37.0/20.0/32.5%, versus 53.5/9.0/30.5% for CFG-1.5, despite both selecting banana only 1.5% of the time. To separate this effect from AUS's preference for uniformity, we calculated total variation (TV) distance to the base policy's renormalized distribution over allowed successful objects. At matched observed avoidance, DNG-3 has TV 0.063 versus 0.190 for CFG-1.5. CFG-0.5 has similarly low TV (0.065), but selects banana 9.0% of the time. Hence weaker fixed CFG does not match the same avoidance–fidelity trade-off. We acknowledge CFG-1.5 has higher raw valid success (93.0% versus 89.5%).

This conclusion extends beyond AUS: recomputing TV from the original outcome frequencies, DNG is closer to the renormalized base distribution than CFG in 4/4 object-picking conditions (mean TV 0.259 vs. 0.455) and 9/10 charger conditions (mean 0.202 vs. 0.355). Across these ten charger conditions, mean constraint-valid success is 58.6% for DNG, versus 51.6% for CFG and 33.4% for the unguided policy.

Matched-strength three-target charger insertion. For forbidden receptacles (1,4,5), three priors of 0.2 yield initial DNG coefficient g0=λ0/2g_0=\lambda_0/2 per branch. On the same 80 environment seeds, matched static-CFG comparisons yield:

Initial coefficientDNG valid ↑Static CFG valid ↑
1.062.5%25.0%
1.567.5%31.25%
2.073.75%30.0%

At initial coefficient 1.5, exponential CFG decay to 1.3 or 0.75 yields 28.75% or 23.75% success, versus 67.5% for DNG. This result is from one selected charger configuration, not all settings.

These controls show that the favorable avoidance–fidelity trade-off in picking and valid-success performance in this multi-target charger setting cannot be explained solely by an overly strong CFG baseline or the tested decay schedules. We will report raw success, forbidden rates, and distribution fidelity separately. We do not claim that these controls exhaust all CFG schedules or establish matched-strength superiority for every π0\pi_0 and failure-driven experiment.

AuthorsOpenReview ↗

Q3. Positive-target selection versus negative guidance

We thank the reviewer for proposing this natural baseline. We implemented it for the closed-set charger task using the original pretrained policy, with no additional training. At episode initialization, we select an admissible receptacle either uniformly or proportionally to the unguided policy's observed frequencies, then execute its positive instruction throughout the episode.

Across 200 episodes per method and avoidance condition, we obtain:

Forbidden targetsSelectionValid success ↑Forbidden ↓Failure ↓
(1,2,4)Uniform75.0%7.5%17.5%
(1,2,4)Base-proportional76.5%6.5%17.0%
(1,4,5)Uniform84.5%3.0%12.5%
(1,4,5)Base-proportional84.5%3.0%12.5%

These are strong results, and positive selection can outperform negative guidance in this closed-set setting. We will include this baseline and clarify the scope of our claim.

However, the methods address different deployment information assumptions. Positive selection presumes that (i) the complete set of admissible goals is known, (ii) each has a reliable positive instruction, and (iii) the robot can commit to a selected replacement goal. Base-proportional selection additionally uses estimated base-policy frequencies for those goals. These assumptions hold in our controlled five-receptacle task, but need not hold when a robot is given only what not to do.

This distinction is especially important for deployment failures. A newly encountered obstacle in Push-T may identify unsuccessful trajectories without revealing any successful replacement trajectory, let alone a finite list of positive targets. Our failure-driven approach trains an auxiliary negative model on unsuccessful rollouts and redirects the frozen base policy away from those behaviors, without enumerating or positively labeling successful alternatives. At a single fixed DNG scale of λ0=2\lambda_0=2, this improves performance in all six failure-driven deployment conditions in Tables 5–6.

Thus, positive-target selection is preferable when admissible alternatives are fully specified; negative guidance is valuable when exclusions or failures are known but positive replacements are not. We will present both methods as complementary rather than claim universal superiority of negative guidance.

Reviewer_4r5cOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates inference-time negative guidance for adapting generative robot policies to deployment-time constraints without modifying the pretrained base policy. Building upon Dynamic Negative Guidance (DNG), the authors extend the formulation to multiple simultaneous avoidance targets and investigate two complementary settings: language-conditioned avoidance and failure-driven avoidance using an auxiliary diffusion policy trained on unsuccessful rollouts. The framework is evaluated on several manipulation tasks, including object picking, Push-T, charger insertion, and real-robot bowl placement, with additional experiments on π0 in LIBERO. The results suggest that dynamic negative guidance can suppress undesirable behaviors while better preserving task completion and alternative valid outcomes than static guidance.

Critical strengths
  1. The paper addresses a practical and relevant problem: adapting pretrained robot policies to new deployment constraints without requiring retraining of the base policy.
  2. The framework explores two complementary sources of negative guidance: language-specified undesirable outcomes and an auxiliary policy learned from failure rollouts. This provides flexibility for different deployment scenarios.
  3. The experiments cover multiple manipulation settings, including multi-target avoidance, obstacle-induced failures, and unavailable insertion targets. The inclusion of real-robot experiments and π0 on LIBERO further strengthens the empirical evaluation.
  4. The paper evaluates both avoidance effectiveness and task completion, rather than treating suppression of forbidden outcomes as sufficient. The proposed Avoidance-Uniformity Score also considers the diversity of remaining valid behaviors.
Critical weaknesses
  1. Limited novelty. The method is mainly based on DNG and extends it to robot policies with multiple negative conditions. DynaGuide also studies inference-time guidance for robot diffusion policies. The methodological novelty appears quite limited.
  2. Hyperparameter selection. It is unclear how the guidance scales are selected and whether they are tuned on the test set. The robustness of the method under a fixed guidance configuration across different tasks is also unclear.
  3. Limited generalization. Most experiments use task-specific diffusion policies. Even the π0 experiments require separately trained checkpoints. I think this method cannot be directly applied to general-purpose VLA models without additional training.
  4. Inference efficiency. The method requires additional policy evaluations for negative conditions, which may increase inference costs. However, the authors do not report inference latency or computational overhead.
Questions for authors

see weaknesses.

6 comments
AuthorsOpenReview ↗

W1. Novelty relative to image-generation DNG and robotic DynaGuide

We thank the reviewer for raising the novelty concern. The original DNG was proposed for image generation; we build on its single-target principle, and we acknowledge that DynaGuide already studies inference-time steering of diffusion robot policies. We therefore do not claim to be the first to steer robot policies at inference time. Our contribution lies in the deployment objective, guidance mechanism, and methodological extensions.

(1) Why negative steering differs in robotics.

In image generation, negative guidance suppresses undesirable visual content while balancing sample quality. In robotics, the guided output is physically executed: it changes future observations, robot states, and which actions remain feasible.

A behavior can be successful under the pretrained policy yet become forbidden only at deployment—for example, inserting into a newly unavailable charger receptacle. Merely avoiding that receptacle does not complete the task. The robot must retain executable alternative behaviors through subsequent feedback and replanning.

This trade-off is measurable. With receptacles (1,2,3) forbidden (Table 2), rejection sampling lowers forbidden selections to 12.5% but raises task failures to 46.0%, yielding 41.5% constraint-valid success. DNG obtains 9.5% forbidden selections, 19.5% failures, and 71.0% constraint-valid success.

Our evaluation therefore centers on successful constrained execution rather than suppression alone. We do not claim an explicit cross-chunk posterior or formal safety guarantee.

(2) How our mechanism differs from DynaGuide.

DynaGuide employs a learned dynamics model to predict future outcomes and differentiates a compatibility objective against reference observations.

Our framework instead guides the conditional generative score or velocity field directly, using (i) target-specific language conditions already represented by the policy or (ii) an auxiliary diffusion model trained on deployment failures.

Thus, language-conditioned guidance does not require a forward dynamics predictor or desired-outcome reference images, although it does require grounded negative conditions. Failure guidance requires training an auxiliary model but leaves the base policy frozen.

These are different assumptions and complementary approaches, not a claim of universal superiority.

(3) Technical extensions beyond the original DNG formulation.

For mutually exclusive forbidden outcomes c1,…,cKc_1,\ldots,c_K, we use a joint-complement denominator

1−∑jpt(cj∣xt),1-\sum_j p_t(c_j\mid x_t),

rather than simply combining independent single-target complements. This models exclusion of the union of forbidden modes without enumerating admissible alternatives. We acknowledge its natural derivation from DNG.

For flow-matching π0\pi_0, the original stochastic reverse-diffusion posterior update is inapplicable. We instead derive

v¬−=vb−rt1−rt(vc−−vb)v_{\neg-}=v_b-\frac{r_t}{1-r_t}(v_{c^-}-v_b)

from probability flux and develop an adaptive compatibility surrogate based on the current Euler proposals.

The attenuation is state-dependent, not a predetermined schedule: it depends on local broad-versus-negative velocity disagreement. We do not claim that this surrogate recovers an exact posterior.

Finally, failure feedback may specify what went wrong without identifying a positive replacement trajectory. Learning the negative distribution from such failures enables adaptation while preserving the broad task objective and base-policy weights.

We will make the image-generation-to-embodied-control distinction and the contrast with DynaGuide explicit, while acknowledging existing robotic steering methods and the limits of our evaluations.

AuthorsOpenReview ↗

W2. Hyperparameter selection and robustness

We thank the reviewer for highlighting the importance of guidance-scale selection. We agree that reporting the best scale selected from evaluation results may overestimate performance. However, our experiments also demonstrate consistent improvements using a single fixed guidance scale across different deployment conditions, without condition-specific tuning.

(1) Fixed-scale robustness across tasks. In Tables 5–6, we evaluate DNG with the same guidance scale λ0=2\lambda_0=2 across six deployment conditions: three Push-T obstacle configurations and three charger-insertion constraints. At this fixed scale, DNG improves task performance over the unguided policy in all six conditions. These settings involve different failure mechanisms, including geometric obstacles and unavailable insertion targets. This demonstrates that the observed improvements do not depend solely on selecting the best guidance strength separately for each constraint.

(2) Additional guidance-strength experiments. We also conducted matched-effective-strength comparisons against static CFG on three-target charger insertion, using the same 80 environment seeds for each configuration. At initial effective coefficients 1.0, 1.5, and 2.0, DNG achieves 62.5%, 67.5%, and 73.75% constraint-valid success, respectively, compared with 25.0%, 31.25%, and 30.0% for CFG. These results show that DNG's advantage in this setting persists across multiple guidance strengths rather than depending on a single favorable coefficient.

Our original object-picking and charger experiments use 200 evaluation environment seeds per configuration, and the expanded object-picking CFG comparison uses the same 200 seeds as the original experiment.

(3) Clarifying the selection protocol. We acknowledge that the original guidance-scale sweeps did not consistently use a separate held-out validation set for selecting the reported best scale. We will distinguish fixed-scale results from best-of-sweep results and clarify that multiple environment seeds do not constitute independently trained policy checkpoints.

Overall, the fixed-scale improvements across six deployment conditions provide evidence of practical robustness beyond per-condition scale optimization. We agree that held-out hyperparameter selection and evaluation across independently trained policies would further strengthen the analysis, and we will explicitly acknowledge these limitations.

AuthorsOpenReview ↗

W3. Generalization to general-purpose VLA policies

We thank the reviewer for this important question. We agree that compatibility with general-purpose vision-language-action (VLA) policies is essential for practical deployment. Our π0\pi_0 experiments were specifically designed to investigate whether negative guidance extends beyond conventional diffusion policies, although we acknowledge that they do not establish universal zero-shot compatibility.

(1) Extension to flow-matching VLA policies. The original DNG formulation was developed for image generation using stochastic reverse diffusion. Our earlier manipulation experiments use diffusion-based robot policies, whereas π0\pi_0 employs flow matching and generates actions through velocity-field integration.

To accommodate this difference, we derive a complement-velocity relationship,

v¬−=vb−rt1−rt(vc−−vb),v_{\neg-}=v_b-\frac{r_t}{1-r_t}(v_{c^-}-v_b),

and introduce an adaptive surrogate based on local velocity-field compatibility, rather than directly reusing the stochastic DNG posterior update.

We evaluate this extension on LIBERO-Goal, obtaining:

  • Single-target avoidance (A2): Forbidden outcomes decrease from 10 to 3 out of 100 rollouts, with task failures unchanged at 18.
  • Multi-target avoidance (A4+A2+A5): Forbidden outcomes decrease from 25 to 6, again maintaining 18 failures.

These results demonstrate that complement-based negative guidance can suppress undesirable outcomes while preserving task completion in an evaluated flow-matching VLA architecture.

(2) Deployment requirements and generality. We acknowledge that our π0\pi_0 experiments use separately trained broad-instruction and target-specific checkpoints. Therefore, we do not claim plug-and-play, zero-shot application to arbitrary pretrained VLAs.

More generally, language-conditioned guidance requires access to meaningful broad and negative-condition generative predictions. When such conditioning is unavailable, our complementary failure-driven formulation trains an auxiliary negative diffusion model from deployment failures while leaving the original base policy frozen. This provides an alternative interface for incorporating newly discovered constraints, although we have not yet demonstrated this variant on arbitrary general-purpose VLAs.

(3) Scope of our contribution. Our results establish feasibility across the evaluated diffusion and flow-matching policy architectures, not unrestricted generalization to every VLA or open-ended manipulation task.

We will make these requirements explicit and distinguish demonstrated architectural generality from broader zero-shot applicability. The central contribution is enabling deployment-time behavioral exclusion without modifying pretrained base-policy parameters, with language-conditioned and failure-driven interfaces suited to different information settings.

Reviewer_YA85OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper applies Dynamic Negative Guidance (DNG) to diffusion robot policies so that a frozen policy can be steered away from undesirable outcomes at inference time. It extends DNG to multiple forbidden targets with a shared denominator, and obtains the negative score in two ways: from target-specific language conditions learned during training, or from an auxiliary diffusion policy trained on a small set of failure rollouts.

Critical strengths

Inference-time customization of a fixed robot policy is a practical problem, and the method is simple to apply on top of existing diffusion policies without retraining. Using target-specific language conditions as negative directions, and turning a handful of failure rollouts into a repulsive signal, are both convenient interfaces.

Critical weaknesses
  1. The technical novelty is limited. The core mechanism is DNG applied as is, and the multi-target extension simply sums the per-target posteriors in the complement distribution. Language-conditioned avoidance is standard negative prompting, and failure-driven avoidance is negative guidance with a separately trained model. The paper motivates the work by arguing that closed-loop execution makes negative guidance nontrivial, but nothing in the method addresses this: the posterior is reset for every chunk, and there is no treatment or analysis of consistency across chunks. For the π0 variant, the adaptive weight is non-increasing by construction (Eq. 24), so it acts as a fixed decay schedule rather than a state-dependent posterior.

  2. The comparison with static CFG does not support the main claim. Both methods are run at the same nominal scale, but DNG’s effective weight is the scale multiplied by p/(1−p), which starts well below the scale and decays during denoising. CFG is therefore applying a much stronger repulsion. The paper’s own ablation shows CFG at a small scale nearly matching DNG, and CFG is not tuned at all on picking and charger insertion. To show that the posterior weighting matters, the paper needs CFG at a matched effective scale and a simple time-decayed CFG schedule.

  3. The empirical evidence is weaker than presented. In multi-target charger insertion, a large share of rollouts still go to forbidden receptacles, and overall valid success is far below the base policy. AUS is a custom metric that rewards uniformity over the remaining modes, whereas the complement distribution should preserve the base policy’s relative proportions; the ranking also changes with the metric’s weight. It seems the best guidance scale is selected per condition on the test results, and the rejection-sampling and PPGuide baselines perform worse than the unguided policy, which may suggest some implementation issues.

Questions for authors

How does DNG compare to CFG at a matched effective guidance strength, and to CFG with a time-decaying schedule?

Can you report results over multiple seeds, with the guidance scale selected on held-out seeds?

How does failure-driven guidance compare to simply fine-tuning the base policy on the same failure and success rollouts?

7 comments
AuthorsOpenReview ↗

W1. Technical novelty: From image-generation DNG to closed-loop robotic steering

We thank the reviewer for this important concern. We acknowledge that the original Dynamic Negative Guidance (DNG) was developed for image generation and that our single-target score correction builds on that work. Our claim is not that negative guidance itself is new. Rather, transferring it to physically executed, closed-loop robot actions changes the deployment objective and motivates our multi-target, failure-driven, and flow-matching extensions.

(1) Negative concept suppression is not sufficient for robot control.

In image generation, guidance steers the generated content away from undesirable concepts while balancing image quality. In robotics, guided actions instead change the environment and the states encountered at subsequent replanning steps. Moreover, a forbidden outcome may have been a successful behavior of the pretrained policy before a deployment constraint was introduced. Steering away from that behavior is useful only if the robot can still complete its task through another feasible behavior.

The relevant objective is therefore

P(task success∧constraints satisfied),P(\text{task success}\land\text{constraints satisfied}),

not forbidden-outcome suppression alone.

Our results demonstrate this distinction. When charger receptacles (1,2,3) are forbidden (Table 2), rejection sampling lowers forbidden selections from 53.5% to 12.5% but increases failures from 16.5% to 46.0%, achieving only 41.5% constraint-valid success. DNG attains 9.5% forbidden selections, 19.5% failures, and 71.0% constraint-valid success. Thus, stronger avoidance can actually produce worse closed-loop task completion.

We agree that the posterior is reinitialized at every action chunk. We do not claim to solve cross-chunk posterior consistency or provide formal safety guarantees. Our contribution is to investigate whether negative guidance preserves task competence under physical execution, feedback, and replanning—not to propose a trajectory-level inference mechanism.

(2) Beyond single-target image-generation DNG.

For multiple mutually exclusive forbidden outcomes, our joint-complement coefficient is

gi=λ0pt(ci∣xt)1−∑jpt(cj∣xt),g_i=\lambda_0\frac{p_t(c_i\mid x_t)}{1-\sum_jp_t(c_j\mid x_t)},

where the shared denominator accounts for the combined excluded probability mass; independent single-target composition instead uses 1−pt(ci∣xt)1-p_t(c_i\mid x_t). This is a principled extension of DNG, not an entirely new probabilistic identity.

Our failure-driven setting further uses deployment failures, rather than only predefined semantic concepts, to train an auxiliary negative model while retaining the frozen base policy. It does not require positively labeled replacement trajectories.

(3) Flow-matching π0\pi_0 requires a different update.

Original DNG estimates negative posteriors through stochastic reverse-diffusion transitions, which cannot be directly reused with π0\pi_0's deterministic flow. We derive the complement-velocity relation

v¬−=vb−rt1−rt(vc−−vb)v_{\neg-}=v_b-\frac{r_t}{1-r_t}(v_{c^-}-v_b)

from probability-flux decomposition and implement a state-dependent compatibility surrogate.

Although its weight is non-increasing, it is not a fixed function of time: Eq. 24 attenuates according to the local velocity disagreement ∣vc−(xt,t)−vb(xt,t)∣2|v_{c^-}(x_t,t)-v_b(x_t,t)|^2, so trajectories can follow different attenuation schedules. We acknowledge that this is a surrogate, not a calibrated posterior.

In LIBERO-Goal, π0\pi_0 avoidance of A2 reduces forbidden outcomes from 10 to 3 while retaining 18/100 failures; same-nominal-scale CFG produces 98/100 failures. This supports feasibility for flow matching, not superiority over every tuned CFG variant.

We will revise our novelty claims accordingly: our contribution is complement-based behavioral exclusion for embodied, closed-loop control, with multi-target and flow-matching developments and a failure-feedback interface—not the invention of negative guidance itself.

AuthorsOpenReview ↗

W2. Fair comparison with static and time-decayed CFG

We agree that equal nominal guidance scales do not imply equal effective strength. For single-target DNG, gt=λ0pt/(1−pt)g_t=\lambda_0 p_t/(1-p_t); with p0=0.25p_0=0.25 and λ0=3\lambda_0=3, its initial coefficient is 1, not 3. Thus, the original same-nominal-scale comparison alone cannot isolate adaptive weighting. We added smaller static-CFG coefficients and a time-decayed CFG control.

Four-object picking: complete banana-avoidance sweep. The following combines all configurations from Appendix Table 8 with our new CFG controls. Every row represents 200 rollouts; the new controls use the same environment seeds (10001–10200) as Table 8. All outcome columns are percentages of episodes; valid success is Cube + Apple + Cup.

MethodScaleCubeAppleBanana ↓CupFail ↓Valid ↑AUS ↑
DNG13123.5728.51083.811
DNG337201.532.5989.5.851
DNG538.521.52.528.5988.5.844
DNG1053.5151.524.55.593.829
CFG0.5 (new)3518.59307.583.5.793
CFG14811.53325.591.5.830
CFG1.5 (new)53.591.530.55.593.829
CFG2 (new)53.510.51.527.5791.5.812
CFG3606127693.800
CFG5656.51.522593.5.783
CFG1069.54.51.5177.591.734
CFGlinear 1→0.5 (new)43.5134.529.59.586.790

What DNG improves here is chiefly preservation of alternative successful choices—not the highest raw success rate. As CFG strengthens, it heavily favors the cube (48% at scale 1, 69.5% at scale 10) and suppresses the apple. DNG at scale 3 instead retains substantial selection of all three allowed objects (cube 37%, apple 20%, cup 32.5%), with only 1.5% forbidden banana selections. We agree that CFG can achieve higher valid success (up to 93.5% versus DNG-3's 89.5%) and will not claim otherwise.

Crucially, balanced does not necessarily mean uniform or faithful to the original policy. We therefore compute total variation (TV) distance between the successful allowed-object distribution and the unguided distribution renormalized over cube/apple/cup. The base proportions are approximately 39.9/28.7/31.5%; DNG-3 gives 41.3/22.3/36.3% (TV 0.063), while CFG-1.5 gives 57.5/9.7/32.8% (TV 0.190). DNG-3 and CFG-1.5 both select banana in 1.5% of episodes, so DNG more faithfully preserves the original allowed-mode preferences at matched observed avoidance, despite lower valid success. Weak CFG-0.5 reaches similar TV (0.065) but allows banana in 9% of episodes. The tested linear decay has AUS 0.790 and banana rate 4.5%.

Three-target charger insertion: complete matched-strength/decay sweep. We exclude receptacles (1,4,5), leaving (2,3) allowed. Three negative priors of 0.2 give the initial DNG coefficient per branch g0=λ0(0.2)/(1−3(0.2))=λ0/2g_0=\lambda_0(0.2)/(1-3(0.2))=\lambda_0/2. All rows below use the same 80 environment seeds (1–80). Receptacle entries are episode counts out of 80 (R1/R4/R5 forbidden; R2/R3 allowed), and Valid is the percentage inserted into R2 or R3. No recorded failures occurred in this subset.

MethodInitial g0g_0R1R2R3R4R5FailValid ↑\uparrow
DNG 21.017252549062.5%
DNG 31.512282659067.5%
DNG 42.011253455073.75%
CFG fixed 11.045317510025.0%
CFG fixed 1.51.54261958031.25%
CFG fixed 22.04571756030.0%
CFG fixed 33.04741829027.5%
CFG exp 1.5 to 1.31.54481576028.75%
CFG exp 1.5 to 0.751.545415511023.75%

At matched initial strength 1.5, DNG yields 54/80 valid insertions, versus 25/80 for fixed CFG and 23/80 or 19/80 for the tested decays. The fixed CFG results also concentrate heavily on forbidden R1 (42/80 at coefficient 1.5, versus 12/80 for DNG). This is one selected constraint configuration and an easy 80-seed subset, not a claim about all deployment states.

Together, these controls show that DNG's observed avoidance–distribution-fidelity advantage in picking and valid-success advantage in the tested multi-target charger setting are not reproduced by simply weakening CFG or using the tested deterministic decays. They do not establish that DNG maximizes raw success or beats every possible CFG schedule, nor do they establish matched-CFG superiority for π0\pi_0 or failure-driven tasks.

AuthorsOpenReview ↗

W3. Empirical evidence, AUS, and baseline implementation

We appreciate this concern and agree that the paper should distinguish forbidden-outcome suppression, constraint-valid completion, and preservation of allowed behaviors. DNG does not guarantee perfect avoidance, and we do not claim that it maximizes raw success in every condition.

Multi-target charger insertion. The newly constrained objective differs from the original unconstrained task: success now requires insertion into an allowed receptacle. For forbidden receptacles (1,4,5), Table 9 gives 32.5% constraint-valid success for the unguided base policy, 55.5% for DNG, and 23.5% for CFG. Across the ten three-target constraint sets, mean constraint-valid success is 33.4% (base), 58.6% (DNG), and 51.6% (CFG). These results demonstrate improvement over the constrained baseline, while the residual forbidden selections remain an important limitation. We will foreground raw outcome rates rather than suggesting complete enforcement.

Does AUS merely favor uniformity? We agree that the ideal complement distribution should preserve the relative probabilities of allowed modes, which need not be uniform. AUS deliberately measures an additional diversity preference; it is not an estimator of exact complement-distribution fidelity. Moreover, Appendix Table 7 shows that the AUS ranking depends on its weight: CFG slightly leads on average picking AUS at weight 0.1, while DNG leads at the tested weights 0.2-0.4. We will explicitly report that sensitivity.

We therefore computed a distinct fidelity metric from the existing outcome frequencies in Tables 1 and 9. Let AA be the allowed successful outcomes and let q(i)=p(i)/∑j∈Ap(j)q(i)=p(i)/\sum_{j\in A}p(j). We report

TV(qguided,qbase)=12∑i∈A∣qguided(i)−qbase(i)∣.\mathrm{TV}(q_{\mathrm{guided}},q_{\mathrm{base}})=\tfrac12\sum_{i\in A}|q_{\mathrm{guided}}(i)-q_{\mathrm{base}}(i)|.

TV rewards fidelity to the base policy's relative allowed-mode preferences, not uniformity. Averaged across the four picking constraints, DNG achieves TV 0.259 versus 0.455 for CFG (lower in all 4 cases); across ten charger constraints, DNG obtains 0.202 versus 0.355 (lower in 9/10 cases). These are post-hoc calculations from the reported distributions, not new rollouts. We will report TV together with raw valid success and failure rates: fidelity conditional on success does not itself imply high task success. The expanded banana-avoidance CFG controls additionally show that at equal 1.5% forbidden selection, DNG has TV 0.063 compared with 0.190 for fixed CFG at 1.5.

Rejection sampling and PPGuide. These baselines do not universally perform worse than the unguided policy. In Table 5, PPGuide improves all three Push-T scores: 0.4957 to 0.5450, 0.6821 to 0.7250, and 0.3083 to 0.3900. Its weaker results are concentrated in the charger deployment conditions. A plausible reason is that formerly successful insertions become invalid solely because a receptacle is unavailable, which differs from detecting ordinary execution failures. Importantly, our charger PPGuide adaptation uses a different training-data collection protocol from the original method; therefore, we cannot attribute all performance differences to the guidance algorithm. We will document the protocol and qualify comparisons accordingly.

Rejection sampling also exposes the distinction between avoidance and task completion: when (1,2,3) are forbidden (Table 2), it reduces forbidden choices to 12.5%, but failures rise to 46.0% versus 19.5% with DNG. This is why we evaluate the joint outcome of constraint satisfaction and task completion, rather than forbidden suppression alone.

We agree that evaluation-selected best scales can be optimistic. Our separate response to Q2 clarifies the selection protocol and distinguishes multiple environment seeds from independent policy-training seeds. We will revise the manuscript to present the strongest supportable interpretation: an avoidance-completion-fidelity trade-off, not unconditional superiority on any single metric.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.