acceptodds
Under review as a conference paper at ICLR 2027

Noise Space Learning for VLA Adaptation

Abstract

Pretrained vision-language-action (VLA) models encode broad manipulation priors, yet adapting them to specific tasks requires balancing limited robot interaction, computational resources, and safty constraint. Noise space learning (NSL) adapts flow-matching-based VLAs by learning a context-conditioned distribution over source noise while keeping the pretrained VLA fixed. Building on prior noise-steering methods, we formulate NSL as an interface for VLA adaptation that accommodates different learning algorithms and instantiate it with PPO, SAC, and AWR. Theoretically, we establish conditions under which NSL attains the minimum information cost of a target behavioral change. Empirically, we study these NSL algorithms on LIBERO, RoboTwin, and four real-robot manipulation tasks under interaction, computation, and execution constraints. Across these settings, we develop a suite of effective NSL variants and derive practical recommendations for choosing among them under different resource and deployment regimes.

Primary area:
applications to robotics, autonomy, planning
Keywords:
VLA, Robotics, Steering

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_kFDDOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies Noise Space Learning (NSL), which adapts a frozen flow-matching VLA by learning a context-conditioned distribution over its input noise. It implements this interface using PPO, SAC, and success-filtered AWR, and evaluates their tradeoffs across interaction and update budgets. Experiments cover LIBERO, RoboTwin, and four real-robot tasks. An information-theoretic analysis characterizes when unrestricted noise reweighting can achieve a desired behavioral change with minimum KL cost.

Critical strengths
  1. Useful empirical comparison. The paper studies how optimizer choice, update budgets, rollout counts, and training components affect noise-space adaptation.
  2. Substantial update-time savings. Unlike the previous paper, it directly measures update time against adaptation baselines.
  3. Meaningful experimental coverage. Results include two VLA backbones, two simulation benchmarks, and four real-robot tasks.
  4. Carefully qualified theoretical analysis. The appendix explicitly discusses restricted coverage, deterministic-kernel singularities, and why behavioral anchoring does not establish physical safety.
Critical weaknesses
  1. Limited statistical evidence for superiority over strong baselines. Table 4 reports one training seed, without variability across independent runs. The average improvements over RLT are modest: 2.5 percentage points for NSL-SAC on LIBERO and 1.3 points for autonomous NSL-AWR on real robots. Table 1 also reports the best checkpoint within 60 epochs, leaving the robustness of these margins uncertain.
  2. Different evaluation modes complicate the base-policy comparison. Appendix B.1 evaluates NSL using the actor’s deterministic mean, while the base policy samples Gaussian noise. A matched evaluation ablation would help isolate the contribution of learned noise adaptation.
  3. Efficiency measurements cover model updates only. Tables 1 and 3 demonstrate substantial update-time savings, but exclude the full adaptation process. End-to-end wall-clock time, including rollout collection and evaluation, is needed to establish practical speedups.
  4. The comparative safety claim has limited support. The claim of “safer execution” relies partly on an earlier SAC experiment with unintended table contact. Systematic collision or intervention rates under comparable conditions are needed to substantiate AWR’s safety advantage (Section 4.7; Appendix B.6).
Questions for authors
  1. Are the LIBERO training and evaluation reset states disjoint? Can you provide comparisons under matched stochastic and deterministic noise-selection protocols, including a zero-noise base-policy reference? Could you provide LIBERO Pro evaluation results?
  2. Does repeating a 32-dimensional noise vector across the action horizon reproduce the native VLA prior? What exactly are the noise space and reference distribution ρ\rho in the theoretical analysis and implementation?
  3. Which parts of the minimum-information result remain applicable to the restricted Gaussian actors and deterministic evaluation used experimentally? Can you clarify the practical interpretation of the comparison when alternative deterministic interventions have infinite joint KL?
  4. What quantitative evidence supports the safer-execution recommendation—for example, intervention counts or unintended-contact rates from the existing experiments under comparable conditions?
5 comments

Thank you for taking the time to review our paper in detail and share your thoughtful comments. Below, we respond to your questions and concerns with additional explanations and experimental results.

W1: Limited statistical evidence for superiority over strong baselines

[Seed] Reporting the training seed in Table 4 is for strict reproducibility. Based on our prior experiments, our method works quite robust for different seeds. Since re-running experiments under different seeds for different algorithms takes long time, we will add these results in our revision.

[Advantage] Compared with baselines such as RLT, the main advantage of NSL is training efficiency. Specifically, NSL-SAC achieves an 11.2× speedup in model training in simulation, and NSL-AWR achieves a 15.2× speedup in real-robot experiments, while achieving better performance.

[Robustness] To further assess the robustness of our results, we compare the average success rates of all five methods over the final five epochs of the 60-epoch training runs, and show the results in the table, which demonstrate that NSL-SAC consistently outperforms RLT across both models, with an average improvement of 2.82 percentage points.

ModelAction-AWRRLTNSL-PPONSL-AWRNSL-SAC
SmolVLA-Full92.56%92.24%78.68%92.96%94.80%
π₀.₅-One63.64%66.88%52.84%63.00%69.96%
Avg.78.10%79.56%65.76%77.98%82.38%

W2: Different evaluation modes complicate the base-policy comparison.

To address the reviewer’s concern, we provide a matched evaluation ablation on SmolVLA-Full and π₀.₅-One. For the untrained noise actor, deterministic evaluation at its zero mean yields performance similar to the native Gaussian reference. For the trained noise actor, stochastic and deterministic evaluation also produce similar success rates.

We compare the native Gaussian reference with zero-noise evaluation. For trained NSL-AWR policies, we compare mean and stochastic evaluation using the same best checkpoint selected by mean evaluation for each model.

ModelNative Gaussian Initial SRZero Noise Initial SRNSL-AWR Mean Evaluation SRNSL-AWR Stochastic Evaluation SR
SmolVLA-Full50.6%51.6%94.2%93.8%
π₀.₅-One40.4%40.6%64.0%64.4%

W3: Efficiency measurements cover model updates only.

To address the reviewer’s concern, we additionally report rollout collection time and evaluation time for different methods in simulation. We use model-update time for comparison in the paper because all methods share identical rollout and evaluation settings, resulting in similar time costs for these two stages. Model-update time therefore directly captures the computational differences between training methods. Including rollout collection and evaluation, NSL-SAC reduces the total time from 406.70 to 286.53 seconds per batch compared with RLT, a 29.5% reduction. NSL-AWR reduces the total time from 511.40 to 280.49 seconds per batch compared with Action-AWR, a 45.2% reduction.

MethodModel-update time (s/batch)Rollout collection time (s/batch)Evaluation time (s/batch)Total time (s/batch)
Action-AWR210.20224.4076.80511.40
RLT111.60217.2077.90406.70
NSL-PPO2.34203.4074.20279.94
NSL-AWR1.09206.3073.10280.49
NSL-SAC9.93204.2072.40286.53

W4&Q4: The comparative safety claim has limited support.

In the additional real-robot bead-placement experiment, NSL-AWR improves task success without triggering emergency stops, whereas NSL-SAC’s emergency-stop rate increases over successive update rounds. To address the reviewer’s concern, we report emergency-stop statistics during online adaptation, with 48 collection attempts scheduled per round.

MethodCompleted update roundsCollection attemptsSuccesses (SR)Emergency stops (rate)
NSL-SAC0480 (0%)0 (0%)
NSL-AWR0480 (0%)0 (0%)
NSL-SAC1482 (4.2%)17 (35.4%)
NSL-AWR1485 (10.4%)0 (0%)
NSL-SAC2230 (0%)23 (100%)
NSL-AWR2488 (16.7%)0 (0%)

Round 0 denotes initial collection before any model update. After two update rounds, all 23 NSL-SAC collection attempts triggered emergency stops, prompting early termination of the experiment.

Emergency stops include both operator-initiated stops in response to unsafe behavior and automatic stops triggered by excessive torque. Unsafe behaviors include unintended table collisions, large-amplitude hazardous motions, and knocking away beads or the bead-placement board.

Q1-1: Are the LIBERO training and evaluation reset states disjoint?

No. We use the same 50 benchmark-provided reset states per task for LIBERO training and evaluation, consistent with prior work such as SimpleVLA-RL [1] and LaST-R1 [2]. Evaluation covers all 500 task–reset pairs across the ten tasks.

For performance with disjoint reset states, we refer the readers to RoboTwin experiments in the appendix, where each task and difficulty setting uses 50 training scenarios and 50 held-out evaluation scenarios with disjoint reset seeds. All three NSL algorithms consistently improve performance under this setting. At epoch 30, averaged equally across the eight tasks and the Easy/Hard settings, NSL-PPO, NSL-AWR, and NSL-SAC improve SR over the base policy reference by 37.75, 37.13, and 48.63 percentage points, respectively.

[1] SimpleVLA-RL. SimpleVLA-RL: Scaling VLA Training via Reinforcement Learning. arXiv:2509.09674. https://arxiv.org/abs/2509.09674

[2] LaST-R1. LaST-R1: Reinforcing Robotic Manipulation via Adaptive Physical Latent Reasoning. arXiv:2604.28192. https://arxiv.org/abs/2604.28192

Q1-2: Can you provide comparisons under matched stochastic and deterministic noise-selection protocols, including a zero-noise base-policy reference?

Yes, please refer to our response to W2 for the evaluation settings and detailed results. The zero-noise base policy achieves similar performance to the native Gaussian reference, and trained NSL-AWR policies perform similarly under deterministic and stochastic evaluation.

Q1-3: Could you provide LIBERO Pro evaluation results?

Yes, we evaluate the best checkpoints selected on LIBERO-10 for the five adaptation methods on the LIBERO-10 suite of LIBERO-Pro-1.0, and present the results in the following table. We find that NSL consistently improves performance over the base policies on LIBERO-Pro, with NSL-SAC achieving the highest success rate on both models.

ModelBaseAction-AWRRLTNSL-PPONSL-AWRNSL-SAC
SmolVLA-Full10.40%21.40%20.15%17.80%22.30%24.20%
π₀.₅-One14.05%22.55%24.05%19.60%23.45%27.35%
Avg.12.23%21.98%22.10%18.70%22.88%25.78%

Q2-1: Does repeating a 32-dimensional noise vector across the action horizon reproduce the native VLA prior?

No. Native VLA noise is sampled independently across the action horizon, whereas repetition shares one 32-dimensional vector across time. Our base policy reference uses native Gaussian noise.

Q2-2: What exactly are the noise space and reference distribution (\rho) in the theoretical analysis and implementation?

The noise space contains the generator’s noise inputs, and ρ\rho is a fixed reference distribution on that space. Our theory does not require a specific form or dimensionality of ρ\rho, under the stated assumptions. The 32-dimensional implementation can be represented using a standard Gaussian reference, with repetition and fixed noise transformations included in the generator map.

Q3-1: Which parts of the minimum-information result remain applicable to the restricted Gaussian actors and deterministic evaluation used experimentally?

The information-cost lower bound remains applicable to stochastic Gaussian actors, while deterministic mean evaluation leaves the framework-level analysis unchanged. The minimum-information equality characterizes the NSL framework over unrestricted noise distributions; a Gaussian actor attains this bound when it can represent an optimal noise distribution.

Q3-2: Can you clarify the practical interpretation of the comparison when alternative deterministic interventions have infinite joint KL?

Infinite joint KL reflects a support mismatch caused by changing a deterministic generation mechanism; it does not imply infinite training cost. Our result establishes conditions under which NSL attains the behavioral information lower bound while preserving the pretrained generator. Appendix A provides a shared Gaussian smoothing formulation for finite information-cost comparisons at a specified scale. Computational efficiency is evaluated separately through measured training times.

Thank you again for your careful review and valuable suggestions. We hope our responses can clarify the points raised and address your concerns.

Thank you again for your thoughtful review and feedback. As the discussion period is approaching its end, we wanted to follow up and ask whether our response has addressed your main concerns. We sincerely appreciate the time and care you have given to reviewing our work.

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

This paper studies NSL for adapting pretrained flow-matching VLA models. The method keeps the VLA frozen and learns a context-conditioned noise actor to adjust the input noise distribution. The authors investigate three optimization methods, PPO, SAC, and AWR, and provide a theoretical analysis based on information cost. Experiments on LIBERO, RoboTwin, and four real-robot manipulation tasks show that NSL improves task success rates while reducing model-update cost. NSL-SAC achieves the best performance in simulation, while NSL-AWR is more suitable for settings with limited interaction and compute budgets.

Critical strengths

Training a lightweight noise actor avoids backpropagation through the large VLA, significantly reducing model-update cost. The real-robot experiments show that NSL-AWR improves task success rates with much lower update time than the compared baselines, which is useful for robot adaptation with limited computational resources.

Critical weaknesses

My main concern is the limited evidence supporting the safety comparison between NSL-AWR and NSL-SAC. The paper suggests that NSL-AWR is more suitable for real-robot adaptation, while NSL-SAC may produce unsafe motions. However, this conclusion is mainly based on limited real-robot SAC observations and noise distribution analysis. Without a systematic safety comparison under similar conditions, it is unclear whether NSL-AWR is consistently safer than NSL-SAC.

Questions for authors

Can the authors provide a more systematic safety comparison between NSL-AWR and NSL-SAC on real robots? For example, can they report the frequency of unsafe motions, safety interventions, or emergency stops under comparable training budgets and execution constraints?

3 comments

Evidence supporting the safety comparison between NSL-AWR and NSL-SAC

Thanks for pointing this out. We provide more evidence to support our safety related arguments as follows.

[Experimental evidence] Our main observation for the bead-placement task is that NSL-AWR progressively improves task success without triggering emergency stops, whereas NSL-SAC exhibits an increasing emergency-stop rate over successive update rounds. To support this conclusion, we provide additional emergency-stop statistics during real-robot online adaptation. Each round schedules 48 collection attempts. After two update rounds, all 23 NSL-SAC collection attempts triggered emergency stops, prompting early termination. Detailed results are provided below.

Round 0 denotes initial collection before any model update. Emergency stops include both operator-initiated stops in response to unsafe behavior and automatic stops triggered by excessive torque. Unsafe behaviors include unintended table collisions, large-amplitude hazardous motions, and knocking away beads or the bead-placement board.

MethodCompleted update roundsCollection attemptsSuccesses (SR)Emergency stops (rate)
NSL-SAC0480 (0%)0 (0%)
NSL-AWR0480 (0%)0 (0%)
NSL-SAC1482 (4.2%)17 (35.4%)
NSL-AWR1485 (10.4%)0 (0%)
NSL-SAC2230 (0%)23 (100%)
NSL-AWR2488 (16.7%)0 (0%)

[Mechanism evidence] Unusually large noise values lie in the tails of the standard Gaussian prior used during VLA training, where limited training coverage can lead to unreliable or unsafe motions. NSL-SAC can push noise toward these regions by maximizing overestimated critic values. Our real-robot diagnostics (Section 4.7 and Appendix B.6) show noise approaching the amplitude limit, higher predicted Q values and greater twin-critic disagreement for actor proposals than replay noise, and unintended table contact despite the limit. Bounding noise does not directly constrain decoded motions or contact forces. NSL-AWR grounds its updates in observed successful noise samples through weighted regression, avoiding direct maximization of potentially overestimated critic values.

Thank you once again for your time and constructive feedback. We hope our explanations and additional evidence can address your concerns.

Thank you again for taking the time to review our work and share your feedback. With the discussion period coming to a close, we wanted to follow up and ask whether our response has helped address your main concerns. Thank you very much for your care and consideration.

Thank you very much for taking the time to consider our response. We are glad that the additional emergency-stop results helped address your concern. We wish you all the best.

Reviewer_9YgcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies Noise Space Learning (NSL), which adapts a frozen flow-matching vision-language-action model by learning a context-conditioned distribution over its input noise. It implements this interface using PPO, SAC, and a success-filtered variant of advantage-weighted regression (AWR), without backpropagating through the VLA.

Critical strengths
  • Table 1 reports competitive success rates with considerably cheaper gradient updates.
  • The real-robot experiments show substantial improvement over the initial policies at low model-update cost.
  • The rollout-count, update-budget, and component ablations examine when different NSL implementations work well.
  • The critic and noise-drift diagnostics help identify potential failure mechanisms beyond aggregate success rates.
Critical weaknesses
  • The incremental contribution over prior noise-steering methods is insufficiently isolated. DSRL already formulates frozen-generator adaptation through noise-space RL, permits different RL algorithms, evaluates SAC with the flow-based π0 model, and repeats lower-dimensional noise across the action horizon. Thus, applying the interface to flow-based VLAs is already established.[1] Human-action inversion and supervision of the noise actor are also established in UniSteer[2], whose inversion procedure this paper adopts. The authors acknowledge these precedents, but stronger attribution experiments are needed to identify the contribution of their particular recipes, especially advantage weighting versus unweighted regression on successful noise samples.
  • The evaluation does not yet establish the reliability of the comparative conclusions. Appendix B.1 states that NSL uses a 32-dimensional noise vector repeated across the prediction horizon and deterministic actor-mean evaluation, whereas the base reference uses native Gaussian sampling. Real-robot comparisons have similar uncertainty: the autonomous AWR–RLT difference corresponds to only two additional successes across 160 attempts per method. The substantial gains over the initial policies are encouraging, but finer method rankings remain insufficiently supported.

[1] A. Wagenmaker, M. Nakamoto, Y. Zhang, S. Park, W. Yagoub, A. Nagabandi, A. Gupta, and S. Levine, “Steering Your Diffusion Policy with Latent Space Reinforcement Learning,” arXiv preprint arXiv:2506.15799, 2025. doi: 10.48550/arXiv.2506.15799.

[2] J. Lu, X. Qin, Y. Jiang, K. Wang, C. Zhang, B. Liang, J. Yang, M. Xu, and L. Zhao, “UniSteer: Unified Noise Steering for Efficient Human-Guided VLA Adaptation,” arXiv preprint arXiv:2605.10821, 2026. doi: 10.48550/arXiv.2605.10821.

Questions for authors
  • Does the base reference use the same temporally repeated noise construction as NSL?
  • What performance is obtained with an untrained zero-mean noise actor, and when trained NSL policies are evaluated stochastically?
  • How does NSL-AWR compare with unweighted maximum likelihood on exactly the same successful replay samples?
  • How does NSL-SAC differ empirically from a matched DSRL-SAC implementation?
3 comments

We sincerely thank the reviewer for the time and effort devoted to reviewing our paper and for the thoughtful comments and constructive suggestions. Below, we respond to each comment and question point by point.

W1: Incremental contributions

Building on prior noise-space RL methods, our work focuses on systematically investigating the theoretical properties, optimization strategies, and practical design choices for efficient VLA adaptation. Our contributions are three-fold, each of which clearly distinguishes our work from prior ones. In addition, recent popularity of this learning paradigm (e.g., [1-4]) also supports the significance of our contributions. Specifically:

  1. Theoretical formulation and analysis. We develop an information-theoretic formulation that compares noise adaptation with generator and residual adaptation. Our analysis establishes the information-theoretic advantages of NSL under the stated assumptions.
  2. Systematic comparison across resource budgets. Unlike prior work (e.g., [1]) adopts only SAC as the RL algorithm, we show other RL algorithms such as PPO and AWR could also be effective. Further, we compare these variants (PPO, SAC, and AWR) under different rollout and computation budgets. AWR achieves strong performance when budget is limited, whereas SAC benefits more from larger budget. Our real-robot experiments further demonstrate AWR’s effectiveness in real-world online adaptation.
  3. Substantial engineering improvements. We develop techniques to improve the performance of different NSL variants, such as task balancing and success/failure sampling balancing. Removing these components individually reduces SAC’s SR within 30 epochs by 9.5 and 6.5 percentage respectively, and PPO’s by 7.9 and 9.0 percentage respectively (see Table 2).

[1] DSRL. Steering Your Diffusion Policy with Latent Space Reinforcement Learning. arXiv:2506.15799. https://arxiv.org/abs/2506.15799

[2] UniSteer. UniSteer: Unified Noise Steering for Efficient Human-Guided VLA Adaptation. arXiv:2605.10821. https://arxiv.org/abs/2605.10821

[3] FlowDAgger. FlowDAgger: Human-in-the-Loop Adaptation of Generative Robot Policies in Latent Space. arXiv:2607.08877. https://arxiv.org/abs/2607.08877

[4] FRS / DSBC. Improving Robotic Generalist Policies via Flow Reversal Steering. arXiv:2606.13675. https://arxiv.org/abs/2606.13675

W2: Reliability in evaluation

In simulation, we choose the evaluation protocol that works better for the baseline. Specifically, we evaluate the base policy using native Gaussian noise to provide a strong reference because it yields higher success rates than repeated noise. For trained NSL policies, we use the actor mean to remove randomness introduced by noise sampling during evaluation. Results for the repeated-noise base policy and stochastic evaluation of trained NSL-AWR policies are provided in our responses to Q1 and Q2.

On real robots, the clear advantage of NSL-AWR compared with RLT is its efficiency and compatibility with human guidance. NSL-AWR achieves 15.2× training speedup under the same computational resources while achieving a slightly higher success rate than RLT. Incorporating human guidance further improves NSL-AWR’s performance, achieving a mean success rate of 83.8%, exceeding autonomous RLT by 13.8 percentage points.

Q1: Does the base reference use the same temporally repeated noise construction as NSL?

No. We use native standard Gaussian noise for the base-policy reference since it works better than tiled/repeated Gaussian noise.

ModelNative Gaussian Noise Initial SRRepeated Gaussian Noise Initial SR
SmolVLA-Full50.6%31.8%
π₀.₅-One40.4%40.0%

Q2: What performance is obtained with an untrained zero-mean noise actor, and when trained NSL policies are evaluated stochastically?

Stochasticity in evaluation does not lead to significant performance change. We evaluate with different protocols on SmolVLA-Full and π₀.₅-One respectively and present the results as follows.

ModelNative Gaussian Initial SRZero Noise Initial SRNSL-AWR Mean Evaluation SRNSL-AWR Stochastic Evaluation SR
SmolVLA-Full50.6%51.6%94.2%93.8%
π₀.₅-One40.4%40.6%64.0%64.4%

Q3: How does NSL-AWR compare with unweighted maximum likelihood on exactly the same successful replay samples?

NSL-AWR outperforms unweighted maximum likelihood on both evaluated models. We compare the two methods on successful replay samples using SmolVLA-Full and π₀.₅-One over 30 training epochs. The table below reports success rates at epochs 10, 20, and 30, together with the best success rate within this training budget.

ModelMethodEpoch 10 SREpoch 20 SREpoch 30 SRBest SR (E1–E30)
SmolVLA-FullUnweighted Maximum Likelihood78.6%81.0%82.6%82.6%
SmolVLA-FullNSL-AWR80.6%84.0%87.2%87.2%
π₀.₅-OneUnweighted Maximum Likelihood52.8%53.0%54.0%54.4%
π₀.₅-OneNSL-AWR53.6%55.4%57.2%59.0%

Q4: How does NSL-SAC differ empirically from a matched DSRL-SAC implementation?

Compared with DSRL-SAC, NSL-SAC incorporates two additional components: task balancing and success/failure sampling balancing. We also reuse pooled hidden features from the frozen VLA instead of the separate CNN visual feature extractor used in DSRL-SAC. Our ablation study in Table 2 shows that removing task balancing reduces the two-model average best SR within 30 epochs by 9.5 percentage points, while removing success/failure sampling balancing reduces it by 6.5 percentage points.

Thank you again for your thoughtful feedback. We hope our responses can address your questions and concerns.

Thank you again for your time and thoughtful feedback. As the discussion period is coming to an end, we wanted to follow up and ask whether our response has addressed your main concerns. We sincerely appreciate your time and consideration.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response and Summary of Revisions

We sincerely thank all reviewers and the Area Chair for their time, thoughtful feedback, and constructive suggestions throughout the discussion.

During the rebuttal period, we provided additional experimental results and clarifications on the following aspects:

  1. Contributions and methodology: We further clarified NSL's theoretical formulation, systematic comparison of optimization algorithms, and practical engineering improvements. We also highlighted the differences between NSL-SAC and DSRL-SAC, including task balancing, success/failure sampling balancing, and the use of pretrained VLA features, supported by component ablations.

  2. Theoretical analysis: We further clarified the scope and practical interpretation of our information-theoretic results, including their applicability to restricted Gaussian actors and deterministic evaluation.

  3. Additional experimental evaluations: We provided results on matched noise evaluation protocols, weighted versus unweighted regression, final-five-epoch average success rates, LIBERO-Pro generalization, and held-out evaluation scenarios in RoboTwin.

  4. Computational efficiency: We extended our efficiency analysis beyond model updates to include rollout collection and evaluation, providing total wall-clock time comparisons.

  5. Real-robot experiments: We provided additional emergency-stop statistics comparing NSL-AWR and NSL-SAC, alongside further clarification of real-world adaptation efficiency and the benefits of human guidance.

We greatly appreciate the reviewers' valuable feedback and will incorporate these additional results and clarifications into the revised manuscript.

Thank you again for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.