acceptodds
Under review as a conference paper at ICLR 2027

ProactiveVLA: Augmenting Embodied Memory through Proactive Environment Exploration

Abstract

Rapid adaptation to a new environment requires a robot to acquire useful knowledge about local objects, states, and interactions from limited experience. Systems that use language models to organize perception, planning, and control through executable robot interfaces can adapt through execution feedback and memory, making the choice of experience central to their effectiveness. Repeated practice of a target task may refine a familiar solution while leaving other interactions relevant to changed conditions untested. We introduce ProactiveVLA, which uses proactive environment exploration to acquire reusable knowledge for deployment-time adaptation. After completing an initial task, the agent allocates the remaining interaction budget to self-proposed goals covering object affordances, state-changing interactions, and compositions of interactions. It verifies execution outcomes and consolidates both task-directed and exploratory experience into memory that guides subsequent planning and control. ProactiveVLA outperforms the baselines under the same turn budget on LIBERO-Pro and RoboCasa365 Composite-Seen. On LIBERO-Pro Goal-T, with at most one VLA primitive invocation allowed during evaluation, ProactiveVLA completes 48% of instances, compared with 19% for the Harness VLA task-refinement baseline.

Primary area:
applications to robotics, autonomy, planning
Keywords:
vision-language-action models, embodied AI, autonomous exploration

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_G6amOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies rapid deployment-time adaptation of robots to new environments. The authors argue that repeatedly refining a provided reference task may concentrate experience around a single solution and fail to acquire knowledge useful under changed conditions. To address this, the paper introduces ProactiveVLA, which allocates part of the limited interaction budget to proactive environment exploration and acquires reusable knowledge about object affordances, state-changing interactions, and interaction compositions. The acquired experience is then consolidated into memory to guide subsequent task execution.

Critical strengths
  1. Compared with Harness VLA, ProactiveVLA extends deployment-time adaptation from pure task refinement to proactive environment exploration, and combines visual verification with memory consolidation to reuse the acquired experience in subsequent tasks. The overall motivation is intuitive and well aligned with the goal of rapid adaptation to new environments.

  2. The method is evaluated on LIBERO-Pro and RoboCasa365 Composite-Seen. On LIBERO-Pro, ProactiveVLA improves over Harness VLA in seven of the eight settings, increasing the overall success rate from 72.1% to 76.4%. In addition, on LIBERO-Pro Goal-T, when at most one VLA primitive invocation is allowed during evaluation, ProactiveVLA achieves 48% success compared with 19% for Harness VLA, an improvement of 29 percentage points.

Critical weaknesses
  1. Presentation and clarity

1.1. The adaptation-budget protocol appears internally inconsistent. Sec. 3.1 suggests that a fixed b_init turns are allocated to the provided task, after which the remaining (B - b_init) turns are used for exploration. However, Algorithm 1 does not explicitly contain b_init, while the implementation section states that exploration uses the remaining budget “after task success.” It should be clarified whether exploration begins after a fixed number of initial turns or only after successful task completion, and what happens if the reference task cannot be completed within the budget.

1.2. The variable e_i in Algorithm 1 is not clearly defined. The paper should explicitly state whether it represents a scene, an environment instance, or another entity.

1.3. Figure 3 is difficult to interpret. The task instruction and target object are not clearly shown, and the localization outputs are not visualized. The key distinction between “RE-LOCALIZE,” “NO RE-LOCALIZE,” and “WRONG POSITION” is mainly conveyed through textual annotations rather than observable evidence. Therefore, the figure does not convincingly demonstrate that an exploration-derived recovery rule caused the successful behavior. Showing the retrieved memory rule, localization/tool traces, and the supporting adaptation evidence would make the analysis more convincing.

1.4. The auxiliary model configuration is insufficiently specified. Codex is identified as the high-level agent, but it is unclear whether the language-model proposer uses exactly the same model/configuration. More importantly, the specific model used as the independent visual checker is not identified.

  1. Experimental validation

2.1. The sets of comparison methods differ substantially across the two benchmarks. This is not necessarily inappropriate, but the paper should better explain which baselines are available or applicable to each benchmark and why some related agentic methods are absent from the RoboCasa365 comparison.

2.2. The statistical support on RoboCasa365 is limited. GR00T N1.5 and the StarVLA variants are evaluated with 30 or 50 rollouts per task, whereas Harness VLA and ProactiveVLA use only five. Although the main Harness VLA vs. ProactiveVLA comparison is matched, the resulting estimates have substantially different statistical precision across methods.

2.3. Five rollouts per task make the per-task success rates very coarse. A single additional success changes the reported rate by 20 percentage points. Therefore, improvements such as 20%→40% correspond to only one additional successful rollout. More rollouts, confidence intervals, or statistical tests would be needed to support fine-grained per-task conclusions.

2.4. The current experiments do not fully establish that the improvement specifically comes from proactively acquired environment knowledge. ProactiveVLA simultaneously changes the exploration prompt, memory organization, high-level task decomposition, and tool-use behavior. The paper itself shows that ProactiveVLA produces substantially more fine-grained action-level VLA instructions and uses analytic control more frequently. A matched baseline such as Harness VLA with explicitly enforced fine-grained action-level decomposition/prompting would help disentangle the effect of proactive environment knowledge from improved prompting and tool coordination.

2.5. It would also be valuable to study whether the advantage of proactive exploration persists as ordinary task-execution experience accumulates over time, for example by comparing task-only continual memory with task memory + proactive exploration. This would clarify whether the method primarily benefits cold-start rapid adaptation or remains useful during longer-term deployment.

  1. Novelty

3.1. ProactiveVLA is largely instantiated on top of the Harness VLA agentic execution stack, with the main additions being proactive exploration, three-level exploration objectives, visual verification, and global memory consolidation. The system is reasonably structured, but the technical novelty appears relatively incremental. In particular, the ablation results show only modest drops when removing initial-task guidance or memory management, from 71% to 68% and 67%, respectively, whereas removing the three-level exploration structure reduces performance to 42%. This suggests that a substantial portion of the gain comes from the manually specified three-level exploration scaffold/prompt. Since the exploration categories, ordering, and minimum exploration requirements are explicitly encoded in the prompt, it remains unclear how much of the contribution should be regarded as a new algorithmic mechanism rather than carefully engineered prompting and memory organization.

Questions for authors

For the submission, the following should be addressed:

  1. For presentation and clarity, the issues summarized in the Weaknesses section should be clarified, particularly the adaptation-budget protocol, notation in Algorithm 1, interpretation of Figure 3, and the auxiliary model configurations.

  2. The evaluation protocol on RoboCasa365 would benefit from further clarification and stronger statistical support. The compared methods use substantially different numbers of rollouts, while Harness VLA and ProactiveVLA are evaluated with only five rollouts per task, meaning that a single additional success changes the reported success rate by 20 percentage points. Could the authors explain the choice of evaluation protocol and provide more rollouts, confidence intervals, or statistical tests to support the per-task conclusions?

  3. The current results do not fully disentangle whether the improvements come from proactively acquired environment knowledge or from changes in prompting, task decomposition, and tool usage. Since ProactiveVLA produces more fine-grained action-level VLA instructions and uses analytic control more frequently, it would be helpful to include a matched Harness VLA baseline with similarly fine-grained action-level decomposition or prompting.

  4. It would also be valuable to study whether the advantage of proactive exploration persists as ordinary task-execution experience accumulates over time. For example, how does task-only continual memory compare with task memory + proactive exploration? Such an experiment could clarify whether the proposed method mainly benefits cold-start rapid adaptation or remains useful during longer-term deployment.

24 comments
AuthorsOpenReview ↗

1. The adaptation-budget protocol and when exploration begins (Weakness 1.1)

We sincerely thank you for your effort in reviewing our paper. We sincerely thank you for the detailed review and for recognizing the motivation of proactive exploration and the gains under limited VLA calls. During the discussion period, we revised the method description and recovery figure to improve clarity, and conducted additional experiments, including a real-robot comparison, evaluations with an alternative planner and VLA controller, and an expanded 800-rollout RoboCasa365 evaluation. We further introduced a matched memory-processing baseline, repeated memory acquisition, and pooled-memory evaluation to better characterize the contribution of proactive exploration in the revised paper. We address each point below.

Total budget for adaptation. The adaptation budgets are:

BudgetPlanner requestsRobot-action calls
Total adaptation per reference task9648
Each initial-task attemptAt most 24At most 12

Each attempt in adaptation uses min(24, remaining_turns) and min(12, remaining_actions). Actual consumption is deducted from the shared totals, so the procedure is not equivalent to exactly four attempts. Following an unsuccessful attempt, the agent resets and retries the provided task while both total allowances remain positive.

When exploration begins. After the first success, ProactiveVLA resets once to the seed-0 initial state and enters autonomous exploration if at least 4 requests and 1 action call remain. Exploration receives only the unused allowance. HarnessVLA continues refining the provided objective under the same total limits.

AuthorsOpenReview ↗

2. What happens if the reference task cannot be completed within the budget? (Weakness 1.1, Question 1)

Thank you for raising this question. In practice, the initial task is successfully completed in the vast majority of adaptation instances, as the allocated interaction budget allows multiple attempts and recovery from execution failures. Therefore, ProactiveVLA can typically proceed to proactive exploration as intended.

For the small fraction (about 2.5%) of cases where the initial task cannot be completed within the budget, the run does not enter autonomous exploration, but the agent still preserves the failed attempts and their execution feedback as negative experience for subsequent memory construction. These records provide useful information about unsuccessful interactions and physical constraints, even without a completed reference task. Importantly, all attempts share the same interaction budget, ensuring a consistent evaluation setting.

We will clarify this behavior and the transition to proactive exploration in the revised paper. As future work, we plan to investigate more adaptive strategies for cases where the initial task cannot be completed, such as targeted failure diagnosis and exploration aimed at identifying the underlying physical constraints and possible recovery strategies.

AuthorsOpenReview ↗

3. Presentation and clarity (Weakness 1.2, Weakness 1.3, Weakness 1.4, Question 1)

Notation of the variable e_i in Algorithm. We will define e_i as the reference environment instance, including its scene and initialization, paired with the provided task instruction g_i.

Presentation of Figure 3. We sincerely thank you for the specific suggestions and have revised the paper (now it's Figure 4). The key decision is what happens after the manipulator is moved away: ProactiveVLA localizes the target again before continuing, whereas HarnessVLA proceeds without that update in the illustrated execution. We have revised our paper to make this comparison directly inspectable by showing the task instruction and target, the localization outputs at that decision point, the retrieved memory rule, and its supporting adaptation evidence.

Report of the auxiliary model configuration. Thanks for your feedback, we have carefully revised our paper to include them. The task proposer, visual checker, and memory-processing modules use GPT-5.6 Sol with high reasoning effort, with verification and evidence review performed in separate contexts. For VLA control, we use the RLinf-released π0.5-SFT checkpoint on LIBERO-Pro and RLDX-1 on RoboCasa365.

Reviewer_9S6sOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

ProactiveVLA studies how a robot should use a limited period of practice before evaluation. After attempting the given task, it explores other goals involving individual objects, changes to the scene, and combinations of actions. It checks the outcomes and turns the experience into memory. This memory and the VLA remain fixed during evaluation. Compared with Harness VLA, success rises from 72.1% to 76.4% on LIBERO-Pro and from 30.0% to 41.3% on RoboCasa365 Composite-Seen.

Critical strengths
  1. The paper asks a useful question: should an agent keep practicing the given task, or spend some of its remaining budget exploring related interactions? ProactiveVLA is built on the Harness VLA execution stack, and the paper states that Harness VLA was run with the same Codex planner, which makes this a relevant comparison. However, Section 4.1 also says the baselines use their publicly reported results, so the authors should confirm that the Harness VLA results in Table 1 were re-run with this planner. Freezing the acquired memory before evaluation also helps separate preparation from later task execution.

  2. Table 3 gives useful evidence that the choice of exploration goals matters. On Goal-S, “w/o three-level exploration” scores 42%, compared with 71% for the full method. This variant keeps the interaction budget and memory-processing pipeline fixed. It therefore provides fairly direct support for structured goal generation over free-form exploration. Improvements in seven of eight LIBERO-Pro settings provide further support for the approach.

Critical weaknesses
  1. The budget needs a clearer definition. The paper emphasizes equal adaptation budgets, but does not give the actual adaptation-turn and action limits. The appendix leaves these as {max_turns} and {max_actions}. The only concrete figure, 30 agent turns per rollout on RoboCasa365, is an evaluation-time budget. The exploration prompt also requires at least three Level-1 tasks, two Level-2 tasks, and one Level-3 task, with at least three action attempts per task. This implies a minimum budget that the paper does not reconcile with B. It is also unclear how proposer and visual-checker calls are counted. The one-VLA-call result is useful, but measures only one part of execution cost. Other control and perception tools remain available. Reporting total robot actions, model usage, and elapsed time would give a fuller picture of efficiency.

  2. The main comparison does not fully separate exploration from memory processing. ProactiveVLA adds a substantial pipeline for organizing and reviewing experience. The paper does not show whether applying that same pipeline to Harness VLA’s task-refinement records would recover some of the gain. The component tests are also limited to Goal-S. Their large exploration effect is encouraging, but the smaller gains from initial task guidance and memory management need uncertainty estimates. Goal-S contains 100 rollouts, so the 71%, 68%, and 67% results differ by only three or four successful rollouts. The same caution applies to RoboCasa365: with five rollouts per task, the gain from 30.0% to 41.3% corresponds to 24 versus 33 successes out of 80. Each memory is built from a single reference instance per task, using only seed-0 evidence. Repeating memory acquisition would help show whether the results depend on a particular exploration run. The drop on L10-T, from 52% for Harness VLA to 44%, also deserves analysis.

Questions for authors
  1. What are the exact adaptation-turn and action limits? How are proposer, checker, and memory-processing calls counted? Section 3.1 says exploration begins after some attempts at the initial task, while Section 4.1, the abstract, and the provided-task memory prompt assume that the task was completed. What happens if the initial task is not completed?

  2. How does task refinement perform when its records are processed through the same memory pipeline as ProactiveVLA? Were the Harness VLA results in Table 1 rerun with the same Codex planner, or taken from the published Harness VLA results?

5 comments
AuthorsOpenReview ↗

1. A clearer definition of the budget (Weakness 1, Question 1)

We sincerely thank you for your effort in reviewing our paper. We appreciate your recognition of the preparation-budget question and of the controlled evidence for three-level exploration. Your comments help us distinguish the contribution of exploration from memory processing and clarify the experimental settings. During the discussion period, we conducted several additional experiments, including a matched memory-processing baseline, three independent memory-acquisition runs, an experiment pooling memories across 80 tasks, and an expanded 800-rollout RoboCasa365 evaluation. We also added real-robot experiments and an alternative agent configuration, and clarified the adaptation budgets, model configurations, and elapsed time in the revised paper. We address each point below.

Total budget of requests and robot action calls for adaptation. The adaptation budgets are:

BudgetPlanner requestsRobot action calls
Total adaptation per reference task9648
Each initial-task attemptAt most 24At most 12

Each attempt in adaptation uses min(24, remaining_turns) and min(12, remaining_actions). Actual consumption is deducted from the shared totals, so the procedure is not equivalent to exactly four attempts. Following an unsuccessful attempt, the agent resets and retries the provided task while both total allowances remain positive.

Minimum budget during exploration. After the first success, ProactiveVLA resets once to the seed-0 initial state and enters autonomous exploration if at least 4 requests and 1 action call remain. Exploration terminates when either budget is exhausted, even if the requirements in the prompt remain unmet. Exploration receives only the unused allowance. HarnessVLA continues refining the provided objective under the same total limits.

Proposer, visual-checker and memory-processing calls. Proposer, visual-checker and memory-processing calls are counted in total requests.

Report of model usage and elapsed time. Thanks for your feedback, we have carefully revised our paper to include them. The planner, task proposer, visual checker, and memory-processing modules use GPT-5.6 Sol with high reasoning effort, with verification and evidence review performed in separate contexts. For VLA control, we use the RLinf-released π0.5-SFT checkpoint on LIBERO-Pro and RLDX-1 on RoboCasa365.

We report the average wall-clock time per rollout on RoboCasa365 Composite-Seen:

ComponentTime (s)
VLA tool execution8.39
Other robot tools10.70
LLM, observations, SAM3, memory, and finalization772.11
Total791.21

Most of the elapsed time is spent on model inference, perception, and system processing, while robot tool execution accounts for a relatively small fraction. These results provide a more complete picture of the computational and execution costs of our framework. All values are rounded independently.

Also, we added the average evaluation-time cost per rollout for both methods on Goal-S:

Metric (per rollout)Harness VLAProactiveVLA
Model requests28.824.8
Robot action calls15.213.0
Tool calls in total73.260.0
Input tokens1.88M1.47M
Output tokens68.0k62.0k
Elapsed time788 s749 s

ProactiveVLA completes more tasks with fewer requests, actions, tokens, and less time per rollout. The injected memory lengthens each prompt, but execution reaches completion in fewer steps, so total usage is lower.

AuthorsOpenReview ↗

2. Separating exploration from memory processing, more rollouts, uncertainty estimates, etc. (Weakness 2, Question 2)

Thank you for distinguishing the contribution of exploration from memory processing. We apply the same memory management pipeline to Harness VLA’s task-refinement:

MethodGoal-T
HarnessVLA81.0%
HarnessVLA + Memory Pipeline83.0%
ProactiveVLA88.0%

Improved memory organization alone cannot fully explain the gains of ProactiveVLA. Importantly, memory processing and experience acquisition address different aspects of adaptation. Applying a stronger memory pipeline can better organize existing task-refinement records, but it cannot provide direct interaction evidence about object affordances, state changes, or compositions that were never explored. ProactiveVLA explicitly allocates its remaining preparation budget to acquiring such additional evidence.

Uncertainty estimates of initial task guidance and memory management. Thank you for pointing this out. Each LIBERO-Pro setting is evaluated over 100 rollouts (10 tasks × 10 seeds), providing a substantial evaluation budget. While the ablation results consistently favor the full method, we acknowledge that the modest 3–4% differences do not, on their own, establish statistically significant improvements. We therefore interpret these results as evidence of positive contributions rather than conclusive estimates of their individual effects. We have carefully revised our paper to include this.

The component tests in Goal-S. Thank you for this valuable suggestion. Due to the limited rebuttal period, we were unable to extend the component ablations beyond Goal-S, but we will update our submission accordingly.

Repeating memory acquisition. Thank you for this helpful suggestion. We conducted three separate memory-acquisition runs for both HarnessVLA and ProactiveVLA on LIBERO-Pro Goal-T and evaluated the resulting memories. The success rates are reported below.

MethodRun 1Run 2Run 3Average
HarnessVLA81.0%83.0%78.0%80.7%
ProactiveVLA88.0%87.0%85.0%86.7%

ProactiveVLA consistently outperforms HarnessVLA across all three runs, with an average gain of 6.0 percentage points. These results suggest that the benefit is not driven by a single favorable exploration run.

Evaluation with more rollouts for RoboCasa365. Thank you for highlighting the importance of evaluation precision. We conducted an expanded RoboCasa365 Composite Seen evaluation with 800 rollouts for Harness VLA and ProactiveVLA (50 rollouts per task, therefore more comparable to other baselines), using RLDX-1 as the controller. On expanded evaluations, ProactiveVLA outperforms existing baseline methods:

MethodSuccess rate95% CI
Harness VLA32.9%29.7–36.2%
ProactiveVLA40.9%37.5–44.3%

The 8.0-point difference has a 95% CI of 3.3–12.7 points (two-proportion z = 3.32, p < 0.001).

Analysis of the drop on L10-T. Thank you for highlighting this result. Our inspection identified a failure case in which the bottle’s placement inside the drawer prevented subsequent closure. This illustrates that placement must satisfy the conditions required for the next action, beyond getting the object into the container.

Under the same interaction budget, HarnessVLA’s repeated task refinement provides more opportunities to reinforce such specific constraints. ProactiveVLA acquires broader interaction knowledge, but even its compositional exploration may not revisit each difficult transition sufficiently. This trade-off between exploration breadth and task-specific refinement may contribute to the lower L10-T performance, highlighting the need to balance broader exploration with additional practice of difficult transitions.

Thanks for your feedback again, we have carefully revised our paper to include them.

AuthorsOpenReview ↗

3. What happens if the initial task is not completed? (Question 1)

Thank you for raising this question. In practice, the initial task is successfully completed in the vast majority of adaptation instances, as the allocated interaction budget allows multiple attempts and recovery from execution failures. Therefore, ProactiveVLA can typically proceed to proactive exploration as intended.

For the small fraction (about 2.5%) of cases where the initial task cannot be completed within the budget, the run does not enter autonomous exploration, but the agent still preserves the failed attempts and their execution feedback as negative experience for subsequent memory construction. These records provide useful information about unsuccessful interactions and physical constraints, even without a completed reference task. Importantly, all attempts share the same interaction budget, ensuring a consistent evaluation setting.

We will clarify this behavior and the transition to proactive exploration in the revised paper.

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

ProactiveVLA uses a limited interaction budget to gather reusable environment knowledge. After attempting a provided task, it explores object affordances, state-changing interactions, and interaction compositions; it then verifies outcomes and consolidates the experience into memory for later execution. The paper reports gains over Harness VLA on LIBERO-Pro and RoboCasa365 Composite-Seen, including 48% versus 19% success on LIBERO-Pro Goal-T when evaluation permits at most one VLA primitive invocation.

Critical strengths
  1. The paper compares proactive exploration with task refinement under a shared agentic execution framework and interaction-budget framing. Its ablations suggest that structured three-level exploration matters: removing it substantially lowers Goal-S performance.
  2. The low-VLA-call analysis tests whether acquired memory changes later execution behavior, rather than measuring only eventual task success. This makes the proposed adaptation objective concrete.
Critical weaknesses
  1. The method is evaluated only in simulation and on a single agentic VLA execution stack. The paper therefore does not establish whether its memory and exploration strategy transfers to real robots or different planners and controllers.
  2. RoboCasa results use five evaluation rollouts per task, while the reported external baselines use substantially more rollouts. The small sample makes the per-task and aggregate comparisons noisy, and the paper does not report uncertainty estimates for them.
Questions for authors
  1. What are the per-condition adaptation budgets and rollout counts for the matched ProactiveVLA versus Harness VLA comparisons, and are uncertainty estimates available for the reported gains?
  2. Is the memory used at evaluation built only from adaptation evidence that is independent of the evaluation seeds and task instances?
6 comments
AuthorsOpenReview ↗

1. Real-world evaluation (Weakness 1)

We greatly appreciate your valuable comments and insightful feedback. Also, thank you for recognizing the value of comparing proactive exploration with task refinement and of examining execution under limited VLA calls. During the discussion period, we conducted additional experiments to strengthen the validation of our method, including a quantitative real-robot comparison, evaluations with an alternative planner and VLA controller, and an expanded RoboCasa365 evaluation with 800 rollouts per method. We also clarified the adaptation budgets and the independence of memory acquisition from evaluation in the revised paper. We address each point below.

We have added a quantitative real-robot comparison. On the ARX dual-arm platform, ProactiveVLA achieves 80.0% successes, compared with 56.7% for Harness VLA. We evaluate two atomic tasks and one composition task, each with 10 epsiodes (randomized initialization):

TaskHarnessVLA success rateProactiveVLA success rate
Place the pink jar inside the red metal box80.0%90.0%
Place the red metal box on the plate60.0%80.0%
Place the jar inside the box, then place the loaded box on the plate30.0%70.0%
Overall56.7%80.0%

The largest improvement occurs on the composite task, from 30% to 70%. This task captures an important intermediate-state dependency: after the jar is inserted, some grasp regions of the box become obstructed. ProactiveVLA switches to an accessible right-rim grasp and completes the transfer, whereas HarnessVLA remains unsuccessful.

ProactiveVLA actively proposes and grasps various objects (such as cups) and combines different tasks, thereby mastering a broader range of grasping techniques and contact points. In contrast, Harness VLA repeatedly practices the same task, resulting in a more limited scope of experience.

Thanks for your feedback again, we have carefully revised our paper to include them.

AuthorsOpenReview ↗

2. Evaluations for different planners and controllers (Weakness 1)

For different planner, we additionally evaluated Harness VLA and ProactiveVLA using deepseek-v4.1-flash through Claude Code agent harness, with the high reasoning configuration during both adaptation and evaluation. Evaluated on tasks 0–9 × seeds 1–10 on LIBERO-Pro Goal-S:

MethodSuccess rate
Harness VLA57%
ProactiveVLA67%

For different controller, we use RLDX-1 as the controller to conduct a RoboCasa365 Composite Seen evaluation with 50 rollouts per task. (In the evaluation of Libero-Pro, RLinf-released pi05_libero130_fullshot SFT checkpoint is used as controller for both Harness VLA and ProactiveVLA.) Evaluated on 16 tasks, 50 seeds per task, 800 rollouts per method:

MethodSuccess rate
Harness VLA32.90%
ProactiveVLA40.90%

We have carefully revised our paper to include them.

AuthorsOpenReview ↗

3. Evaluation with more rollouts for RoboCasa365 (Weakness 2)

Thank you for highlighting the importance of evaluation precision. We conducted an expanded RoboCasa365 Composite Seen evaluation with 800 rollouts for Harness VLA and ProactiveVLA (50 rollouts per task, therefore more comparable to other baselines), using RLDX-1 as the controller. On expanded evaluations, ProactiveVLA outperforms existing baseline methods:

MethodSuccess rate95% CI
Harness VLA32.9%29.7–36.2%
ProactiveVLA40.9%37.5–44.3%

Thanks for your feedback again, we have carefully revised our paper to include them.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Revisions

We thank all reviewers for their constructive feedback. During the discussion period, we conducted additional experiments and revised the manuscript to address four main areas of concern.

1. Real-robot validation and alternative agent configurations (Sections 4.1–4.2; Appendices B.1 and B.4).

To address limited validation beyond the original simulation setup (BKXP W1), we added an ARX dual-arm comparison on two atomic tasks and their composition, with 10 randomized trials per task. ProactiveVLA achieves 80.0% overall success versus 56.7% for HarnessVLA, with the largest improvement on the composite task (70.0% versus 30.0%). We also evaluated both methods using DeepSeek-v4.1-flash through Claude Code, obtaining 67.0% versus 57.0% on Goal-S.

2. Larger-scale evaluation and repeated memory acquisition (Table 2; Appendices A, B.2, and D).

For evaluation precision and dependence on a particular exploration run (BKXP W2&Q1; 9S6s W2; G6am W2.2–2.3&Q2), we expanded RoboCasa365 evaluation from 5 to 50 rollouts per task, totaling 800 per method, and evaluated both methods with the RLDX-1 controller. ProactiveVLA achieves 40.9% versus 32.9% for HarnessVLA. We additionally report our standalone RLDX-1 evaluation. Appendix A reports 95% confidence intervals for the 800-rollout comparison, and Appendix D reports the per-task results. Across three separate memory-acquisition runs on Goal-T, ProactiveVLA consistently outperforms HarnessVLA, averaging 86.7% versus 80.7%.

3. Contribution of exploration and accumulated experience (Section 5; Appendix B.3).

To separate experience acquisition from memory organization (9S6s W2&Q2; G6am W2.4&Q3), we applied the same memory-processing pipeline to HarnessVLA’s task-refinement records. Goal-T success increases from 81.0% to 83.0%, compared with 88.0% for ProactiveVLA, supporting the value of acquiring additional interaction evidence. We also clarified that both methods share the same evaluation-time execution instructions, without an additional task-decomposition or tool-selection mechanism in ProactiveVLA. We further evaluated Harness VLA with an explicitly fine-grained execution prompt (G6am W2.4&Q3), where fine-grained prompting alone does not recover the gain. This is reported in our response 6 to Reviewer G6am and will be added in the final version.

To examine accumulated experience (G6am W2.5&Q4), we pooled each method’s memories across all 80 LIBERO-Pro tasks. ProactiveVLA retains an advantage on Goal-T (90.0% versus 87.0%). In our response on novelty (G6am W3), we further clarified the central contribution: under a limited interaction budget, structured exploration beyond a completed task can provide more useful preparation than continued task refinement.

4. Presentation, experimental settings, and failure analysis (Sections 3–5; Figure 4; Appendices A, B.1, and C).

We clarified adaptation and evaluation budgets, the transition to exploration, initial-task failure handling, and the separation of adaptation evidence from evaluation (BKXP Q1–Q2; 9S6s W1&Q1; G6am W1.1&Q1). We specified auxiliary model configurations and added per-rollout model requests, robot actions, token usage, and elapsed time for both methods (Appendix B.1). We also clarified notation and revised Figure 4 to connect the recovery behavior with the retrieved memory and supporting exploration evidence (G6am W1.2–1.4).

We distinguished locally evaluated results from published baselines and explained benchmark-specific checkpoint availability (9S6s Q2; G6am W2.1). Finally, we added an L10-T failure analysis concerning placement constraints that prevent subsequent drawer closure, and tempered the interpretation of the smaller component-ablation gains (9S6s W2).

Given the 10-page main-text limit, the main text reports average success rates for the RoboCasa365 evaluation and the component ablations. The per-task RoboCasa365 results are in Appendix D.

We sincerely thank the Area Chair and the reviewers for their time and valuable feedback, which has helped us strengthen the manuscript. Please let us know whether these revisions address your concerns. We welcome further questions and discussion.

Best regards,

Authors of Submission 24364

AuthorsOpenReview ↗

Summary of the contribution

As the discussion period closes, we would like to summarize what we believe is the central message of this paper.

The paper asks how a robot should spend a limited amount of interaction in a new environment. The common answer is to keep practicing the task it was given. We show that proactively exploring the environment prepares the agent better for later tasks than continued practice of the same task. On LIBERO-Pro Goal-T, success rises from 75% to 88%; on a real ARX dual-arm robot, the composite task rises from 30% to 70%. When only one VLA call is allowed at evaluation, ProactiveVLA completes 48% of instances versus 19% for the task-refinement baseline.

The broader point is that useful adaptation comes from actively acquiring knowledge about the environment. During the discussion we added experiments showing that this benefit cannot be reproduced by better memory organization or by more detailed prompting alone, and we believe the question and the evidence are of interest to the community.

We thank the reviewers and the Area Chair for their time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.