acceptodds
Under review as a conference paper at ICLR 2027

CARE-VLA: Context-Aware Retrieval of Execution Experience for Agentic VLA Planning

Abstract

Agentic vision–language–action (VLA) systems use external planners and execution memories to support frozen control policies. However, memories organized around complete tasks or episodes do not directly align with local planning decisions, and irrelevant or redundant experience can distract the planner. We propose CARE-VLA, a training-free framework for context-aware retrieval of step-level execution experience. CARE-VLA constructs a read-only multimodal experience bank from reference trajectories, linking each primitive execution to its preceding context, outcome, and before–after observations. At each decision boundary, it filters records by execution-context compatibility, fuses task-and-context and visual rankings, and selects a small set of relevant records with diverse next primitives and distinct source episodes. The selected experience forms a Memory Card that supplements the planner's original context without updating planner or VLA parameters. On LIBERO-Pro and RoboTwin C2R, CARE-VLA achieves success rates of 80.3% and 61.6%, respectively, exceeding the published Harness VLA results with the corresponding Codex planners by 8.2 and 3.6 percentage points. Compared with Harness VLA, CARE-VLA reduces the average number of primitive executions by 8.9% on episodes successfully completed by both methods. These results support step-level experience retrieval as a means of improving frozen agentic VLA planning.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Agentic Planning, Context-Aware Retrieval, Multimodal Memory, Training-Free Adaptation, Robotic Manipulation

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 4 reviews, above 68% of the others.
Reviewer_QW5iOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes CARE-VLA, a training-free framework that improves agentic VLA planning by retrieving step-level execution experience from a read-only multimodal bank. The key idea is to decompose reference trajectories into primitive-level transitions, each linking the pre-decision execution context, the executed primitive, its outcome, and before–after observations.

At each decision boundary, CARE-VLA performs context-compatibility filtering, dual-route ranking (task+context semantics via Qwen3-Embedding and visual similarity via SigLIP2), reciprocal rank fusion, and diversity-constrained selection to produce a compact "Memory Card" of at most three records. This card supplements the planner's existing context without updating any parameters.

Evaluated on LIBERO-Pro (80.3% vs. 72.1% for Harness VLA with Codex, +8.1 pp), RoboTwin C2R (61.6% vs. 58.0%, +3.6 pp), and three real-robot pick-and-place tasks (85.0% vs. 73.3%, +11.7 pp), CARE-VLA demonstrates consistent improvements while also reducing average primitive executions by 8.9% on shared successful episodes.

Critical strengths
  1. The identification of decision-level misalignment between memory granularity and planning granularity (§1, p.1–2) is a genuine and underappreciated problem in agentic VLA systems. The step-level decomposition (§3.2) directly addresses this by aligning memory records with primitive decision boundaries, enabling local evidence retrieval across tasks and episodes. The training-free design, which preserves the planner and VLA parameters while only augmenting the planner's context, is both principled and practically valuable. This is consequential because it demonstrates that meaningful improvements can be achieved through better experience organization alone, without the cost and risk of retraining.

  2. The paper demonstrates improvements on two simulation benchmarks (LIBERO-Pro: +8.1 pp; RoboTwin C2R: +3.6 pp) and real-robot tasks (+11.7 pp), all using the same frozen planner and VLA (Tables 1–4). The ablation studies, particularly the diversity ablation (Table 3: 84.3% → 72.5% without primitive diversity) and the experience-source study (Table 3: 79.8% → 84.3% from 1-suite to 8-suite), provide meaningful evidence for the individual design choices. The efficiency result (8.9% fewer primitive executions, Table 6) further supports that the gains come from better decision-making rather than simply more computation.

Critical weaknesses

W1. Uncontrolled experience-budget asymmetry. CARE-VLA builds its bank from 80 reference episodes (1,093 records; Table 5), whereas Harness VLA's task-specific memory uses one trace and one summary per task. No baseline gives Harness VLA an equivalent amount of structured experience (e.g., more traces per task or richer summaries). Without this control, the observed gains cannot be cleanly attributed to the step-level retrieval mechanism rather than to having substantially more experience available. The paper acknowledges this asymmetry only in the real-robot section (§4.5) but does not address it in the simulation experiments.

W2. Absence of statistical significance reporting. No confidence intervals, standard deviations, or significance tests are reported for any experiment. On RoboTwin C2R (250 episodes), the 3.6 pp improvement over Harness VLA corresponds to roughly 9 episodes, which may fall within random variation. On LIBERO-Pro (800 episodes), the 8.1 pp gap is larger but still unverified. Without at least a paired bootstrap or McNemar test on the shared task–seed pairs, the reliability of these improvements—especially on RoboTwin—remains uncertain.

W3. Execution-context filtering is never independently ablated. The paper motivates context-compatibility filtering as essential for distinguishing scenes that look similar but differ in execution state (§3.3). However, no ablation removes only the filtering step while keeping ranking, fusion, and diversity selection unchanged. The existing ablations (text-only in Table 1; w/o diversity in Table 3) change different components. Since filtering is presented as a core design principle, its isolated contribution should be directly demonstrated.

W4. Numerous hyperparameters without sensitivity analysis. The retrieval pipeline introduces at least ten tunable constants: (βctx,βhist,βfail)=(0.50,0.30,0.20)(\beta_{\text{ctx}},\beta_{\text{hist}},\beta_{\text{fail}})=(0.50,0.30,0.20), λ=0.15\lambda=0.15, N=12N=12, c=10c=10, α=0.5\alpha=0.5, τabs=0.55\tau_{\text{abs}}=0.55, Δ=0.15\Delta=0.15, Kmax⁡=3K_{\max}=3, and history weights (0.6,0.3,0.1)(0.6,0.3,0.1). The paper does not explain how these were selected, whether they were tuned on a held-out set or on the evaluation tasks, or how sensitive results are to their values. Even a brief sweep of Kmax⁡∈{1,2,3,5}K_{\max}\in\{1,2,3,5\} and τabs∈{0.45,0.55,0.65}\tau_{\text{abs}}\in\{0.45,0.55,0.65\} would substantially strengthen confidence.

W5. Limited real-robot validation scope. The real-robot evaluation covers only 3 pick-and-place tasks with 20 trials each (60 trials total; Table 4). All tasks share the same primitive structure (approach → grasp → transport → place), leaving untested whether the retrieval mechanism helps with more diverse manipulation (e.g., assembly, tool use, or contact-rich tasks). No per-trial outcomes, failure-mode breakdown, or cross-day robustness are reported. While the 11.7 pp gain over Harness VLA is encouraging, the small and homogeneous task set limits the strength of the real-world claim.

Questions for authors
  1. Can you provide a baseline where Harness VLA is given additional reference experience (e.g., multiple task-specific traces per task instead of one, or more detailed summaries) matched to the information content of CARE-VLA's 80-episode bank? Alternatively, can you run CARE-VLA with a reduced bank (e.g., 10 reference episodes, ~130 records) to assess how much of the improvement persists? This would help isolate whether the gain comes from the retrieval mechanism or from having more experience.

  2. Could you provide confidence intervals or results of statistical significance tests (e.g., paired bootstrap, McNemar's test) for the main results on LIBERO-Pro and RoboTwin C2R? In particular, is the 3.6 pp improvement on RoboTwin C2R (approximately 9 episodes out of 250) statistically significant?

  3. Could you report an ablation that removes only the execution-context filtering step (allowing all records to enter ranking, without requiring previous primitive/outcome/failure type/gripper state to match), while keeping all other components unchanged? This would directly test the paper's central claim that context compatibility is essential.

  4. What is the success rate of the 80 reference episodes used to construct the LIBERO-Pro experience bank? Are all reference episodes successful, or do they include failure–recovery sequences? How does the quality of reference episodes affect retrieval quality?

  5. Could you provide a brief analysis of failure cases where CARE-VLA underperforms or where retrieved evidence is misleading? For instance, are there cases where context-compatible records suggest an inappropriate next primitive due to subtle scene differences not captured by the visual encoder?

  6. For RoboTwin C2R, the bank uses "one non-evaluation initialization seed per task across all 50 tasks" (§4.1). Could you clarify the total number of step-level records in this bank and whether the cross-task retrieval (across all 50 tasks) is a significant factor in the observed improvement?

7 comments
AuthorsOpenReview ↗

Experience Controls, Statistical Reliability, Retrieval Analysis, and Real-Robot Evaluation

Q1 and W1: Experience-budget fairness and the contribution of retrieval

Q1 — Reviewer’s question: Could Harness VLA receive additional reference experience, such as multiple traces per task or richer summaries, matched to the information content of CARE-VLA’s 80-episode bank? Alternatively, could CARE-VLA be evaluated with a reduced bank of approximately 10 reference episodes and 130 records to distinguish retrieval gains from the benefits of more experience?

W1 — Reviewer’s concern: CARE-VLA uses 80 reference episodes and 1,093 records, whereas Harness VLA uses one task-specific trace and summary per task. Without a baseline given an equivalent amount of structured experience, the gains cannot be cleanly attributed to step-level retrieval. This asymmetry is acknowledged for the real robot but not addressed in simulation.

Our response to Q1 and W1:

Thank you for your attention to experience budgets and performance attribution. We address Q1 and W1 through matched-experience controls, small-bank experiments, and experience-expansion experiments to distinguish the effects of experience volume from those of retrieval.

1. Providing Harness VLA with additional experience from the same sources

The original bank contains one reference trajectory from each of 80 tasks. CARE-VLA segments these trajectories into step-level records that can be retrieved across tasks. As GPT-5.5 was no longer available during the rebuttal period, the supplementary experiments use GPT-6 Astra with low reasoning on Goal-T/S and Long-T/S and are reported separately from the original GPT-5.5 results.

All groups use matched planners, VLA backends, software, task–seed pairs, execution budgets, and evaluation protocols. The added experience is collected by Astra, with one non-evaluation reference episode for each of 40 tasks, and is provided to the planner in different forms.

Method / settingAccessible reference experienceAverage success rate (%)
Harness VLA (internally reproduced)Task-specific memory for the current task and the original global memory83.5
Harness VLA + expanded task experienceHarness-format experience from all 40 tasks76.8
Harness VLA + full step-level experience bankOriginal memories + the entire step-level experience bank, with retrieval disabled85.8
CARE-VLAThe same original memories + the same step-level experience bank, with retrieval and filtering of relevant records89.3

Expanding access to Harness-format experience decreases the success rate from 83.5% to 76.8%. This setting covers the same reference tasks as CARE-VLA but organizes and uses the experience differently, showing that more experience does not automatically yield gains. Irrelevant experience and the additional selection burden may affect current decisions.

Providing the full step-level experience bank increases the success rate to 85.8%, supporting the value of step-level organization for using added experience. With exactly the same original memories and step-level experience bank, CARE-VLA further achieves 89.3%, outperforming the retrieval-disabled control by 3.5 percentage points, directly supporting the role of retrieval (Appendix E.2, Table 9).

2. Substantial gains remain with only 10 reference episodes

Table 3 of the original paper already includes a small-bank setting that uses only 10 reference episodes for each target suite. The record counts for Spat-T, Spat-S, Goal-T, and Goal-S are 152, 137, 108, and 135, respectively, averaging 133 records, consistent with the scale suggested by the reviewer.

The following results are averaged over these four suites using GPT-5.5 with medium reasoning. All CARE-VLA settings use the same retrieval parameters and a Memory Card budget of at most three records.

Method / settingSize of the added step-level experience bankAverage success rate (%)
Harness VLA (published baseline)No added step-level experience bank72.8
CARE-VLA: 1-suite bank10 reference episodes from the current target suite79.8
CARE-VLA: 4-suite bank40 reference episodes from Spat-T/S and Goal-T/S81.8
CARE-VLA: 8-suite bank80 reference episodes from all eight suites84.3

With only 10 reference episodes, CARE-VLA still improves by 7.0 percentage points, retaining approximately 61% of the full-bank gain. Thus, the complete 80-episode bank is not necessary to obtain improvements.

3. Expanding experience sources can further improve performance

We progressively expand the experience sources from Long-T/S to Goal-T/S and Long-T/S, then to all eight LIBERO-Pro T/S suites, and finally add reference experience from the four task suites under Object (Obj) perturbations. These configurations contain 20, 40, 80, and 120 reference episodes, respectively.

The retrieval parameters and card budget remain unchanged. Evaluation is fixed to the same 200 task–seed pairs on Long-T/S, consisting of 2 suites × 10 tasks × 10 seeds.

Total reference episodes204080120
CARE-VLA success rate82.0%83.5%84.5%85.0%

The success rate increases from 82.0% to 85.0%, showing the value of expanding experience sources for further performance improvements. In future work, we will investigate the relationships among experience scale, diversity, and retrieval effectiveness on larger benchmarks.

In summary, directly expanding access to Harness-format experience does not yield gains, whereas CARE-VLA outperforms the retrieval-disabled control by 3.5 percentage points with the same step-level bank. With only 10 reference episodes per suite, it still improves by 7.0 percentage points and retains approximately 61% of the full-bank gain. Expanding experience sources further improves performance. These results indicate that the gains arise from both experience accumulation and effective experience organization and retrieval.

Q2 and W2: Statistical significance and uncertainty of the main results

Q2 — Reviewer’s question: Could confidence intervals or statistical significance tests, such as paired bootstrap or McNemar’s test, be provided for LIBERO-Pro and RoboTwin C2R? In particular, is the 3.6-percentage-point improvement on RoboTwin C2R, corresponding to approximately nine out of 250 episodes, statistically significant?

W2 — Reviewer’s concern: The paper does not report confidence intervals, standard deviations, or significance tests. The RoboTwin C2R gain may reflect random variation, and the larger LIBERO-Pro gap is also unverified. Paired analysis on shared task–seed pairs is needed to assess the reliability of the improvements.

Our response to Q2 and W2:

Thank you for your suggestion. Using the historical execution logs, we conducted episode-level paired bootstrap analysis and two-sided exact McNemar tests on success/failure outcomes for the same task–seed pairs. The supplementary results are reported in Appendix L, Table 17.

1. Statistical tests use the internally reproduced baseline

The published Harness VLA results provide only aggregate success rates, which cannot be used for paired tests. We therefore use the internally reproduced baseline with matched configurations and per-episode execution records. Its success rates on LIBERO-Pro and RoboTwin C2R are 71.9% and 57.2%, respectively, close to the published 72.1% and 58.0%.

2. The paired improvements on both benchmarks are statistically significant

QuantityLIBERO-ProRoboTwin C2R
Matched evaluation episodes800250
Harness VLA successes575143
CARE-VLA successes642154
Both methods succeed548142
Only CARE-VLA succeeds9412
Only Harness VLA succeeds271
Success-rate gain (percentage points)8.44.4
Paired bootstrap 95% CI (percentage points)[5.75, 11.00][1.60, 7.20]
Two-sided exact McNemar testp=7.14×10−10p=7.14\times10^{-10}p=0.00342p=0.00342

For RoboTwin C2R, which is of particular concern to the reviewer, statistical significance depends not only on the net increase in successful episodes but also on the differences between the methods on the same episodes: 12 episodes succeed only with CARE-VLA, while just one succeeds only with Harness VLA. This paired difference is statistically significant (p<0.01p<0.01), and the 95% confidence interval for the gain excludes zero. LIBERO-Pro similarly shows a paired advantage of 94 versus 27 episodes, with the statistical test also supporting the reliability of the improvement.

In summary, compared with the internally reproduced Harness VLA baseline under matched configurations, CARE-VLA improves by 8.4 and 4.4 percentage points on LIBERO-Pro and RoboTwin C2R, respectively. Both improvements are statistically significant (p<0.01p<0.01), and both paired-bootstrap 95% confidence intervals exclude zero, further supporting the reliability of the gains.

Q3: Ablation of execution-context compatibility filtering

Reviewer’s question: Could an ablation remove only execution-context filtering, allowing all records to enter ranking without matching the previous primitive, outcome, failure type, or gripper state, while leaving all other components unchanged? This would directly test the central claim about context compatibility.

Our response:

Thank you for your suggestion. We have added an ablation that removes only execution-context compatibility filtering while keeping all other components unchanged, directly testing its contribution (Appendix E.2, Table 9).

The experiment uses GPT-6 Astra with low reasoning on Goal-T/S and Long-T/S in LIBERO-Pro. Both settings share the same original memories, step-level experience bank constructed from 40 reference episodes, retrieval parameters, Memory Card budget, and evaluation settings. The ablation no longer requires historical records to match the current previous primitive, outcome, failure type, and gripper state, allowing all records to enter subsequent ranking. Multimodal ranking and diversity-based selection remain unchanged.

MethodGoal-TGoal-SLong-TLong-SOverall
CARE-VLA without context compatibility filtering8897867486.3
CARE-VLA9298897889.3

Removing filtering reduces the average success rate from 89.3% to 86.3%. This shows that context compatibility filtering provides additional gains even when multimodal ranking and diversity-based selection are retained. Historical records with similar tasks or visual observations may differ in action outcomes, failure types, and gripper states. Filtering out incompatible records before ranking helps provide experience better suited to the current execution state.

In summary, with the other components and experience budget unchanged, context compatibility filtering provides an average gain of 3.0 percentage points, supporting its importance for retrieving decision-relevant experience.

Q4: Reference-episode success, failure–recovery sequences, and experience quality

Reviewer’s question: What is the success rate of the 80 reference episodes used to construct the LIBERO-Pro bank? Are they all successful, or do they include failure–recovery sequences? How does reference-episode quality affect retrieval quality?

Our response:

Thank you for your attention to reference-experience quality. All 80 stored reference episodes ultimately complete their tasks, giving a task success rate of 100%, and are segmented into 1,093 step-level records. We follow Harness VLA’s reference initialization procedure to collect one ultimately successful trajectory per task. This 100% refers to the selected trajectories stored in the bank, not to every attempt during collection.

An ultimately successful episode can still contain local failures and recovery, such as adjusting the gripper after an empty grasp or changing the action after stalled movement. We retain these steps, their execution feedback, subsequent actions, and before–after observations. Thus, the bank contains failure–recovery sequences, rather than only successful actions.

Reference-experience quality depends not only on final success but also on record accuracy, state diversity, and applicability to the current decision. Accurate feedback and before–after observations help distinguish effective actions from failed attempts, while local failure and recovery provide experience absent from smoothly executed trajectories. Conversely, even when the source episode ultimately succeeds, highly repetitive records or records mismatched to the current state may occupy limited retrieval slots and reduce the value of the evidence. CARE-VLA therefore uses context compatibility filtering, multimodal ranking, and diversity-based selection to provide complementary experience applicable to the current state.

In future work, the bank could also include ultimately unsuccessful reference episodes, using explicit failure sequences as negative examples to help the planner identify actions to avoid. For example, under similar contact-induced stalling, a historical record showing that repeatedly issuing the same movement command produces no progress could prompt the planner to avoid repeating ineffective actions and instead observe again or adjust its strategy. We will study combinations of successful experience, recovery experience, and negative failure examples, and their effects, on larger benchmarks.

In summary, the current 80 reference trajectories all ultimately succeed but include local failures and recovery. The value of experience should be assessed by the execution evidence it provides. Including negative failure examples in the future could allow the bank to support decisions about both how to act and which actions to avoid.

Reviewer_64iBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper uses a few relevant examples to determine what step to take next based upon the use of pre-train VLA. This is done in place of the complete understanding of the entire experiment; however, it will use a short example of the exact scenario to provide accuracy. It achieves this through a process of breaking down the demonstration into each decision that is made, identifying those examples which are similar to the present condition and then developing them into index cards (i.e. presenting the planner with up to three examples) and finally allowing the current planner to decide on its own. Their findings indicated that the simulation success rate increased from 72.1% to 80.3% and resulted in eight more successful task completions per 100 attempts, while their small real robot experiment increased the success rate from 73.3% to 85%. Although the underlying AI remains unchanged, the robot can potentially make better decisions by recalling the appropriate examples for a given situation.

Critical strengths

The paper addresses a core planning problem. Ablation study and empirical gains are meaningful. It is consequential to me whether the method solves an important problem and the evidence supports it.

Critical weaknesses

Baseline comparison are not fully controlled and the core experiments do not isolate core contribution significantly. It is consequential to my overall assessment because the evidence supports a narrower claim than the paper suggests.

Questions for authors

The results of the comparison appear to be very encouraging; however, I am concerned about how similar the comparison groups were. For example, did Harness VLA and CARE-VLA have exactly the same planner versions, software, evaluation seed, execution budget, and reference experience? In addition, since the baseline performance data for the two systems was collected in previous studies, why do we think that the performance differences can be attributed solely to CARE-VLA?

2 comments
AuthorsOpenReview ↗

Controlled Baselines and Attribution of Retrieval Gains

W1 and Q1: Controlled comparisons and attribution of performance gains

W1 — Reviewer’s concern: The baseline comparisons are not fully controlled, and the core experiments do not sufficiently isolate the main contribution. This affects the overall assessment because the evidence supports a narrower claim than the paper suggests.

Q1 — Reviewer’s question: The results are encouraging, but did Harness VLA and CARE-VLA use exactly the same planner versions, software, evaluation seeds, execution budgets, and reference experience? Given the reliance on baseline results from previous studies, how can the performance differences be attributed solely to CARE-VLA?

Our response to W1 and Q1:

Thank you for your attention to comparison fairness and performance attribution. We agree that comparing success rates reported in different studies alone does not sufficiently isolate the contribution of the method itself. We therefore added internally matched baselines and retrieval ablations using the same reference experience to address two questions: whether the overall gains hold and whether they arise solely from access to more experience.

1. Which experimental conditions are matched?

Our implementation is based on the official Harness VLA repository. CARE-VLA and the internally reproduced Harness VLA baseline use the same Codex CLI version, GPT-5.5 model identifier, reasoning settings, base prompts, original task-specific and global memories, VLA backend, primitive implementations, feedback rules, evaluation task–seed pairs, execution budgets, and evaluation protocol. Specifically, LIBERO-Pro uses πRLinf\pi_{\mathrm{RLinf}} with medium reasoning, while RoboTwin C2R uses LingBot-VLA with xhigh reasoning. CARE-VLA adds the step-level experience bank, retrieval, and evidence injection through a Memory Card to this setup.

The exact CLI version and model snapshot used in the official historical runs were not disclosed, so we cannot confirm that these conditions exactly match those of the published results. We report the published and internally matched results separately and assess the performance gains against the internally reproduced baseline (Appendix E.1, Table 8):

MethodLIBERO-ProRoboTwin C2R
Harness VLA (published)72.1%58.0%
Harness VLA (internally reproduced)71.9% (575/800)57.2% (143/250)
CARE-VLA80.3% (642/800)61.6% (154/250)
Gain over the internally reproduced baseline8.4 percentage points4.4 percentage points

The gains are computed from success counts and rounded to one decimal place. Thus, CARE-VLA’s improvements are supported not only by comparisons with published results but also by internal runs under matched configurations. The primitive usage statistics in Table 6 likewise come from the execution logs of these internal runs.

2. Is the reference experience the same? How do we distinguish the effects of more experience from those of retrieval?

CARE-VLA retains Harness VLA’s original memories and segments reference execution trajectories into step-level records that can be retrieved across tasks. Both methods follow the same reference initialization and reference–evaluation separation protocol. However, the original baseline and CARE-VLA differ in how experience is organized and in the scope of accessible experience, so the above comparisons alone cannot quantify the contribution of retrieval in isolation. For example, LIBERO-Pro uses seed 0 to collect reference experience and seeds 1–10 for evaluation. The bank is frozen before evaluation, and evaluation trajectories are not fed back into it.

To directly control experience availability, we added the following comparisons. As GPT-5.5 was no longer available during the rebuttal period, the additional experiments use GPT-6 Astra with low reasoning. Reference experience was also collected by rerunning non-evaluation reference initializations with Astra. The experiments cover Goal-T, Goal-S, Long-T, and Long-S. All internal groups use matched planners, software, VLA backends, task–seed pairs, execution budgets, and evaluation protocols.

SettingUse of original memories and added experienceSuccess rate
Harness VLA (internally reproduced)Uses the original task-specific and global memories83.5%
Full-bank exposure without retrievalRetains the original memories and makes the entire step-level experience bank accessible to the planner85.8%
CARE-VLA without context compatibility filteringRetains the original memories, experience bank, and remaining retrieval components86.3%
CARE-VLARetains the original memories and provides relevant evidence through the complete retrieval pipeline89.3%

The last three settings use exactly the same added experience bank, constructed by segmenting one non-evaluation reference episode from each of 40 tasks. They share the same set of reference executions, rather than using separately collected experience (Appendix E.2, Table 9).

These controls further distinguish the contributions of the individual components:

  • Increased experience availability: Full-bank exposure raises the success rate from 83.5% to 85.8%, showing that the added step-level experience itself is useful.
  • Retrieval and evidence selection: With both the original memories and the added experience bank held fixed, CARE-VLA improves from 85.8% to 89.3%, a gain of 3.5 percentage points, directly supporting the role of selecting experience for the current decision.
  • Context compatibility filtering: Removing only this component reduces the success rate from 89.3% to 86.3%, showing that it contributes even when multimodal ranking and diversity-based selection are retained.

In summary, our performance attribution is based on internally matched comparisons rather than solely on published baselines. With the planner, software, and evaluation settings fixed, CARE-VLA improves by 8.4 and 4.4 percentage points on the two GPT-5.5 benchmarks, respectively. Further fixing the original memories and added experience bank, retrieval and evidence selection still provide an additional gain of 3.5 percentage points with Astra. These results support the paper’s core conclusion: organizing execution experience into step-level evidence relevant to the current state can improve planning without updating the planner or VLA weights, and the gains cannot be explained solely by access to more experience.

Reviewer_kXQ2OpenReview ↗
Rating
3
Confidence
4
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The work proposed CARE-VLA, a variant of VLA model for robotic manipulation with an enhanced version of harness. The work is built upon a previous work called harness VLA. The key idea is to build a memory offline that stores previous episodes and tasks with great details. In reasoning time, the model first retrieves related records from the above-mentioned memory bank and keepts top-k most informative records. Afterwards, the model is executed conditioned on both the retrieved information and other context. Experiments are conducted on several simulation benchmarks and real-robot deployment.

Critical strengths

Below I would summarize the strong points of this work.

  • The paper is well motivated and with good writing. Most key designs of the proposed model are presented with sufficient details or illustrations. The field of embodied robotics is witnessing a blossom of agent-VLA/WAM hybrid methods. It well addresses some important research problems in such a context.

  • The experimental results are positive. On most benchmarks, it surpassed the key baseline, namely harness VLA, with non-trivial margins. Experiments are conducted on both simulation tasks and real-world challenges on robots.

Critical weaknesses

I find this work to be interesting. Nevertheless, its technical part is somewhat incremental given the high similarity to the existing work harness VLA. Below I would summarize my concerns.

My major concern is regarding the construction of the memory. Does it also contain the past episodes for the same task or even the same configuration? In that case, one may wonder whether the comparison is fair enough, since the proposed model essentially executes multiple times and learn from past success and failures. More key details are desired about the source of memory. It would be also useful to illustrate it intuitively by presenting some retrieved memory cards (including failure mode).

It is reported that the retrieval costs roughtly 0.1s per query, which is non-trivial given the essence of rea-time requriement for VLA. The configuration for estimating this time complexity is specified in the main text. The scheme that a real robot works with such a large latency is also not presented. The experiments validate with roughly 10K decisions. However, the scale of the memory is not revealed nor discussed. Will it be updated by including more records brought by previous decisions? What about its time complexity and possible solution with million-scale records?

The experiments are somewhat inconsistency. The conclusions with Harness VLA (CC) are diffierent on two benchmarks.

Questions for authors

See my detailed comments in the "critical weaknesses" section.

3 comments
AuthorsOpenReview ↗

Memory Construction, Runtime, Scalability, and Cross-Planner Comparisons

W1: Memory construction, evaluation fairness, and retrieved Memory Card examples

Reviewer’s concern: Does the memory contain past episodes from the same task or even the same configuration? If so, is the comparison fair, given that the proposed method may effectively execute tasks multiple times and learn from past successes and failures? Please clarify the sources of memory and provide examples of retrieved Memory Cards, including failure modes.

Our response:

Thank you for your attention to memory sources and comparison fairness. We clarify three aspects: the sources of reference experience and their separation from evaluation, local failures within successful trajectories, and actual Memory Card examples.

1. Sources of reference experience and comparison fairness

Yes, the experience bank contains historical reference executions of the same tasks, following the practice of Harness VLA. CARE-VLA retains the original task-specific and global memories and segments the original reference trajectories into step-level records. The 80 successful reference episodes in LIBERO-Pro correspond to 80 tasks and 1,093 records (Appendix G, Table 12).

Both methods collect reference experience using seed 0 and evaluate using seeds 1–10. Reference and evaluation episodes may share task definitions, object categories, and scene templates, but their initialization seeds do not overlap. The experience bank is frozen before evaluation; evaluation trajectories are not added, and model parameters are not updated.

Therefore, the comparison is fair with respect to the separation between reference experience and evaluation that you are concerned about. Same-task reference experience and the seed split are shared settings for both methods. CARE-VLA reuses existing reference trajectories; it does not update experience using the success or failure outcomes of evaluation episodes, nor does it obtain additional information by repeatedly executing evaluation episodes.

2. Local failures within successful reference trajectories

An ultimately successful reference episode does not imply that every primitive execution succeeds. It may still contain empty grasps, stalled movements, and subsequent recovery, consistent with the reference-trajectory setting in Harness VLA.

The Memory Card retains the local execution context, historical actions and outcomes, and before–after observations for the planner to assess applicability. Thus, failure or no_progress in a card denotes a local failure, rather than final failure of the source episode.

3. Memory Card examples

Same-task recovery (Appendix H.1, Figure 4). In the task of placing a wine bottle on top of a drawer, the current evaluation is Goal-S task 2, seed 4. After VLA_ACT returns an empty grasp with the gripper open, the system retrieves a record from reference seed 0 of the same task, in which recovery is achieved through SET_GRIPPER. Based on the current image, the planner determines that the gripper fingers are positioned on either side of the bottle neck, then closes the gripper, lifts the bottle, and completes the placement. This example illustrates recovery under different initializations; the original task-specific memory also contains similar guidance.

Cross-task recovery (Appendix H.2, Figure 5). In the task of placing a salad dressing bottle in a basket (Obj-T task 8, seed 6), a downward MOVE_TO returns failure/no_progress, although the bottle has already entered the basket opening. The system retrieves a record from placing a book in a caddy (Long-S task 5, reference seed 0), which has the same local feedback followed by a successful RELEASE. The current planner explicitly combines this experience with the current image, releases the bottle, and retracts, ultimately completing the task. The reused information is local handling experience, rather than a complete trajectory or historical spatial arguments.

In summary, we follow Harness VLA’s reference-trajectory setup and seed split, maintaining fairness in the separation between reference experience and evaluation. CARE-VLA segments existing trajectories into step-level records, and evaluation outcomes are not fed back into the bank. The two actual card examples illustrate same-task recovery and cross-task reuse after local failures, showing how this experience informs current decisions.

W2: Retrieval latency, real-robot execution, memory size, and scalability

Reviewer’s concern: Retrieval takes roughly 0.1 seconds per query, which may be significant for real-time VLA execution. Although the computational configuration is specified, how does a real robot operate with this latency? The experiments involve roughly 10,000 decisions, but the memory size is unclear. Is the memory updated with records from previous decisions? What is the retrieval time complexity, and how could the method scale to millions of records?

Our response:

Thank you for your attention to retrieval latency, robot execution, and experience-bank scalability. We address these issues from the following four perspectives.

1. Impact of approximately 0.1 seconds of retrieval latency on the actual decision process

CARE-VLA performs retrieval at high-level primitive decision boundaries. According to the logs, each agent inference call takes approximately 10–40 seconds. This duration includes network transfer and server-side waiting and is affected by network fluctuations and service load. As a reference for the runtime scale, the official Harness VLA paper reports an episode-level mean planner duration of 529.23 seconds, or approximately 8.82 minutes, for Codex with high reasoning on LIBERO-Pro (Harness VLA, Table 7 and Appendix G).

At approximately 0.1 seconds, retrieval overhead corresponds to only about 0.25%–1% of a single agent call. Thus, the added retrieval latency is small relative to the decision timescale of the current agentic VLA system. Our work focuses on improving the frozen planner’s decisions through context-aware selection of step-level experience. Large-model interaction and robot execution efficiency are also directions for further optimization.

2. How does the real robot execute under this latency?

The real robot follows this closed-loop process:

Observation and execution feedback → Retrieval and Memory Card generation → Agent selects a primitive → Robot executes → New observation and feedback.

Retrieval is invoked at high-level decision boundaries. Actions within a primitive are executed by the existing VLA policy or control program; retrieval is not repeated before each low-level control command (Appendix I, Figure 6).

We completed three pick-and-place tasks involving an apple, a banana, and a milk container on a Franka Research 3. In the recorded full-execution videos, the individual completion times for these three tasks are 193.7 seconds, 157.2 seconds, and 176.6 seconds, respectively.

Thus, during real-robot execution, retrieval is invoked only before high-level action decisions and does not change the control process within a primitive. Relative to the 10–40 seconds required for each agent inference call, the approximately 0.1 seconds of additional retrieval latency is small.

3. How large is the experience bank, and is it continuously updated during execution?

Table 5 of the original paper already reports that the LIBERO-Pro experience bank contains 80 reference episodes and 1,093 step-level records. The 10,553 figure is the number of decision queries, representing repeated accesses to the bank rather than the number of stored records.

In the current experiments, the bank is frozen before evaluation, and newly generated evaluation trajectories are not added. This controls the sources and budget of experience, keeping the comparison fair and reproducible.

Continuous updating is a valuable research direction. In future work, we will explore adding verified new trajectories to the bank online after segmentation, filtering, and deduplication. CARE-VLA has already improved decisions through experience reuse. Online updating would allow the system to continuously incorporate execution experience from new tasks and scenes, providing richer and more up-to-date references for subsequent decisions.

4. What is the current time complexity, and how could it scale to millions of records?

Let MM denote the number of records in the experience bank, C≤MC\leq M the number of context-compatible records, and dd the total dimensionality of the text and visual features. With cached historical features and full sorting of candidates, the current retrieval complexity is:

O(M+Cd+Clog⁡C).O(M + Cd + C\log C).

The three terms correspond to context scanning, candidate similarity computation, and sorting, respectively. Query encoding and card generation are accounted for separately. In the worst case, C=MC=M, giving a complexity of O(Md+Mlog⁡M)O(Md+M\log M).

For experience banks containing millions of records, we can draw on RAG (Lewis et al., NeurIPS 2020) and established large-scale vector retrieval techniques, using the following pipeline:

Historical feature caching → Context-index filtering → Approximate nearest-neighbor retrieval → Multimodal reranking of a small candidate set → Memory Card generation.

Specific options include HNSW, Faiss inverted indexes and vector compression, and SSD-backed DiskANN (NeurIPS 2019). For example, on the billion-scale SIFT1B dataset, DiskANN achieves a mean query latency below 3 ms, more than 5,000 queries per second, and over 95% 1-recall@1 using 16 CPU cores, 64 GB of RAM, and an SSD. These are vector-index-stage results that provide a technical basis for scaling, rather than measured performance of CARE-VLA’s complete retrieval pipeline.

In our source-expansion experiment, increasing reference experience from 20 to 120 episodes raises the Long-T/S success rate from 82.0% to 85.0%, while still providing at most three records to the planner per decision. This supports expanding the scope of retrievable experience while keeping the evidence input compact for each decision. Efficient retrieval from million-record banks and online experience accumulation are valuable directions for further research.

In summary, retrieval takes approximately 0.1 seconds at high-level decision boundaries, adding about 0.25%–1% overhead relative to 10–40-second agent calls without changing the control process within a primitive. The three real-robot video examples demonstrate the system’s actual execution. The bank contains 1,093 records and remains frozen during the current evaluation; its retrieval complexity is O(M+Cd+Clog⁡C)O(M+Cd+C\log C). In future work, we will combine online experience accumulation with efficient indexing on larger-scale benchmarks to further study retrieval speed, experience reuse, and task performance with large experience banks.

W3: Different rankings against Harness VLA (CC) across the two benchmarks

Reviewer’s concern: The experimental results appear inconsistent because the comparisons with Harness VLA (CC) lead to different conclusions on the two benchmarks.

Our response:

Thank you for pointing this out. The different cross-planner rankings on the two benchmarks are related to the existing performance gaps between the planner configurations in Harness VLA. On RoboTwin C2R, Harness VLA (Codex, GPT-5.5) and Harness VLA (CC) achieve 58.0% and 58.4%, respectively, differing by only one successful episode out of 250 evaluations. Thus, CARE-VLA’s improvement over the Codex baseline allows it to surpass CC, reaching 61.6%. On LIBERO-Pro, the two Harness VLA configurations achieve 72.1% and 82.4%, respectively, a larger gap; CARE-VLA’s 80.3% remains below CC.

The main question examined in this paper is whether step-level experience retrieval improves performance with the planner held fixed. In the internally matched comparisons, CARE-VLA improves from 71.9% to 80.3% on LIBERO-Pro and from 57.2% to 61.6% on RoboTwin C2R, yielding gains of 8.4 and 4.4 percentage points, respectively. In addition, using GPT-6 Astra with low reasoning on the four suites Goal-T/S and Long-T/S, CARE-VLA increases the success rate from 83.5% to 89.3%, a gain of 5.8 percentage points (Appendix E, Tables 8–9).

In summary, the change in cross-planner rankings does not constitute a contradiction in the method’s effectiveness. Under matched planner configurations, CARE-VLA outperforms the corresponding baseline on both benchmarks and also achieves gains with Astra, further supporting its effectiveness with different underlying models.

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

CARE-VLA is a training-free framework for retrieving execution experience for agentic VLA systems. Built on Harness VLA, it adds a step-level multimodal experience bank containing pre-execution context, historical primitives and outcomes, and before–after observations. At each decision boundary, it filters records by execution-context compatibility, fuses task-and-context and visual rankings through reciprocal rank fusion, and selects up to three records with distinct next-primitive types and source episodes. These form a Memory Card supplementing the planner’s existing memories. CARE-VLA achieves 80.3% and 61.6% success on LIBERO-Pro and RoboTwin C2R, exceeding the published Harness VLA (Codex) results by 8.1 and 3.6 percentage points. The paper also studies visual information, experience sources, and primitive diversity through ablations and evaluates three real-robot pick-and-place tasks.

Critical strengths
  1. Clear motivation and practical design. Organizing execution experience into step-level records helps the planner retrieve evidence relevant to its current decision.
  2. The paper is clearly written and easy to follow, with an understandable workflow and implementation details.
Critical weaknesses
  1. Limited technical novelty. CARE-VLA is largely an incremental extension of Harness VLA, adding a step-level experience bank and retrieval, filtering, and evidence-selection mechanisms. The design is reasonable, but the technical advance appears limited.
  2. The main experimental evidence remains insufficient.
    1. On LIBERO-Pro, CARE-VLA (Codex) outperforms Harness VLA (Codex) but trails Harness VLA (CC) on most suites and overall. Although cross-planner comparisons cannot invalidate the method, its gains may depend on the coding agent or model. Could the authors compare both methods using Claude Code or Codex with another model?
    2. The main comparisons use published baseline results. How are experimental conditions matched, including Codex CLI versions, GPT-5.5 snapshots, prompts, and memories? Given the incremental design, disabling the added retrieval module in the same implementation would provide a direct controlled comparison. If such tests were conducted, or historical execution logs with identical configurations were reused, please clarify.
  3. End-to-end efficiency remains unclear. The paper reports retrieval-core latency and fewer primitive executions, but not Memory Card rendering costs or the additional planner inference time. Does the time saved through fewer planning steps and primitive executions outweigh these costs? Comparing end-to-end task duration with Harness VLA under matched conditions would clarify the practical efficiency gains.
Questions for authors
  1. Table 3 shows increasing success rates as the experience bank expands. Would further expansion cause saturation or degradation from irrelevant or redundant experience? How would Harness VLA perform if it used the same expanded reference experience to update its task-specific and global memories? The current study increases both experience volume and task coverage. Could the authors vary only experience volume while fixing task coverage, retrieval parameters, and the Memory Card budget? Further clarification or targeted experiments addressing these concerns and Weaknesses 2–3 could lead me to raise my score.
  2. GPT-6 Astra reportedly performs well on embodied tasks. If feasible, could the authors provide end-to-end evaluations on these benchmarks using GPT-6 Astra as the planner?
5 comments
AuthorsOpenReview ↗

Technical Contribution, Controlled Comparisons, Efficiency, and Experience Scaling

W1: Limited technical novelty

Reviewer’s concern: CARE-VLA appears to be an incremental extension of Harness VLA, adding a step-level experience bank and retrieval, filtering, and evidence-selection mechanisms. Although the design is reasonable, the technical advance appears limited.

Our response:

Thank you for your suggestion. We would like to clarify the technical significance of this work from three perspectives: its core contribution, experimental results, and future value.

Core contribution. CARE-VLA segments reference execution trajectories into step-level experience that can be reused across tasks. Through execution-context matching, multimodal ranking, and diversity-based selection, it provides up to three relevant pieces of evidence for the current decision. Its contribution is to enable historical execution experience to support current action selection and local recovery while keeping the planner fixed and the VLA frozen.

Experimental results. In the Astra controls with the same original memories and added experience bank, CARE-VLA improves the success rate from 85.8% with full-bank exposure to 89.3%, supporting the role of evidence selection (Appendix E.2, Table 9). The mean latency from retrieval to Memory Card readiness is only 119.3 ms, and primitive usage decreases by 8.9% on episodes successfully completed by both methods (Appendix F and Table 6).

Future value. CARE-VLA has shown that reusing execution experience can improve decisions without updating the planner or VLA weights. In the future, newly generated and verified execution trajectories could be segmented into step-level records and added to the experience bank online after quality filtering and deduplication, allowing robots to continuously accumulate reusable execution knowledge during deployment. Drawing on the external-memory perspective of RAG (Lewis et al., NeurIPS 2020), indexing methods such as HNSW (Malkov and Yashunin, TPAMI 2020) could support efficient retrieval from larger experience banks while maintaining an input budget of at most KK pieces of evidence per decision, enabling further exploration of experience sharing across tasks and scenes.

W2.1: Dependence on the planner or underlying model

Reviewer’s concern: CARE-VLA (Codex) outperforms Harness VLA (Codex) on LIBERO-Pro but trails Harness VLA (CC) on most suites and overall. Could the gains depend on the coding agent or underlying model? Could both methods be compared using Claude Code or Codex with another model?

Our response:

Thank you for your suggestion. The official Harness VLA paper has reported evaluation results with GPT-6 Astra, achieving 92.63% across all eight LIBERO-Pro suites and an average success rate of 86.0% on Goal-T/S and Long-T/S (Harness VLA, Appendix H, Table 21). As GPT-5.5 was no longer available during the rebuttal period, we conducted additional matched comparisons using GPT-6 Astra with low reasoning on these four suites. CARE-VLA achieves 89.3%, improving over our internally reproduced Harness VLA baseline (83.5%) by 5.8 percentage points (our paper, Appendix E.2, Table 9). Due to the rebuttal time window, service availability, and evaluation costs, we were unable to complete matched experiments with Claude Code.

W2.2: Matched experimental conditions and direct controls

Reviewer’s concern: The main comparisons use published baseline results. How are Codex CLI versions, GPT-5.5 snapshots, prompts, memories, and other experimental conditions matched? Could the added retrieval module be disabled in the same implementation for a direct controlled comparison? Please clarify whether such tests were conducted or historical logs with identical configurations were reused.

Our response:

Thank you for your suggestion. We have added internally matched GPT-5.5 baselines and controls with retrieval disabled under Astra to examine the overall performance gains and the role of the retrieval mechanism, respectively.

1. Configuration matching and internal reproduction

Our implementation is based on the official Harness VLA repository. CARE-VLA and the internally reproduced baseline use the same Codex CLI version, GPT-5.5 model identifier, reasoning settings, base prompts, original task-specific and global memories, VLA backend, primitive implementations, feedback rules, task–seed pairs, and execution budgets. LIBERO-Pro uses πRLinf\pi_{\mathrm{RLinf}} with medium reasoning; RoboTwin C2R uses LingBot-VLA with xhigh reasoning.

The exact CLI version and model snapshot used in the official historical runs were not disclosed. We therefore report the published results and internally matched results separately (Appendix E.1, Table 8):

MethodSpat-TSpat-SObj-TObj-SGoal-TGoal-SLong-TLong-SLIBERO-Pro OverallRoboTwin C2R
Harness VLA (GPT-5.5, reported)816994917566524972.158.0
Harness VLA (GPT-5.5, reproduced)837094926869554471.957.2
CARE-VLA (GPT-5.5)929093958075625580.361.6

Values are computed from success counts and rounded to one decimal place. The primitive usage statistics in Table 6 also come from the execution records of these internally reproduced runs. Thus, the performance gains are also supported by internally matched comparisons.

2. Controls with the same experience and retrieval disabled

As GPT-5.5 was no longer available during the rebuttal period, the additional experiments use GPT-6 Astra with low reasoning on Goal-T/S and Long-T/S, matching the planner, VLA, software, task–seed pairs, execution budgets, and evaluation protocol:

MethodGoal-TGoal-SLong-TLong-SOverall
Harness VLA (Astra, reported)8899857286.0
Harness VLA (Astra, reproduced)8495837283.5
Without retrieval (full-bank exposure)8795877485.8
Replace memories (direct replacement)4564434248.5
CARE-VLA (Astra)9298897889.3

The added experience bank is constructed from one non-evaluation reference episode for each of 40 tasks, segmented into step-level records. Full-bank exposure, memory replacement, and CARE-VLA use the same added experience. All settings retain the original memories except the replacement control. Making the full bank “accessible” does not mean inserting all records into a single prompt.

CARE-VLA improves over the internal baseline by 5.8 percentage points. With the same original memories and added experience, it improves over the retrieval-disabled control by 3.5 percentage points, supporting the contribution of selective retrieval beyond increased experience availability. Memory replacement achieves only 48.5%, further supporting the complementarity between step-level evidence and the original memories (Appendix E.2, Table 9).

In summary, the internal controls match the CLI version, model identifier, reasoning settings, base prompts, original memories, VLA, and execution and evaluation settings. However, we cannot confirm that the exact CLI version and model snapshot match those used in the official historical runs. We therefore assess the performance gains against the internally reproduced baseline: CARE-VLA improves by 8.4 and 4.4 percentage points on the two benchmarks with GPT-5.5, and by 5.8 percentage points with Astra. With the same experience bank, it also improves over the retrieval-disabled control by 3.5 percentage points, supporting the role of retrieval and evidence selection.

W3: End-to-end efficiency

Reviewer’s concern: The paper reports retrieval-core latency and fewer primitive executions, but does not account for Memory Card rendering or additional planner inference time. Do the savings from fewer planning steps and primitive executions outweigh these costs? A comparison of end-to-end task duration under matched conditions would clarify the practical efficiency gains.

Our response:

Thank you for your suggestion. According to the logs, agent inference calls take approximately 10–40 seconds, including network transfer and waiting overhead. Primitive execution time is also affected by hardware and service load. As these conditions were not strictly matched across the historical runs, we cannot use them to provide a controlled end-to-end runtime comparison. We therefore address end-to-end runtime from another perspective, analyzing efficiency in terms of the added retrieval cost and the reduction in decision cycles.

Using the current implementation, we replayed 10,553 decision queries from 800 episodes in the historical execution logs to provide an additional end-to-end latency measurement for the retrieval stage. On a system with eight RTX 4090 GPUs and warmed models, timing starts at query preparation and ends when the Memory Card text and images are ready. This covers visual encoding, retrieval, evidence selection, and card rendering. The mean latency is 119.3 ms, and P95 is 168.54 ms (Appendix F; Appendix I, Table 14).

Across the 548 episodes successfully completed by both methods, the average number of primitive executions decreases from 15.34 to 13.97, an 8.9% reduction (Table 6). Assuming one retrieval before each primitive, this adds approximately 1.67 seconds of retrieval overhead per episode while eliminating 1.37 planning, execution, and feedback cycles.

Let tbaset_{\mathrm{base}} denote the average duration of a base decision cycle, calculated by dividing the cumulative time spent on planner inference, primitive execution, and feedback processing by the number of cycles. The condition for offsetting the retrieval cost can be estimated as:

\frac{13.97 \times 0.1193}{15.34 - 13.97} \approx 1.22 \mathrm{s}.$$ Agent inference calls alone currently take 10–40 seconds, well above this threshold, even before accounting for primitive execution and feedback processing. Therefore, under the above conditions, the reduction in decision cycles can offset the added retrieval overhead and yield time savings. In summary, Memory Card rendering costs are included in the supplementary measurements, the reduction in execution steps is empirically supported, and the conditional analysis further illustrates the feasibility of offsetting the added overhead. In future work, we will further quantify the additional inference overhead introduced by the Memory Card under matched hardware, service deployment, and concurrency conditions, and conduct paired evaluations of total task duration to study CARE-VLA's end-to-end efficiency gains more systematically.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.