acceptodds
Under review as a conference paper at ICLR 2027

Inline Memory Meets Reusable Skills: Memory-centric Framework for Vision-Language-Action Model

Abstract

Vision-Language-Action (VLA) models have shown strong promise for general-purpose robotic manipulation, yet adapting them to new tasks and domains remains inefficient: existing methods often rely on parameter tuning, incurring substantial costs and risking catastrophic forgetting of previously learned tasks. To address this, we propose Optimus-R, a memory-centric VLA framework that formulates robotic adaptation as explicit query-skill memory tuning. Optimus-R introduces: (i) An Inline Memory Interface for skill extraction. It inserts learnable memory tokens into the VLA prefix stream, allowing the backbone to derive control-aware query and skill representations within the native action-conditioning pathway. (ii) A Query-Skill Memory Bank for skill learning. It externalizes skills into query prototypes for deciding what to retrieve and skill values for specifying how to act, supporting skill reuse and expansion with limited parameter updates. (iii) A lightweight Bridge-and-Adapt mechanism for skill updating. It aligns target-domain queries and skills with the existing memory space through a lightweight adapter and residual memory updates. Experiments on in-domain adaptation, cross-domain adaptation, and lifelong learning show that Optimus-R enables data-efficient skill learning while mitigating catastrophic forgetting.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Model, Robotic Manipulation

Reviews

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

This paper introduces Optimus-R, a memory-centric VLA framework that formulates robotic adaptation as explicit query-skill memory tuning. The framework introduces an Inline Memory Interface for skill extraction, a Query-Skill Memory Bank for skill learning, and a lightweight Bridge-and-Adapt mechanism for skill updating. Experiments on in-domain adaptation, cross-domain adaptation, and lifelong learning show that Optimus-R enables data-efficient skill learning while mitigating catastrophic forgetting.

Critical strengths
  1. The paper proposes an effective memory-centric framework for VLA adaptation, addressing an important problem in robot learning. The framework demonstrates computational efficiency and can be applied to multiple adaptation settings.

  2. The empirical evaluation shows consistent improvements in success rates across different settings, including in-domain adaptation, cross-domain adaptation, and lifelong learning. The method is evaluated on multiple simulation benchmarks (LIBERO, CALVIN, and RoboTwin 2.0) and nine real-world manipulation tasks.

Critical weaknesses
  1. Insufficient evidence to support the core claims.

The paper claims that the proposed memory-centric framework enables efficient adaptation while mitigating catastrophic forgetting, but the experiments do not fully substantiate these claims. In particular, Stage B updates the VLA backbone during cross-domain adaptation, making it difficult to isolate the contribution of the memory mechanism from conventional fine-tuning. It also remains unclear whether the framework can effectively mitigate forgetting under domain shifts, as the forgetting analysis is limited to in-domain lifelong learning. Furthermore, the current ablations do not establish whether the explicit Query-Skill Memory Bank is necessary, as there is no comparison with memory-free latent conditioning under a matched training protocol. Consequently, it remains unclear whether the reported improvements primarily stem from the proposed memory mechanism or from additional model adaptation and training.

  1. Insufficient comparisons with relevant baselines and unclear novelty.

The experimental comparisons are insufficient to establish the advantages of the proposed framework over existing continual learning and memory-based adaptation methods. In the lifelong learning experiments, the main direct baseline is $\pi$0.5 fine-tuning, without comparisons to established continual learning approaches such as experience replay, regularization-based methods, or task-specific adapters. Moreover, the proposed approach shares conceptual similarities with prompt tuning and existing memory-based methods, but its distinction from these approaches is not sufficiently clarified or experimentally validated. The current experiments also do not adequately distinguish whether new-task improvements arise from reusing previously learned skills, expanding the memory bank with new prototypes, or latent exemplar replay. These limitations weaken the evidence for the claimed benefits and novelty of the proposed framework.

Questions for authors
  1. Some symbols in Figure 2 are difficult to interpret, particularly the question marks shown in the memory bank and retrieval components. What do these symbols represent? Are they placeholders for unknown values?
  2. How does Optimus-R compare with recent continual learning methods, particularly replay-based approaches or task-specific adapters?
  3. Many of the real-world tasks involve pick-and-place operations or their compositions. How well does the proposed memory mechanism generalize to more complex tasks requiring long-term dependencies or diverse skill compositions?
4 comments
AuthorsOpenReview ↗

Response to Reviewer njgE (1/2)

Thank you for the thoughtful review and constructive suggestions. We have responded to your concerns as follows, and we have revised the manuscript accordingly. We would greatly appreciate it if you could consider giving us a higher rating.

W1: In particular, Stage B updates the VLA backbone during cross-domain adaptation, making it difficult to isolate the contribution of the memory mechanism from conventional fine-tuning. Furthermore, the current ablations do not establish whether the explicit Query-Skill Memory Bank is necessary, as there is no comparison with memory-free latent conditioning under a matched training protocol.

R1: (1) The comparison with π0.5 measures the gain from adding the complete memory mechanism. Optimus-R conditions the action expert on retrieved skills while preserving π0.5’s inference inputs and outputs. Tables 1 and 2 of the main text compare Optimus-R with π0.5 at matched demonstration budgets. These comparisons evaluate the complete memory-augmented adaptation framework. Table 4 of the main text and Appendix G.1 then ablate individual components and training methods (refer to R1 of Reviewer MHpG).

(2) Stage B updates the backbone to adapt perception and control to the target domain. In Table 4 of the main text, the two Bridge variants fine-tune the same backbone parameters. Adding the Bridge raises real-world success from 35.6% to 61.7% with 50 demonstrations per task (refer to R4 of Reviewer TEuW). In the training cost comparison in Table 3 of the main text, Optimus-R uses 86 GPU hours versus 128 for π0.5-LoRA, with about 0.3B trainable parameters each (refer to R3 of Reviewer MHpG).

(3) Table R4 compares bank retrieval with direct latent conditioning. Both variants start from π0.5-base and train on LIBERO-Long. The memory-free variant conditions the action expert on latents derived directly from the backbone’s memory tokens, while Optimus-R conditions the action expert on skills retrieved from the Query–Skill Memory Bank. Bank retrieval achieves 94.2% success versus 91.4% for direct latent conditioning, a gain of 2.8 percentage points.

Table R4. Bank retrieval versus direct latent conditioning on LIBERO-Long.

VariantSuccess rate (%)
Memory-free latent conditioning91.4
Optimus-R94.2

W2: It also remains unclear whether the framework can effectively mitigate forgetting under domain shifts, as the forgetting analysis is limited to in-domain lifelong learning.

R2: (1) We measure forgetting across task changes in simulation and in the real world. In Table R2 (refer to R5 of Reviewer TEuW), we evaluate LIBERO-Spatial before and after learning LIBERO-Goal. Table 13 in Appendix G.5 also reports 0.009 NBT over ten successive LIBERO-Spatial tasks (refer to R2 of Reviewer MHpG). In the real-world evaluation, Figure 4 of the main text measures how Category A success changes after learning Category B.

(2) The sim-to-real experiment evaluates adaptation to the real-world target tasks. In Table 2 of the main text, we adapt RoboTwin-trained models to the real world and evaluate the target tasks. With 20 real demonstrations per task, Optimus-R reaches 33.9% versus 19.4% for π0.5. The before-and-after retention comparisons are reported in Table R2 and Figure 4 for the Spatial-to-Goal and Category A-to-B sequences, respectively.

W3 & Q2: In the lifelong learning experiments, the main direct baseline is π0.5 fine-tuning, without comparisons to established continual learning approaches such as experience replay, regularization-based methods, or task-specific adapters. How does Optimus-R compare with recent continual learning methods, particularly replay-based approaches or task-specific adapters? The current experiments also do not adequately distinguish whether new-task improvements arise from reusing previously learned skills, expanding the memory bank with new prototypes, or latent exemplar replay.

AuthorsOpenReview ↗

Response to Reviewer njgE (2/2)

W3 & Q2: In the lifelong learning experiments, the main direct baseline is π0.5 fine-tuning, without comparisons to established continual learning approaches such as experience replay, regularization-based methods, or task-specific adapters. How does Optimus-R compare with recent continual learning methods, particularly replay-based approaches or task-specific adapters? The current experiments also do not adequately distinguish whether new-task improvements arise from reusing previously learned skills, expanding the memory bank with new prototypes, or latent exemplar replay.

R3: (1) Optimus-R has lower measured forgetting than both evaluated continual-learning baselines. In Table R3 (refer to R7 of Reviewer TEuW), all three methods use the π0.5 backbone. Optimus-R has lower NBT and higher final success in the ten-task simulation sequence. In the real world, its Category A forgetting is 5.00 points versus 8.75 and 10.00, while its Category B success is 57.0% versus 55.0% and 53.0%. It therefore has the lowest measured forgetting and the highest new-task success, exceeding raw replay by two points and prototype-routed adapters by four points. Moreover, in Table 9 of Appendix G.1, removing exemplar storage reduces full-data LIBERO success from 97.8% to 96.3%, supporting its contribution in that separate setting.

(2) Cross-suite transfer provides evidence consistent with selective skill reuse before bank expansion. In Table 12 of Appendix G.4, we train on LIBERO-Spatial and test on LIBERO-Goal. Before Goal training, the shared bowl-on-plate operation reaches 68% success versus 8% suite-wide. After adaptation, suite-wide success reaches 96% and the bank grows from 16 to 27 entries. The zero-shot result is consistent with transfer of a familiar operation, and the broader gains coincide with bank expansion and target-task training. As described in Sec. 3.4 of the main text and Appendix C.4, a close match updates an existing prototype and a sufficiently supported novel cluster adds an entry.

W4: Moreover, the proposed approach shares conceptual similarities with prompt tuning and existing memory-based methods, but its distinction from these approaches is not sufficiently clarified or experimentally validated.

R4: The revision discusses L2P and IsCiL in Sec. 2 of the main text (refer to R6 of Reviewer TEuW). L2P retrieves prompt tokens, and IsCiL retrieves a prototype to select a skill-specific adapter. As described in Secs. 3.1–3.3 of the main text, Optimus-R retrieves an action-supervised skill value that conditions a shared action expert. Table R3 (refer to R7 of Reviewer TEuW) reports lower forgetting than the prototype-routed adapter baseline, with higher final simulation success and higher real-world new-task success. Table R4 reports higher LIBERO-Long success than direct latent conditioning.

Q1: Some symbols in Figure 2 are difficult to interpret, particularly the question marks shown in the memory bank and retrieval components. What do these symbols represent? Are they placeholders for unknown values?

R5: The small lowercase qq in Figure 2 of the main text denotes a query. The question marks may result from how the PDF viewer renders the figure font. We verified that the labels render correctly in Google Chrome. The current query is qtq_t, and each bank entry pairs the retrieval key pjqp_j^q with its skill value pjsp_j^s. The shapes in the retrieval panel illustrate groups of latent embeddings. As described in Sec. 3.3 of the main text, the aligned query selects keys whose skill values condition the action expert.

Q3: Many of the real-world tasks involve pick-and-place operations or their compositions. How well does the proposed memory mechanism generalize to more complex tasks requiring long-term dependencies or diverse skill compositions?

R6: (1) As described in Appendix F, Category B tests bimanual coordination and sequences of dependent actions. Table 8 and Figure 8 in Appendix F show tasks such as Sequential Pick and Place, which requires moving multiple objects to different targets in order. In Table 2 of the main text, we compare methods with 80 demonstrations per task. Optimus-R reaches 55% on Sequential Pick and Place and 50% on Bimanual Coordination Sequential Task, versus 45% and 40% for π0.5. These results show gains on both coordinated and sequential tasks.

(2) Category A tests whether the same operation transfers across object instances and layouts. The operations are listed in Table 8 in Appendix F. At each rollout, we vary lighting and object shape, color, and position. In the 80-demo comparison in Table 2 of the main text, Optimus-R reaches 85–90% success across the four Category A tasks under these variations.

Thank you for the responses and additional experiments. I appreciate the new comparisons with memory-free latent conditioning and continual learning baselines. These results partly address my concerns.

I'm still concerned about the novelty and advantages of the proposed memory mechanism. While Table R4 demonstrates improvements over direct latent conditioning, the gain is relatively modest and evaluated only on LIBERO-Long. Although the cross-task transfer results are promising, they do not fully isolate the contribution of reusing existing skill prototypes from the generalization capabilities of the pretrained VLA.

Overall, the additional experiments strengthen the empirical evaluation, but I believe the evidence for the distinct strengths of the proposed memory-centric design remains somewhat limited. I will maintain my original score.

Thank you for your follow-up comments. We address the remaining concerns below.

(1) Our comparisons evaluate the memory mechanism against both π0.5 and direct latent conditioning. Tables 1 and 2 of the main text compare the complete framework with π0.5, while the memory-free control tests bank retrieval against direct conditioning. We have extended this comparison to all four LIBERO suites.

Table R5. LIBERO success rates (%). The π0.5 and Optimus-R entries are from the full-data setting in Table 1 of the main text.

MethodSpatialObjectGoalLongAverage
π0.598.898.298.092.496.9
Memory-free latent conditioning98.698.497.691.496.5
Optimus-R99.299.498.494.297.8

The experimental results show that direct latent conditioning degrades VLA performance, whereas explicit skill modeling and the use of a memory bank improve it.

(2) The component ablations test how the memory mechanism learns, organizes, and adapts skills. In Table 4 of the main text, removing Inline Memory tests the inline representation pathway, coupling query and skill tests their separation, removing prototype residuals tests local skill refinement, and removing the Stage-B Bridge tests cross-domain alignment under the same backbone updates. In Table 9 of Appendix G.1, removing skill-aware clustering tests prototype construction, removing exemplar storage tests latent replay, and uniform downsampling tests motion-aware sample selection. Each ablation reduces success. Table 10 of Appendix G.2 additionally examines sensitivity to bank capacity and retrieval configuration.

(3) The π0.5 comparison measures transfer beyond the pretrained VLA baseline. In Table 15 of Appendix G.7, both methods start from π0.5-base and follow the same Spatial→Goal sequence with matched demonstration budgets and training steps. Optimus-R achieves 95.0% SR on Goal vs. 93.8% for π0.5 and reduces Spatial forgetting from 7.2 to 2.8 points. These results demonstrate additional adaptation and retention benefits over the pretrained baseline.

We hope these clarifications address the reviewer’s concerns, and we would be happy to provide further clarification if the reviewer has any additional concerns. If your concerns have been addressed, we would greatly appreciate it if you could consider giving us a higher rating.

Reviewer_TEuWOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes Optimus-R, a memory-centric framework for adapting VLA models. Learnable memory tokens are inserted into the prefix of π0.5, and pooled readouts yield a query and a skill embedding. Clustered query-skill prototypes are stored in a memory bank, retrieved by similarity, and projected into tokens that condition the flow policy. Three training stages pretrain this interface, align target-domain representations for sim-to-real transfer, and update or expand the bank for lifelong learning. Experiments on LIBERO, CALVIN, RoboTwin 2.0, and nine real-world tasks show gains over π0.5 in low-data, sim-to-real, and lifelong settings.

Critical strengths

The evaluation is broad: three simulation benchmarks (LIBERO, CALVIN, RoboTwin 2.0), nine real-world tasks on a bimanual robot, several data budgets, ablations in simulation and the real world (Table 4), and a cost comparison with full fine-tuning and LoRA (Table 3). Compared to π0.5 fine-tuned with the same data budgets, the low-data gains are large and consistent, e.g., 88.6% vs. 72.1% on LIBERO with 30% of the data and 33.3% vs. 19.4% in the real world with 20 demonstrations per task. If these gaps hold under a clean protocol (see below), they would be a useful result for data-efficient VLA adaptation.

Critical weaknesses
  • The comparisons behind all three research questions have issues. For Q1, Section 4.2 adapts "the same Stage-A pretrained model" with 30% to 100% of the demonstrations, but the paper never states which data Stage A uses. Since Stage A also trains the last backbone layers, the low-data gains would partly reflect extra data if Stage A saw the full training set. For Q2, it is not stated whether π0.5 and the other baselines in Table 2 received the same RoboTwin 2.0 pretraining as Optimus-R. Also, Stage B trains the full backbone on the real-world data (Table 5), so the transfer gains cannot be attributed to the memory alone, and it is unclear whether the "w/o Stage-B Bridge" variant (12.8% vs. 33.3% with 20 demos, Table 4) also drops the backbone fine-tuning. For Q3, the new-task success rates in Figure 4 are exactly 5 points (one of 20 rollouts) below the corresponding Category-B entries of Table 2 for 47 of the 50 method, budget and task combinations, and two of the other three are already 0 in Table 2. Independent real-robot evaluations would not produce such a constant offset, and I found no explanation in the paper. Even at face value, the reduction in forgetting from 15.0 to 10.0 points rests on 80 rollouts before and after adaptation, which gives a binomial standard error of roughly 9 points for the difference. More generally, no result comes with multiple seeds, standard deviations, or confidence intervals. The low-data gains are well outside binomial noise, but other claims are not, e.g., 97.8% vs. 96.9% on LIBERO with full data. This is the main reason for my rating, since each headline claim rests on one of these comparisons.
  • Conceptually, the method is close to existing retrieval-based continual learning, and some close baselines are missing. Querying a memory with features from a frozen pretrained model and inserting the retrieved entries into the input sequence is the core idea of prompt-pool methods such as L2P (Wang et al., 2022), and IsCiL (Lee et al., 2024) retrieves incrementally learned skills through prototypes in a state-embedding space for continual imitation learning. Neither is discussed in Section 2. The VLA-specific design is a reasonable contribution, but the paper should position it against these methods. Moreover, no continual learning strategy is described for the baselines in Figure 4, which suggests plain sequential fine-tuning, whereas Optimus-R replays stored latent pairs. Liu et al. (2026), whose task order Appendix G.5 adopts, report that simple experience replay with two pretrained VLAs already reaches NBT values of 0.019 and 0.007 on LIBERO-Spatial. The NBT of 0.009 in Appendix G.5, reported without any baseline, therefore does not show that the memory bank is what limits forgetting. This matters because mitigating forgetting through explicit memory is a central claim of the paper.

References:

  • Z. Wang, Z. Zhang, C.-Y. Lee, H. Zhang, R. Sun, X. Ren, G. Su, V. Perot, J. Dy, and T. Pfister, "Learning to prompt for continual learning", IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2022.
  • D. Lee, M. Yoo, W. K. Kim, W. Choi, and H. Woo, "Incremental learning of retrievable skills for efficient continual task adaptation", Advances in Neural Information Processing Systems (NeurIPS), 2024.
  • H. Liu, C. Kim, B. Liu, M. Liu, and Y. Zhu, "Pretrained vision-language-action models are surprisingly resistant to forgetting in continual learning", International Conference on Machine Learning (ICML), 2026.
Questions for authors

Clarification requests:

  • How were the new-task success rates in Figure 4 obtained? Besides the constant offset to Table 2, the forgetting values in Figure 4 are reproduced exactly if S_before is taken from the Category-A entries of Table 2. Please state which model and rollouts each number comes from and, if possible, share per-episode evaluation records.
  • Which data is Stage A trained on in each experiment? If the Stage-A model for the 30-70% settings in Table 1 has seen the full training set, please report results in which Stage A uses only the same subset as π0.5. Also, do the 86 GPU hours in Table 3 include Stage A?
  • Were the baselines in Table 2 and Figure 4 pretrained on RoboTwin 2.0 before real-world fine-tuning, and how were they trained in the lifelong setting?
  • Is the backbone still fine-tuned on real-world data in the "w/o Stage-B Bridge" variant? If not, a variant that fine-tunes the backbone but drops A_mem and the Skill Adapter would isolate the effect of the alignment.
  • Could you report results over multiple training seeds with standard deviations, at least in simulation, and confidence intervals for the real-world success rates?

Suggestion for future work:

  • Comparing against continual learning baselines on the same π0.5 backbone, e.g., experience replay with a small buffer of raw demonstrations and a retrieval-based method like those above, both in the real-world lifelong setting and on the LIBERO-Spatial sequence of Appendix G.5.
4 comments
AuthorsOpenReview ↗

Response to Reviewer TEuW (1/3)

Thank you for the thoughtful review and constructive suggestions. We have responded to your concerns as follows, and we have revised the manuscript accordingly. We would greatly appreciate it if you could consider giving us a higher rating.

W1 & Q1: How were the new-task success rates in Figure 4 obtained? Besides the constant offset to Table 2, the forgetting values in Figure 4 are reproduced exactly if S_before is taken from the Category-A entries of Table 2. Please state which model and rollouts each number comes from and, if possible, share per-episode evaluation records.

R1: We checked the rollout records and found three reporting errors in Figure 4 and Table 2 of the main text. Thank you for identifying the discrepancies. Both experiments start from RoboTwin 2.0 checkpoints and use the same real-world evaluation setup, with separate CSVs recording each of the 20 rollouts per task and checkpoint as a success (1) or failure (0). Table 2 of the main text evaluates a policy trained jointly on all nine tasks. Figure 4 of the main text evaluates a policy that learns Category A first and then adapts to Category B. However, we made the following errors when converting the raw data into experimental results:

(1) Table 2 of the main text undercounted one Optimus-R success on Sequential Pick and Place. At 20 demonstrations per task, the original records contain three successes out of 20. The table counted two and therefore reported 10% instead of 15%.

(2) The new-task panel in Figure 4 of the main text used the cross-domain rollout files. The original result-aggregation pipeline filtered for successful episodes and then misread the first success in each nonempty list as a header. Dropping one success out of 20 accounts for 47 five-point offsets. Two entries with no successes stayed at zero. For Optimus-R on Sequential Pick and Place with 20 demonstrations, the aggregation pipeline and the table each counted two of the three successes, so both reported 10%. The corrected panel uses Category B rollouts from the checkpoint after lifelong adaptation.

(3) The original forgetting score was incorrectly calculated using the cross-domain adaptation results. Table R1 gives the Category A scores before learning Category B, and the forgetting values are recalculated from the corrected before-adaptation scores.

We have corrected these errors in the revision. The corrected results retain the advantages in new-task success and old-task retention.

Table R1. Category A success rates (%) before learning Category B. Each task is evaluated over 20 rollouts.

MethodDemos per taskR.O.B.S.P.P.P.T.Average
RDT8015.045.025.055.035.00
OpenVLA8010.030.050.045.033.75
OpenVLA-OFT8070.065.075.080.072.50
π08065.080.070.090.076.25
π0.52020.020.030.015.021.25
π0.55055.060.060.065.060.00
π0.58080.065.090.075.077.50
Optimus-R2045.040.055.040.045.00
Optimus-R5065.075.085.070.073.75
Optimus-R8085.080.085.080.082.50

R.O.: Reorientation; B.S.: Block Stacking; P.P.: Place Object onto Plate; P.T.: Place Object onto Tablecloth.

W2 & Q2: Which data is Stage A trained on in each experiment? If the Stage-A model for the 30-70% settings in Table 1 has seen the full training set, please report results in which Stage A uses only the same subset as π0.5. Also, do the 86 GPU hours in Table 3 include Stage A?

R2: (1) Each low-data run uses its own demonstration subset for Stage A, initial bank construction, and adaptation. In Table 1 of the main text, the 30%, 50%, and 70% settings each use only their allotted subset. On LIBERO at 30%, Optimus-R reaches 88.6% versus 72.1% for π0.5. In the sim-to-real experiment, Stage A and the initial bank use RoboTwin 2.0 source data, and the 20, 50, or 80 real demonstrations per task are used for target adaptation. In the lifelong experiment, Category B demonstrations enter only after Category A training.

(2) The reported 86 GPU hours include Stage A and initial bank construction. In Table 3 of the main text, Optimus-R uses 86 GPU hours on full-data LIBERO versus 128 for π0.5-LoRA. The cost difference includes preparing the memory.

AuthorsOpenReview ↗

Response to Reviewer TEuW (2/3)

W3 & Q3: Were the baselines in Table 2 and Figure 4 pretrained on RoboTwin 2.0 before real-world fine-tuning, and how were they trained in the lifelong setting?

R3: (1) Optimus-R and all baselines initialize from RoboTwin 2.0 checkpoints. For Table 2 of the main text, the methods use the same RoboTwin 2.0 source data before joint fine-tuning on all nine real tasks. Figure 4 of the main text also starts from RoboTwin-trained checkpoints and uses a sequential real-world evaluation protocol.

(2) The original lifelong baselines use sequential fine-tuning without replay. After learning Category A, they fine-tune only on Category B demonstrations. Optimus-R also uses only Category B demonstrations for the action loss. Its stored query–skill pairs enter the memory alignment loss (refer to Appendix D.1).

W4 & Q4: Is the backbone still fine-tuned on real-world data in the "w/o Stage-B Bridge" variant? If not, a variant that fine-tunes the backbone but drops A_mem and the Skill Adapter would isolate the effect of the alignment.

R4: Yes, the variant without the Stage-B Bridge fine-tunes the same backbone parameters as Optimus-R. In Table 4 of the main text, removing the Stage-B Bridge changes sim-to-real success with 50 demonstrations per task from 61.7% to 35.6%. The sim-to-real drop shows that the Bridge contributes to cross-domain adaptation under matched backbone updates.

W5 & Q5: Even at face value, the reduction in forgetting from 15.0 to 10.0 points rests on 80 rollouts before and after adaptation, which gives a binomial standard error of roughly 9 points for the difference. Could you report results over multiple training seeds with standard deviations, at least in simulation, and confidence intervals for the real-world success rates?

R5: (1) Across three LIBERO evaluation seeds, Optimus-R shows less mean forgetting on LIBERO-Spatial and higher mean success on LIBERO-Goal. We extend the evaluation of the Spatial-to-Goal setup in Appendix G.4. Both methods start from π0.5-base and learn Spatial before Goal with matched demonstration budgets and training steps. We evaluate the corresponding fixed checkpoints using evaluation seeds 0, 42, and 7. Each suite is evaluated with 500 rollouts per seed. Table R2 reports the mean and sample standard deviation across these evaluation runs. The experimental results show that Optimus-R has lower forgetting than π0.5 baseline.

(2) Optimus-R combines higher new-task success with stronger retention in the real-world evaluation. With 80 demonstrations per task in Figure 4, Optimus-R retains 77.50% success on Category A versus 63.75% for π0.5 and achieves 57.0% on Category B versus 44.0% for π0.5. Its average forgetting is 5.00 percentage points, compared with 13.75 for π0.5: an 8.75-point reduction alongside a 13.0-point gain in new-task success.

(3) We report 95% Wilson confidence intervals for the real-world success rates. For the same Optimus-R run, Category A success is 82.50% before adaptation (95% CI: 72.7–89.3%) and 77.50% afterward (67.2–85.3%). Category B success is 57.0% (47.2–66.3%).

Table R2. LIBERO-Spatial → LIBERO-Goal. Success rates are mean ± sample SD over three evaluation seeds; forgetting is in percentage points.

MethodSpatial before Goal (%)Spatial after Goal (%)Goal after adaptation (%)Mean Spatial forgetting (pp) ↓
π0.5 sequential fine-tuning90.8 ± 1.083.6 ± 1.693.8 ± 1.47.2
Optimus-R93.4 ± 0.890.6 ± 1.295.0 ± 1.02.8

W6: Querying a memory with features from a frozen pretrained model and inserting the retrieved entries into the input sequence is the core idea of prompt-pool methods such as L2P (Wang et al., 2022), and IsCiL (Lee et al., 2024) retrieves incrementally learned skills through prototypes in a state-embedding space for continual imitation learning. Neither is discussed in Section 2.

R6: (1) L2P retrieves prompts, and IsCiL uses retrieval to select skill adapters. L2P matches frozen image features to learned keys and prepends the corresponding prompt tokens for continual classification. IsCiL matches the encoded observation and language subgoal to skill prototypes and activates the corresponding LoRA adapter.

(2) Optimus-R retrieves action-supervised skill values to condition a shared action expert. Future-action reconstruction supervises the skill values, and the retrieved values are projected into tokens for the shared π0.5 action expert. The policy can therefore reuse the same expert as skills are added or refined. We added this comparison to Sec. 2 of the main text in the revision.

AuthorsOpenReview ↗

Response to Reviewer TEuW (3/3)

W7 & Q6: The NBT of 0.009 in Appendix G.5, reported without any baseline, therefore does not show that the memory bank is what limits forgetting. Comparing against continual learning baselines on the same π0.5 backbone, e.g., experience replay with a small buffer of raw demonstrations and a retrieval-based method like those above, both in the real-world lifelong setting and on the LIBERO-Spatial sequence of Appendix G.5.

R7: (1) We compare raw experience replay and IsCiL-style adapter retrieval on the same π0.5 backbone. Raw replay mixes current training data with a buffer of earlier demonstrations. Adapter retrieval uses skill prototypes to select a skill-specific LoRA adapter. Within each setting, we match the backbone initialization, task order, current-task demonstration budget, optimizer steps, and evaluation protocol. We use the ten-task LIBERO-Spatial sequence from Appendix G.5 with 10,000 steps per task, and the real-world Category A → B protocol from Figure 4 of the main text with 80 demonstrations per task.

(2) Optimus-R has the lowest reported forgetting and the highest real-world new-task success among these methods. In Table R3, Optimus-R reaches 97.6% final Spatial success with 0.009 NBT, versus 97.2% SR and 0.013 NBT for experience replay and 95.4% SR and 0.020 NBT for adapter retrieval. In the real-world evaluation, Optimus-R has lower forgetting than other methods. So the results show the benefit of the proposed memory mechanism.

Table R3. Comparison with continual-learning baselines on a shared π0.5 backbone.

MethodLIBERO-Spatial final success (%)LIBERO-Spatial NBT ↓Category B success (%)Category A forgetting (pp) ↓
π0.5 + raw experience replay97.20.01355.08.75
π0.5 + IsCiL-style adapter retrieval95.40.02053.010.00
Optimus-R97.60.00957.05.00

Dear Authors,

Thank you for your additional experiments and clarifications. My concerns about the limited novelty of the work with respect to existing retrieval-based continual learning approaches remain, and I am therefore keeping my rating.

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

The authors propose Optimus-R, a memory-centric VLA framework that formulates robotic adaptation as explicit query-skill memory tuning for in-domain adaptation, cross-domain adaptation, and lifelong learning tasks. The results show improvement on data-efficient skill learning while mitigating catastrophic forgetting.

Critical strengths
  • The paper is well-written and easy to follow.
  • The paper addresses an important problem, and the method introduces a novel perspective.
  • The paper provides extensive experiments across in-domain adaptation, cross-domain adaptation, and lifelong learning tasks and spans across different settings.
Critical weaknesses
  • Is the performance gain from the memory mechanism design? Table 4 provided ablation studies for several components, yet it's unclear to me about the benefit of explicit skill memory combining with the additional supervision and architectural changes.
  • The lifelong-learning experiment seems relatively limited given four initial tasks and five new tasks, it'd be great if the authors can provide more experiments on lifelong adaptation across more tasks and domains.
  • The memory-centric adaptation still requires substantial backbone fine-tuning.
Questions for authors
  • To what degree do the performance gains come from the explicit memory mechanism rather than additional supervision or backbone fine-tuning?
  • Could the authors provide stronger evidence of memory-based adaptation, particularly for lifelong learning across more tasks and domains?
1 comment
AuthorsOpenReview ↗

Response to Reviewer MHpG

Thank you for the thoughtful review and constructive suggestions. We have responded to your concerns as follows, and we have revised the manuscript accordingly. We would greatly appreciate it if you could consider giving us a higher rating.

W1 & Q1: Is the performance gain from the memory mechanism design? Table 4 provided ablation studies for several components, yet it's unclear to me about the benefit of explicit skill memory combining with the additional supervision and architectural changes. To what degree do the performance gains come from the explicit memory mechanism rather than additional supervision or backbone fine-tuning?

R1: (1) Our core contribution is a memory mechanism that learns, retrieves, and updates skills for VLA policies. The added architecture and supervision support that mechanism. Compared with the π0.5 backbone, Optimus-R adds retrieved skill as conditioning for the action expert. In Tables 1 and 2 of the main text, we compare both methods at matched demonstration budgets and settings. These comparisons evaluate the complete memory-augmented adaptation framework.

(2) The ablations test the components of the memory mechanism. In Table 4 of the main text, full-data RoboTwin Hard success drops from 55.0% to 45.2% with detached prompts, 48.0% with a shared query–skill latent, and 46.5% without prototype residuals. In Table 9 of Appendix G.1, removing skill-aware clustering or exemplar storage lowers full-data LIBERO success from 97.8% to 95.8% or 96.3%, respectively. Table R4 (refer to R1 of Reviewer njgE) compares bank retrieval with direct latent conditioning. These results show that the skill representation, bank construction, and update procedure each contribute to performance.

W2 & Q2: The lifelong-learning experiment seems relatively limited given four initial tasks and five new tasks, it'd be great if the authors can provide more experiments on lifelong adaptation across more tasks and domains. Could the authors provide stronger evidence of memory-based adaptation, particularly for lifelong learning across more tasks and domains?

R2: (1) Optimus-R retains previously learned tasks over a ten-task sequence. In Appendix G.5, we start from π0.5-base and train on ten LIBERO-Spatial tasks in sequence. We carry the model and bank forward, train for 10,000 steps per task with latent replay, and reevaluate every task seen so far after each stage. Table 13 of Appendix G.5 reports 97.6% final mean success and negative backward transfer (NBT) of 0.009, showing limited average forgetting over this sequence.

(2) The memory bank can incorporate skills from successive task suites with little added retrieval cost. In Appendix G.3, we initialize on LIBERO-90 and adapt the same model and bank to Spatial, Object, Goal, and Long in order. We measure bank size, retrieval latency, and recall after each suite. Table 11 in that appendix shows that the bank grows from 81 to 109 occupied prototypes, while retrieval latency increases from 1.80 to 1.90 ms and Recall@5 decreases from 95.6% to 94.7%. These results show that Optimus-R can expand its memory during lifelong learning while keeping retrieval latency and recall close to their initial levels.

(3) We add additional lifelong learning experiments. Tables R2 and R3 (refer to R5 and R7 of Reviewer TEuW) show Spatial-to-Goal results across three evaluation seeds and continual-learning baselines in simulation and the real world.

W3: The memory-centric adaptation still requires substantial backbone fine-tuning.

R3: (1) Backbone updates serve domain transfer, while in-domain lifelong learning keeps the backbone frozen. For cross-domain adaptation, we fine-tune selected backbone parameters to accommodate domain changes in perception and control. For lifelong learning, we acquire new tasks by updating query alignment, retrieval temperature, prototype residuals, and bank entries while holding the backbone fixed (Sec 3.4 of the main text and Table 5 in Appendix D).

(2) Optimus-R reduces the total cost of adaptation. In Table 3 of the main text, we compare training cost and success on full-data LIBERO. Optimus-R trains about 0.3B parameters and reaches 97.8% success in 86 GPU hours, while Full π0.5 fine-tuning trains 3.6B parameters and reaches 96.9% in 560 hours. π0.5-LoRA trains about 0.3B and reaches 96.5% in 128 hours. Optimus-R thus achieves higher success with about 85% fewer GPU hours than π0.5 full fine-tuning and 33% fewer than π0.5-LoRA.

Comments on the paper

1 comment

Dear AC and Reviewers,

We sincerely appreciate the Area Chair and all reviewers (#MHpG, #TEuW, #njgE) for their careful evaluation and constructive feedback. We have responded to each of concerns and provided additional experimental results to support our main claims. As the discussion deadline approaches, we would be grateful for your feedback, especially on any concerns that remain, so that we have time to respond before the discussion closes. If your concerns have been addressed, we would greatly appreciate it if you could consider giving us a higher rating.

Best,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.