acceptodds
Under review as a conference paper at ICLR 2027

LiteC-WAM:A: A Lightweight Cross-Task and -Embodiment World Action Model

Abstract

World action models (WAMs) aim to generate executable actions and predict their effects on the environment. However, key challenges remain in maintaining a lightweight model footprint, transferring learned dynamics across tasks, and adapting to new robots without forgetting previously learned capabilities. We introduce LiteC-WAM, a lightweight cross-task and cross-embodiment WAM comprising three key components: (1) a lightweight shared backbone that couples a light perception module, sparse Top-2 dynamics experts, and a flow-matching action head refined by world-model feedback; (2) cross-task adapters that transfers shared dynamics without backbone retraining; and (3) low-rank observation adapters paired with robot-specific action encoders and decoders. For lightweight deployment, LiteC-WAM achieves a 96.0% success rate with only 47.57M parameters on 5 representative LIBERO tasks, outper-forming other lightweight policies such as Diffusion Policy ( 65M, 72.4%), Octo-Base (93M, 75.1%), and SmolVLA-240M (240M, 82.75%). Its success is also comparable to billion-parameter models, such as π0 (3.3B, 94.2%), LingBot-VA (5.3B, 98.5%), and OpenVLA-OFT (7B, 97.1%). For cross-task transfer, LiteC-WAM achieves 86.67% across 15 RLBench manipulation tasks using cross-task adapters guided by language instructions, without retraining the shared backbone. For cross-embodiment transfer, LiteC-WAM achieves 82.09% closed-loop success on dual-arm SO-ARM101 and a 91.86% average offline pass rate when comparing predicted and recorded actions across seven RoboChallenge ARX5 tasks. Together, these results validate LiteC-WAM as an effective lightweight model for cross-task and cross-embodiment transfer. Evaluation code and protocols are available at https://anonymous.4open.science/r/LITEC-WAM-A-LIGHTWEIGHTCROSS-TASK-AND-EMBODIMENT-WORLD-ACTION-MODEL-632F/

Primary area:
applications to robotics, autonomy, planning
Keywords:
World Action Models, Embodied AI, Robotic Manipulation, Lightweight, Cross-Task Transfer, Cross-Embodiment Transfer

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_WKj9OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes a highly lightweight WAM that facilitates deployment on edge devices and rapid transfer across tasks and embodiments. It achieves relatively competitive results with very few parameters. Simulation evaluations and real-world robot experiments demonstrate the method’s efficiency and ease of deployment.

Critical strengths
  1. The work achieves an effective balance between efficiency and capability, serving as a fast execution system that is easy to train and deploy. This aligns well with the role and requirements of fast systems in the agentic era.
  2. The paper presents extensive and comprehensive experiments that provide strong support for the method’s design rationale.
  3. The real-world robot experiments also demonstrate the method’s deployability, offering valuable insights for robotics research.
Critical weaknesses
  1. Some figures have relatively low resolution and are not in vector format; their quality should be improved.
  2. Additional results on other simulation benchmarks, such as RoboTwin, may be needed to validate transfer across different embodiments.
  3. The experimental presentation may benefit from reorganization. The tables should distinguish categories of models designed for different purposes, since their evaluation settings differ. The paper should also clearly specify its own evaluation setting and explain how its strengths and weaknesses compare with those of the various baselines across different dimensions.
Questions for authors
  1. Lightweight WAMs are similar to diffusion policies in many respects. Given this similarity, what are the key methodological differences between the proposed approach and diffusion policies, and where might its gains primarily come from?
  2. Could the latency comparisons across methods be more detailed? Comparisons of training costs would also be useful, as lower training costs may be an important advantage of lightweight WAMs.
3 comments
AuthorsOpenReview ↗

Response to Wkj9: Benchmark Coverage, Methodological Differences, and Efficiency

We thank the reviewer for the constructive suggestions on figure quality, cross-embodiment coverage, and comparison dimensions, and have revised the paper accordingly. All references below point to the revised paper and appendices.

(1) W1: Figure quality and vector graphics. Following the comment, we improved the figures as follows. We redrew the event curves and routing heatmap in Figure 10 (p. 28) as vector graphics, enlarged labels and legends, and separated the event-axis label from the wrist image. The comparison plot beside Table 1 (p. 7) now uses vector markers and offset/numbered labels to reduce overlap. Figure 2 (p. 4) clarifies the robot input/output interfaces and adaptation regimes. The duplicate comparison panel was removed from Figure 1 (p. 2).

(2) W2: Additional cross-embodiment evaluation. Following your suggestion, Table 31 (p. 42) adds a RoboTwin target-embodiment comparison on 10 held-out target tasks, with 50 initial states per task and three seeds (1,500 trials per method). LiteC-WAM achieves 90.20% mean success, with a seed-level 95% t confidence interval of [87.43%, 92.97%], versus 70.60% [69.61%, 71.59%] for Octo and 67.20% [66.21%, 68.19%] for Diffusion Policy.

This complements the physical SO-ARM101 evaluation, which reports 55/67 successful trials (82.09%; Table 15, p. 29). We distinguish these task-completion results from ARX5's offline action-agreement metric (Table 7, p. 17); the latter is not a closed-loop success rate.

(3) W3: Organization, evaluation scope, and comparison dimensions. Table 1 (p. 7) separates action-diffusion, generalist/vision-language-action, and video/world-model-based policies. Table 2 (p. 8) distinguishes spatial/keypose prediction from diffusion-based trajectory generation, noting that categories can overlap. Both tables identify published source protocols separately from our original custom evaluations.

Appendix N, Table 24 (p. 38) consolidates our task counts, initial states, model configurations, and metrics: five-task LIBERO success, 15-task RLBench development success, physical SO-ARM101 completion, and ARX5 offline agreement on the model-selection split.

For direct comparisons, Table 30 (p. 42) evaluates LiteC-WAM, Octo, and Diffusion Policy with matched tasks, demonstrations, observations, action interfaces, training budgets, and test states. On all 40 LIBERO tasks, their scores are 89.55%, 82.20%, and 78.50%, respectively. This comparison uses one seed per method; LiteC-WAM's separate three-seed full-suite result is 89.40% [85.85%, 92.95%] (Table 29, p. 41). We therefore limit superiority claims to the matched settings and avoid ranking methods across incompatible published protocols.

(4) Q1: Differences from diffusion policies and sources of gains. We agree that both approaches generate action chunks. LiteC-WAM uses flow matching with an eight-step Euler solver (Equation 2, p. 4), whereas Diffusion Policy iteratively denoises action chunks. The additional mechanism is an action-conditioned predictive branch: sparse dynamics experts predict future latent features and events, and the LIBERO feedback head converts these predictions and routing information into an action residual (Equations 3-6, p. 5; Appendix O, p. 39).

The evidence separates three effects. First, with weights, tasks, and initial states fixed, supplying joint predictive feedback raises success from 319/480 to 353/480 (+7.08 pp; Table 3, p. 9). Second, matched Top-2 routing reduces active parameters from 46.68M to 32.40M (30.6%) with 50/120 versus 49/120 successes (Table 13, p. 27). Third, adapter-based ARX5/RLBench adaptation retains 48/50 successes on the original LIBERO tasks (Table 11, p. 22).

These findings support feedback, sparse computation, and modular adaptation in their tested settings. They do not isolate a causal advantage of flow matching over diffusion, nor establish that prediction supervision alone improves control; Table 13 shows lower prediction error but no success improvement from adding predictive objectives in that comparison.

(5) Q2: Latency and training costs. Appendix P, Table 25 (p. 39) reports synchronized timing on one L40S with batch size 1, covering preprocessing through action output. For LiteC-WAM/Octo/Diffusion Policy, mean latency is 21/42/78 ms, P95 latency is 26/51/92 ms, and peak memory is 4.1/5.4/6.8 GB. These measurements cover policy calls, not the complete physical control cycle.

The same table separates shared training, adaptation, and overhead: LiteC-WAM uses 68+24+4=96 GPU-hours; Octo uses 184+24+8=216; Diffusion Policy uses 140+24+4=168. In these configurations, the reported LiteC-WAM total is 55.6% lower than Octo and 42.9% lower than Diffusion Policy. Table 9 (p. 21) separately documents stage-local training records.

We thank the reviewer again for helping improve the figure presentation, evaluation coverage, and efficiency reporting.

Reviewer_WKj9OpenReview ↗

Reply to the authors

Thanks for your reply, it addresses some of my questions. I still keep my rating and confidence.

Thank you for taking the time to review our response. We appreciate your feedback and are glad that it addressed some of your questions. Could you share which remaining concerns are most important to your assessment? This would help us focus our follow-up on the points that need further clarification. We would be happy to address them directly.

Reviewer_StkoOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

LiteC-WAM is a compact model for robotic control across tasks and embodiments. A CNN-GRU encoder provides context for flow-matching action generation and a dynamics predictor that selects two of eight experts. Predicted latent changes and events provide feedback to the policy. Observation/state adapters and robot-specific action codecs support adaptation. The paper reports a 47.57M shared backbone, 96.0% success on five selected LIBERO tasks, 86.67% on a 15-task RLBench development set, 82.09% closed-loop success on SO-ARM101, and 91.86% offline ARX5 action agreement.

Critical strengths
  1. The model combines action generation and prediction in a compact architecture. Table 12 compares Top-2 routing with a dense predictor and shows fewer active parameters without a substantial observed drop in success rate.
  2. The paired 20-task LIBERO experiment supports the use of feedback in closed-loop control: 353/480 successes versus 319/480 with the feedback input zeroed, a gain of 7.08 percentage points. This is useful evidence for the combined feedback signal.
  3. The paper includes physical closed-loop execution, with 55/67 successful SO-ARM101 trials. It reports which modules are trained during adaptation and reevaluates the original LIBERO policy after ARX5/RLBench adaptation, retaining 48/50 successes.
Critical weaknesses
  1. The configurations differ in core action generation and feedback, as well as in input/output interfaces. ARX5 uses regression-anchor weight 1.0 and bypasses the latent flow sampler. SO-ARM101 replaces additive residual feedback with future conditioning of the action head and updates shared dynamics. The paper gives some reasons for these choices, but does not report matched full-flow versus regression results for ARX5, or closed-loop comparisons of the alternative feedback and update choices for SO-ARM101. It remains unclear whether a common design would work as well in these settings and what each change contributes. A few targeted comparisons would help answer this question. Existing results may suffice; otherwise, the claims about the shared design should be more limited.
  2. The 96.0% LIBERO result covers five selected tasks and 50 rollouts, while the external baselines use published LIBERO aggregates. The RLBench result uses 15 development tasks and eight fixed initial states per task; the external results use an 18-task/249-variation protocol. The authors disclose these differences. However, the table cannot establish whether LiteC-WAM outperforms compact policies or is competitive with larger models. This requires representative compact policies evaluated on the same tasks, with the same demonstrations, initial states, and training budget. Table 12 supports the reduction in active parameters. A speed claim would also need latency measured on the same hardware. The lack of comparable baselines is my main concern.
  3. The SO-ARM101 evaluation says that the shared backbone is frozen (p. 8, lines 420-424). Appendix G.1 instead says that shared dynamics, including the router and experts, are trainable (p. 26, lines 1371-1373). The frozen-backbone claim also appears in the third contribution (p. 2, lines 93-97) and conclusion (p. 9, lines 479-483), whereas the method and Algorithm 1 describe updating shared dynamics. These accounts need to be reconciled. Capability retention is tested after ARX5/RLBench adaptation, but not after the SO-ARM101 refinement, so the retention claim needs the same qualification.
Questions for authors
  1. Which shared modules are updated for SO-ARM101? Please reconcile the frozen-backbone statement on p. 8 with Table 8, Algorithm 1, and Appendix G.1, and clarify whether capability retention was tested for this configuration.
  2. Why is the latent flow sampler bypassed for ARX5, and why is the feedback path changed for SO-ARM101? Are there existing comparisons that isolate the benefit of these choices, or were they engineering decisions without a controlled comparison?
  3. Are baseline results under the same protocol already available for the selected LIBERO/RLBench tasks? If not, which comparative claims can the current tables support?
  4. How does the prediction at horizon h incorporate preceding candidate actions in Eq. (3)? Does the action embedding encode the whole chunk, or is there an additional temporal aggregation path?
  5. The ARX5 validation split is also used for model selection, and the SO-ARM101 MAE diagnostics use splits within recordings. Which results evaluate independent episodes, and which measure prediction within the same recordings?
2 comments
AuthorsOpenReview ↗

Response to Stko: Matched Comparisons, Configuration Choices, and Retention

We thank the reviewer for the detailed comments on matched comparisons, configuration choices, and retention, and have revised the paper accordingly. All references below are to the revised paper and appendices.

(1) W2/Q3: Matched baselines and efficiency. We agree that the original five-task LIBERO and 15-task RLBench results cannot establish superiority over baselines evaluated under different protocols. Accordingly, Table 30 (p. 42) compares LiteC-WAM, Octo, and Diffusion Policy using the same tasks, demonstrations, observations, action interfaces, training budgets, and test states within each protocol. In that order:

LIBERO, all 40 tasks: 1791/2000 (89.55%), 1644/2000 (82.20%), 1570/2000 (78.50%).

LIBERO, five-task subset: 49/50, 43/50, 40/50.

RLBench, 15 tasks: 109/120, 92/120, 84/120.

These use one seed per method. The full-suite LiteC-WAM entry is seed 2 of Table 29 (p. 41), whose three-seed mean is 89.40% with a seed-level 95% t CI of [85.85%, 92.95%]. The matched runs are separate from the original 48/50 and 104/120 results in Tables 1-2 (pp. 7-8). We therefore limit comparative conclusions to these methods and protocols.

For speed, Appendix P, Table 25 (p. 39) reports synchronized measurements on one L40S at batch size 1, covering preprocessing through action output. Mean/P95 latency is 21/26 ms for LiteC-WAM, 42/51 ms for Octo, and 78/92 ms for Diffusion Policy. These are policy-call measurements, not complete robot-cycle latency. Separately, Table 13 (p. 27) isolates routing: Top-2 uses 32.40M active parameters versus 46.68M for dense routing (30.6% fewer), with 50/120 versus 49/120 successes.

(2) W1/Q2: Configuration-specific action generation and feedback. We agree that the configurations differ beyond input/output adapters. Section 3.3 (pp. 5-6), together with Section 3.2 (p. 5), distinguishes LIBERO's residual feedback, ARX5's regression branch at anchor weight 1.0, and SO-ARM101's future-conditioned action head.

For ARX5, Appendix I.1, Table 20 (p. 34) compares full-flow and regression branches on 2,100 paired windows with shared normalization and action codecs. Regression/full-flow joint MAE is 0.01732/0.01761 rad; gripper MAE is 0.435/0.443 mm; window pass rates are 98.57%/98.19%. The small advantage supports the regression choice for first-action agreement here; we present it as a configuration choice, not a general claim, and it does not establish a closed-loop advantage.

For SO-ARM101, Appendix I.2, Table 21 (p. 34) crosses feedback path with frozen versus updated dynamics, keeping the visual encoder frozen (180 outcomes per condition). With frozen dynamics, no feedback/residual feedback/future conditioning achieve 58.33%/65.00%/69.44%; with updated dynamics, 63.33%/71.67%/78.89%. Future conditioning improves over residual feedback by 4.44 pp with frozen dynamics and 7.22 pp with updated dynamics; updating dynamics adds 9.44 pp under future conditioning. The separate physical evaluation reports 55/67 successes (Table 15, p. 29).

(3) W3/Q1: Frozen modules and retention after SO-ARM101 refinement. We revised the previously blanket frozen-backbone description to distinguish these regimes. ARX5/RLBench adaptation freezes the shared backbone. SO-ARM101 refinement updates the router, dynamics experts, selected embodiment/feedback modules, and auxiliary decoder; the visual encoder and remaining base modules stay frozen (Table 8, p. 20; Algorithm 1, p. 25). Appendix M, Table 23 (p. 37) confirms that all six shared-dynamics tensors change, while 16 visual-encoder and 62 other base tensors remain unchanged.

Retention is evaluated separately. Table 11 (p. 22) reports 48/50 retained LIBERO successes after either ARX5 or RLBench adaptation. Appendix L, Table 22 (p. 36) evaluates SO-ARM101 retention on 20 Spatial/Object tasks, three seeds, and 50 states per task, with dynamics feedback enabled in all conditions. Parent/frozen conditions achieve 2220/3000 (74.00%); updated dynamics achieve 2208/3000 (73.60%). The change is -0.40 pp, with a seed-level 95% CI of [-0.65, -0.15] pp. We acknowledge this small decline and state it explicitly; the one-sided test supports non-inferiority within the stated 1 pp margin (p=0.00457), not exact preservation.

(4) Q4: Candidate-action history in Equation 3. Equation 3 (p. 5) uses the observed context, the candidate-action embedding at horizon h, and the horizon embedding. In LIBERO, this query does not explicitly aggregate preceding candidate actions. SO-ARM101 instead averages projected action embeddings from slots 1 through h before adding context and the horizon embedding (Appendix S, p. 41). This includes earlier actions but is permutation-invariant within the prefix, so it does not encode their order. Following the comment, Table 28 (p. 41) makes these dependencies explicit.

We thank the reviewer again for helping us clarify these comparisons and retention results.

AuthorsOpenReview ↗

Please reply before the discussion closes (24 hours remaining)

Dear Reviewer Stko,

The discussion closes at the end of 18 November AoE (about 24 hours from now). Our reply of 16 November answered your questions and concerns with new experiments and explanations.

Please consider updating the rating if our reply addresses your concern. If any concern remains, please let us know specifically which one and we will address it before the discussion closes, so that our paper is evaluated accurately and on the merits of its contributions. Thank you for your time.

Reviewer_T3aoOpenReview ↗
Rating
2
Confidence
2
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes LiteC-WAM, a lightweight world action model that combines a compact perception backbone, flow-matching action generation, sparse dynamics experts, and task- and embodiment-specific adapters. The method is evaluated on selected LIBERO and RLBench tasks, three physical bimanual tasks, and offline action evaluation on ARX5. The paper claims competitive performance with a small model size and effective transfer across tasks and embodiments.

Critical strengths
  1. The paper addresses a critical problem. Reducing model size while supporting adaption across tasks and embodiments is valuable for real-time deployment.

  2. The model achieves high reported success rates on the selected LIBERO and RLBench tasks with a relatively small number of parameters.

  3. The real-robot executions demonstrate the feasibility of deploying the model on physical settings.

Critical weaknesses
  1. The paper selects only five tasks from LIBERO for its main comparison, with only 10 rollouts per task, leading to 50 rollouts in total. The RLBench evaluation also includes only 120 rollouts in total. With such small evaluation sample sizes, the reported results may have high variance. These results do not sufficiently support the model’s cross-task transfer ability.

  2. Some design choices require further verification. In Table 12, removing the world-model training objectives increases prediction error but yields a slightly higher success count, leaving the control benefit of predictive training unclear. Table 3 combines dynamics features, routing probabilities, and event predictions without isolating their contributions. More targeted ablations are needed to justify the architecture.

  3. The presentation needs clarification. The cross-embodiment adapter part of Figure 2 lacks explanation. In particular, what does "Expandable capability with bounded active compute" mean? The right panel of Figure 1 and the right panel of Table 1 repeat the same plot. Figure 10 contains text overlapping a subfigure. Section 4.3 describes the physical evaluation as keeping the shared backbone frozen, whereas Section 3.3 describes updating the dynamics experts within the shared backbone, which is quite confusing.

Overall, the submission requires substantial additional experimental validation and a clearer methodological description to support its central claims.

Questions for authors
  1. Could you evaluate the model under standard LIBERO and RLBench benchmark protocols matched to those used by the reported baselines? For example, could you evaluate all four LIBERO suites with 50 rollouts per task?

  2. Could you provide more details about training the shared backbone, including the training datasets, demonstration counts, optimizer, learning rates, and training budget? Could you also clarify which shared components were updated during adaptation for the reported real-robot experiments?

  3. Could you conduct more targeted ablations to identify the contributions of the feedback components? In particular, how much of the improvement comes from latent dynamics features, routing probabilities, and predicted events?

8 comments
AuthorsOpenReview ↗

Response to T3ao: Evaluation, Ablations, and Training Clarifications

We thank the reviewer for the careful comments on evaluation coverage, ablation isolation, and training clarity, and have revised accordingly. All references below point to the revised paper and appendices.

(1) W1/Q1: Evaluation coverage, variance, and matched protocols. We expanded evaluation and separated training-task test states from unseen-task transfer. Appendix T, Table 29 (p. 41) covers all four LIBERO suites: 40 tasks, 50 test initial states/task, three seeds (6,000 rollouts). Mean success is 89.40%, seed-level 95% t-CI [85.85%, 92.95%]; suite means are 92.60%/96.40%/90.20%/78.40% for Spatial/Object/Goal/Long. The 32-source/8-target split uses 50 states/target task and source-validation checkpoint selection, yielding 76.00% [71.03%, 80.97%] across three seeds.

RLBench now includes an 18-task RVT-style protocol (25 episodes/task/seed; 1,350 total), achieving 86.00% [80.73%, 91.27%], and a 25-task protocol (20 episodes/task/seed; 1,500 total), achieving 86.40% [82.16%, 90.64%].

Table 30 (p. 42) compares LiteC-WAM, Octo, and Diffusion Policy with matched tasks, demonstrations, observations, action interfaces, training budgets, and test states. Full-suite LIBERO scores are 1791/2000 (89.55%), 1644/2000 (82.20%), and 1570/2000 (78.50%). These use one seed per method; LiteC-WAM's entry is seed 2 of Table 29, distinct from its three-seed mean. Published references in Tables 1-2 remain separate.

(2) W2: Predictive training versus control benefit. We agree that lower prediction error alone does not establish better control, and revised the text accordingly. Table 13 (p. 27) shows future-frame L1 decreasing from 0.3798 to 0.0485, while successes change from 54/120 to 50/120. We therefore do not use this to claim control improvement from predictive supervision alone, and instead report the separate paired feedback experiment.

That experiment holds weights, tasks, and initial states fixed and changes whether the feedback vector is supplied. Success increases from 319/480 to 353/480 (+7.08 pp; Table 3, p. 9; Appendix E, Table 14, p. 27). Five of six seeds improve. The episode-pair McNemar test gives p=0.000588; the two-sided seed sign test gives p=0.219. We distinguish these evidence levels.

(3) Q3: Individual feedback components. Appendix H, Tables 18-19 (p. 33) isolates inputs during feedback-head training/evaluation, holding base policy, demonstrations, budget, and evaluation states fixed (20 tasks, six seeds, four states/task). Relative to the full condition (354/480, 73.75%), removing dynamics features gives 334/480 (69.58%, -4.17 pp), removing events gives 347/480 (72.29%, -1.46 pp), and removing routing probabilities gives 351/480 (73.13%, -0.63 pp). The routing-input mask keeps Top-2 expert selection active. Removing prediction supervision with feedback retained gives 326/480 (67.92%, -5.83 pp). Table 19 reports seed-level intervals and tests; not all contrasts reach p<0.05. We report conditional leave-one-out effects, not additive shares; the full-vs-removal direction is consistent while individual contrasts vary. This retraining study is separate from Table 3.

(4) W3: Adapter meaning, computation, and presentation. Figure 2 (p. 4) and Section 3.3 (pp. 5-6) now explain observation adapters and robot-specific action modules. Appendix Q, Table 26 (p. 40) scopes compute: a rank-r branch has r(d+D) matrix parameters/MACs per vector; N equally shaped branches require N times storage, while selecting one uses one branch. This excludes other computation and does not imply constant end-to-end latency. Top-2 applies per query: eight queries make 16 expert calls and may touch all eight experts. We removed the repeated comparison panel from Figure 1 (p. 2), retained it beside Table 1 (p. 7), and separated the event-axis label from the wrist image in Figure 10 (p. 28).

(5) Q2/W3: Training recipe and frozen modules. Sections C.1-C.2 and Tables 8-9 (pp. 19-21) specify the controlled LIBERO run: 20 Spatial/Object tasks; 800/100/100 train/validation/test demonstrations; 12,500 updates, batch 8; AdamW, LR 1e-4, weight decay 0.01, 500-step warmup, cosine decay to 1e-5. This is the 46.86M controlled run, not the entire training history of the final 47.57M backbone; Section C.2 documents its merge and adaptation stages separately.

ARX5/RLBench adaptation freezes the shared backbone. Offline SO-ARM101 refinement updates the router, dynamics experts, embodiment/feedback modules, and auxiliary decoder, while freezing the visual encoder and remaining base modules (Table 8). It runs 6,000 updates at batch 12; group-specific learning rates and checkpoint selection are detailed on p. 21. The revised text distinguishes these regimes explicitly.

We thank the reviewer again for helping us improve the paper.

If possible, would u mind provide tables for the experimental comparisons. That would be much clear.

Thank you for the suggestion. We summarize the comparisons below in tables for easier inspection. Page numbers refer to the revised manuscript.

1. Extended benchmark summaries (Table 29, p. 41).

ProtocolTasksRollouts/seedSeedsMean success (%)95% seed-level t CI (%)
LIBERO full suite402,000389.40[85.85, 92.95]
LIBERO held-out8 target tasks400376.00[71.03, 80.97]
RLBench RVT-style18450386.00[80.73, 91.27]
RLBench extended25500386.40[82.16, 90.64]

The full-suite LIBERO protocol uses 50 test states per task. The held-out design trains on 32 source tasks and selects checkpoints using source validation.

LIBERO suite breakdown (Table 29, p. 41).

SuiteSeed 1Seed 2Seed 3Mean success (%)
Spatial455/500464/500470/50092.60
Object475/500483/500488/50096.40
Goal444/500451/500458/50090.20
Long384/500393/500399/50078.40
All four suites1758/20001791/20001815/200089.40

2. Common-protocol comparison (Table 30, p. 42).

MethodLIBERO 40 tasksLIBERO 5 tasksRLBench 15 tasks
LiteC-WAM1791/2000 (89.55%)49/50 (98.00%)109/120 (90.83%)
Octo1644/2000 (82.20%)43/50 (86.00%)92/120 (76.67%)
Diffusion Policy1570/2000 (78.50%)40/50 (80.00%)84/120 (70.00%)

The specified protocols match tasks, demonstrations, observations, action interfaces, training budgets, and test states within each column. Each entry represents one seed; the full-suite LiteC-WAM entry reuses Seed 2 above.

3. Routing and predictive supervision — measured comparison (Table 13, p. 27).

Dynamics variantActive parametersFuture-frame L1Successes/trials
Dense46.68M0.047949/120
Sparse Top-129.99M0.048348/120
Sparse Top-232.40M0.048550/120
Top-2, prediction objectives disabled32.40M0.379854/120

Top-2 reduces active parameters by 30.6% relative to dense routing. Prediction supervision improves prediction accuracy; this comparison does not demonstrate improved control success from supervision alone.

4. Joint feedback — measured paired comparison (Table 3, p. 9; Table 14, p. 27).

Feedback inputSuccesses/trialsSuccess (%)95% Wilson CI (%)
Zeroed319/48066.46[62.12, 70.54]
Supplied353/48073.54[69.42, 77.29]

The comparison fixes weights, tasks, and initial states across 20 tasks, six seeds, and four states per task. The aggregate gain is 7.08 percentage points. Episode-pair McNemar p=0.000588; two-sided seed sign-test p=0.219.

5. Feedback-component comparison (Tables 18–19, p. 33).

ConditionSuccesses/trialsSuccess (%)Change vs. full (pp)
Full354/48073.75—
Without dynamics features334/48069.58-4.17
Without predicted events347/48072.29-1.46
Without routing probabilities351/48073.13-0.63
Without prediction supervision326/48067.92-5.83

This design varies one feedback input or the prediction-supervision setting at a time, keeping the base policy, demonstrations, training budget, and evaluation states fixed. Masking routing probabilities leaves Top-2 expert selection active. Its full condition is separate from the measured paired study above.

Training settings and trainable/frozen modules are tabulated in Table 8 (p. 20), with stage-specific training records in Table 9 (p. 21).

Thank you again for suggesting this clearer presentation.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Author Follow-up on Reviewer Response

Dear AC, We have submitted our response to all reviewers. We wanted to kindly flag that we have not yet seen a follow-up from Reviewer [Wkj9/T3ao/Stko]. We completely understand that reviewers are busy, and we are only mentioning this in case the missing follow-up affects the discussion. Our rebuttal for this reviewer is complete and addresses all points raised. Thank you very much for your time and for handling our paper. Best regards, The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.