acceptodds
Under review as a conference paper at ICLR 2027

GenPA: Image Generators as Efficient Robot Policies

Abstract

Pretrained visual generation models have shown great promise in robotic manipulation. However, a representative line of work still treats future visual prediction as the central link between visual generative pretraining and action learning, retaining it as a training objective even when future scenes are no longer generated at inference time. This paradigm organizes policy learning around the proxy task of predicting future visual states. We argue that the instruction-following, spatial understanding, and visual imagination capabilities acquired through image generation pretraining can support control without future visual prediction: image generation models themselves can be directly adapted into powerful action learners. Meanwhile, their native image-editing interface can accommodate diverse dense perception targets, providing additional visual supervision for action learning. Building on this perspective, we propose GenPA, which directly learns actions by fine-tuning an image generation backbone together with a lightweight 19M-parameter regression adapter, while incorporating native perception co-training to strengthen control-relevant visual representations. At deployment, GenPA requires only a single backbone forward pass and action regression, without visual generation or iterative action denoising. Without additional large-scale robot-data pretraining, GenPA demonstrates strong control performance and robustness to perturbations across simulation benchmarks (99.10% success on LIBERO and 83.79% on LIBERO-Plus) and real-world manipulation tasks. Under compiled inference, GenPA takes 20.9 ms per action chunk on a single H100 GPU, achieving 1.7–3.4× speedups over strong VLA and WAM baselines. Our latency analysis further highlights architectural advantages of visual generation backbones for efficient robot control. Ablations across visual tasks and their combinations identify current-depth supervision as a simple, effective training recipe. Together, these results show that image generators can be directly adapted into strong, efficient robot policies without requiring future visual prediction.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robot learning, robotic manipulation, image generation models, perception-action co-training, efficient inference

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_ykMfOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

GenPA fine-tunes a pretrained image-generation backbone for robot control using a lightweight action regression adapter. Its auxiliary visual objective predicts current-scene perceptual targets, such as depth, rather than future observations. Deployment requires one backbone forward pass without image generation or iterative action denoising. Experiments on LIBERO, LIBERO-Plus, RoboTwin 2.0, and a real-world USB manipulation task demonstrate competitive performance and low inference latency. Ablations examine action decoding and alternative visual supervision targets.

Critical strengths
  1. GenPA achieves competitive control performance with a lightweight regression adapter and 20.9 ms compiled inference on H100. The internal regression–flow comparison supports the practical value of simplifying action generation.
  2. Within a common policy architecture, current-depth supervision improves LIBERO-Plus robustness over action-only training and performs competitively with future-RGB supervision.
Critical weaknesses

Fast-WAM already removes future generation at inference, and ImageWAM already adapts image-editing backbones for control. GenPA’s additional contribution centers on current-perception co-training and lightweight action regression. Its significance therefore rests primarily on the empirical effectiveness of this combination.

Questions for authors
  1. Can you provide repeated-seed results for the regression–flow and None/depth/future-RGB comparisons, with evaluation separated from checkpoint selection?
  2. Did the real-world baselines receive the same image augmentations as GenPA?
4 comments
AuthorsOpenReview ↗

Response to Reviewer ykMf

We thank the reviewer for the careful assessment and constructive suggestions. Section and table references below refer to the revised manuscript.

W1. Relation to Fast-WAM and ImageWAM

We agree that Fast-WAM and ImageWAM already enable inference without generating future images; their main formulations retain future-prediction co-training. Our contribution concerns how the generator is trained for control: GenPA combines direct action learning with native current-scene perception co-training. Within this framework, we compare current targets, future targets, and their combinations under a common policy architecture.

These comparisons show that current-scene supervision can improve control robustness without future prediction, while combining individually useful targets does not consistently improve performance (Section 3.6). Repeated training and fixed-final-checkpoint comparisons support the robustness gains, complemented by independent standard-LIBERO evaluations. The action-head ranking also varies across evaluation settings (Q1). This establishes current perception as an effective training alternative, with evidence on its benefits and limits.

We also characterize efficiency beyond the action head: component-level measurements show how recurring observation-processing costs and instruction-encoding reuse affect latency alongside action decoding (Section 3.4). This identifies the computational factors underlying efficient control with image-generation backbones.

Q1. Repeated training and independent evaluation

We now report results from three training seeds for each of the four configurations, complete ten-checkpoint LIBERO scans, results for both selected and fixed-final checkpoints, and separate fixed-policy evaluation on three new initial-state sets disjoint from selection.

Depth and future-RGB supervision improve mean Plus success over action-only training in all four Plus settings. With current-depth supervision, regression has higher standard LIBERO success, but flow leads on Plus under Official instructions and at the endpoint under Clean instructions.

Repeated training. All configurations use action regression except “Depth + flow.” Success rates (%) are mean ± sample SD across the three training runs.

Selected is the checkpoint with the highest standard LIBERO success; Final is the fixed training endpoint. Plus never enters selection. Clean denotes the original paper’s main Plus setting, which removes appended perturbation metadata from instructions for non-language perturbations; Official uses released instructions (Appendices C.1–C.2).

EvaluationAction-onlyCurrent depthFuture RGBDepth + flow
Selected LIBERO98.80 ± 0.2298.87 ± 0.2598.68 ± 0.1098.27 ± 0.12
Final LIBERO98.28 ± 0.1698.53 ± 0.0898.18 ± 0.2598.00 ± 0.40
Selected Plus, Official74.63 ± 4.0378.67 ± 3.9777.74 ± 6.5779.47 ± 2.42
Final Plus, Official74.32 ± 3.4879.27 ± 3.6278.64 ± 5.8279.98 ± 2.60
Selected Plus, Clean79.49 ± 0.8582.33 ± 1.2781.80 ± 2.0581.49 ± 1.18
Final Plus, Clean79.45 ± 0.3982.35 ± 0.8582.15 ± 2.1582.46 ± 0.43

The LIBERO rows reuse selection states; independent evaluation is reported below.

Independent-state evaluation. We fixed the selected checkpoints from training seed 42 and tested them on three shared sets of new states disjoint from selection, totaling 6,000 episodes per policy. Mean ± sample SD across evaluation sets is 97.63 ± 0.24% (action-only), 98.33 ± 0.29% (current depth), 97.80 ± 0.05% (future RGB), and 97.03 ± 0.33% (depth + flow). These results measure evaluation variability for fixed policies; we did not test all three training runs on the new states.

Complete per-seed ten-checkpoint scans are in Appendix C.3, Tables 24–25. The selected/final results across training seeds are in Appendix B.5/Table 17 and Appendix D.1/Table 28; independent-state results are in Appendix C.3, Table 26.

Q2. Image augmentation in real-world baselines

No. Each method used its own image-augmentation recipe; Fast-WAM used deterministic preprocessing without random augmentation. Appendix E.1 now details these differences. The real-world comparison therefore reflects both policy design and augmentation, without isolating their individual effects.

AuthorsOpenReview ↗

Follow-up on our response

Thank you again for your constructive feedback. We would like to briefly follow up on our November 11 response, particularly the additions addressing your two questions:

  1. Three-seed training for the regression–flow and action-only/depth/future-RGB comparisons, with complete checkpoint scans and both selected and fixed-final results.
  2. Independent evaluation of the selected seed-42 policies on three new initial-state sets disjoint from checkpoint selection.
  3. Explicit clarification that the real-world baselines used different augmentation recipes, with these differences documented and the comparison’s interpretation qualified.

We would appreciate your feedback on whether these additions address your concerns, and are happy to clarify any remaining questions.

Thanks for the detailed response. I maintain my rating.

AuthorsOpenReview ↗

Thank you for your feedback

Thank you for reviewing our responses and additional experiments. We appreciate your constructive feedback and continued support.

Reviewer_irNpOpenReview ↗
Rating
4(+1)
Confidence
2
Informativeness
3
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes GenPA, which directly fine-tunes an image-generation backbone for robotic action prediction without requiring future visual prediction. A regression adapter reads instruction-conditioned observation features from FLUX.2 klein Base 4B, while a separate training pass reuses the generator's native image-editing objective for auxiliary perception supervision, primarily current depth. Inference uses one backbone pass and direct action regression. Experiments on LIBERO, LIBERO-Plus, RoboTwin 2.0, and real-world USB manipulation show a favorable accuracy–latency trade-off, with ablations of action readout and visual supervision and detailed inference profiling.

Critical strengths
  1. Simple and effective adaptation of generative pretraining for control: The method removes both future-image generation and iterative action denoising from deployment while retaining perception supervision during training. The separation between action and perception passes is clearly specified, and the action path does not consume depth annotations or future observations at inference (Section 2).

  2. Strong empirical performance with meaningful robustness evaluation: GenPA reports 99.10% success on LIBERO, 83.79% on LIBERO-Plus, and 91.45% average success on RoboTwin 2.0. The LIBERO-Plus comparison re-evaluates released baselines on common instances and processed instructions rather than only collecting published scores. Real-world USB success averages 78%, compared with 70% for ImageWAM, using the same 600 demonstration trajectories (Tables 1–3 and 5).

  3. Useful controlled ablations and unusually detailed latency reporting: The paper compares regression and flow-based action adapters within a common backbone and evaluates multiple perception targets. Appendix B.5 confirms the depth benefit under the main augmented-image configuration, increasing LIBERO-Plus success from 79.07% to 83.79%. Appendix D reports H100/L40S measurements, instruction-encoding overhead, and closed-loop performance after compilation; the approximately 20.9 ms compiled H100 latency is therefore supported by more than a single isolated timing result.

Critical weaknesses
  1. Standard LIBERO results reuse the checkpoint-selection evaluations as the reported test scores: Appendix C.1 explicitly selects the best of ten checkpoints on the same 2,000 LIBERO trials and reports those selection scores. This introduces selection optimism and makes small improvements, such as 99.10% versus 98.40% for the published ImageWAM result, difficult to assess under potentially different selection protocols. All GenPA training configurations also use seed 42, so variation across independently trained policies is not quantified. LIBERO-Plus is not directly used for checkpoint selection and remains useful supporting evidence, but it does not resolve the statistical interpretation of the standard LIBERO comparisons.

  2. The distinctive benefit of native generative perception supervision is not isolated: Parallel action regression is already an established efficient adaptation strategy [1]. The interesting contribution is therefore its combination with an adapted image generator and native perception co-training. The existing regression–flow and current–future target ablations are valuable, but do not compare the native flow-based perception loss against a lightweight direct depth-prediction head trained on the same backbone and labels. Thus, the results establish that the proposed recipe works, but not whether reusing the generative interface provides a benefit beyond ordinary auxiliary geometric supervision. This distinction matters when assessing the specific methodological contribution.

  3. The resource-efficiency evidence is strongest for warm online inference, not the full policy footprint: The timing protocol is well documented, including uncached instruction encoding, but peak memory and training GPU-hours are not reported. The 19M figure describes the LIBERO action adapter, while the method fine-tunes a 4B backbone, uses a separate frozen Qwen3-4B text encoder, and adds a perception pass during training; RoboTwin uses a 71.45M adapter (Appendices A.1–A.5). These are disclosed architectural choices, not flaws in the latency measurement, but a complete resource comparison is needed to assess deployment requirements and the training-cost trade-off behind the faster policy calls.

References

[1] Moo Jin Kim, Chelsea Finn, and Percy Liang. Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. Robotics: Science and Systems (RSS), 2025. DOI: 10.15607/RSS.2025.XXI.017.

Questions for authors
  1. Can the main LIBERO comparison be repeated on evaluation seeds or initial states disjoint from checkpoint selection, with the checkpoint fixed beforehand? A limited repeated-training check for the main configuration would also help determine whether the small accuracy margins are stable.

  2. On a representative LIBERO subset, how does native flow-based depth supervision compare with direct depth regression from the same backbone features, using the same labels and matched action-update budgets? Please report both closed-loop performance and the additional training cost.

  3. What are the peak training/inference memory and total training GPU-hours for action-only versus depth-co-trained GenPA? Please distinguish the resident policy footprint with cached instructions from the resources needed when the text encoder is also loaded.

4 comments
AuthorsOpenReview ↗

Response to Reviewer irNp (Part 1/2: W1, W2, W3, Q1)

We thank the reviewer for the careful assessment and constructive suggestions. Section and table references below refer to the revised manuscript.

W1. Checkpoint selection and training variability

We now label the original 99.10% LIBERO result as a checkpoint-selection score. We added four complementary checks: fixed-checkpoint evaluation on three new initial-state sets disjoint from selection; three training seeds for the main configurations; complete ten-checkpoint LIBERO scans; and results for both selected and fixed-final checkpoints. These address evaluation independence, training variability, and sensitivity to checkpoint selection. Q1 below summarizes the evidence.

W2. Native supervision versus direct depth regression

We agree that parallel action regression is established in OpenVLA-OFT [1]; here we examine the additional value of native current-scene perception co-training for an image-generation backbone.

We added the requested DPT-style depth-regression control with the same backbone, depth labels, and action-update budget. Native supervision achieves higher mean Plus success, with similar standard LIBERO performance and greater training cost (Q2).

W3. Full policy resource footprint

The 19M count refers only to the LIBERO action adapter. We now report the full policy footprint, training and inference memory, and estimated training GPU-hours in Q3. Depth co-training increases training cost but adds no inference-time computation.

Q1. Independent LIBERO evaluation

We fixed the checkpoints before evaluating all six policies on three shared sets of new initial states, disjoint from the selection states and one another, totaling 6,000 episodes per policy.

Fixed policyIndependent success (%)
GenPA: native current-depth supervision, regression actions98.33 ± 0.29
GenPA: action-only training97.63 ± 0.24
GenPA: future-RGB supervision97.80 ± 0.05
GenPA: native current-depth supervision, flow actions97.03 ± 0.33
GenPA: direct depth regression98.12 ± 0.33
ImageWAM: released FLUX.2-4B checkpoint97.05 ± 0.25

Values are mean ± sample SD across the three evaluation sets; the GenPA policies use training seed 42 and regression actions unless stated otherwise. The five GenPA results appear in Appendix C.3, Table 26. ImageWAM uses its released Prefix-Only mode, which predicts actions without image denoising; this differs from the inference protocol behind its published 98.40%.

We also repeated training for action-only, native-depth, future-RGB, and flow-action variants with three seeds. Selected denotes the best standard-LIBERO checkpoint; Final denotes the fixed training endpoint. Native-depth GenPA achieves 98.87 ± 0.25% (Selected) and 98.53 ± 0.08% (Final), versus 98.80 ± 0.22% and 98.28 ± 0.16% for action-only training. Here SD is across training runs, and these scores reuse the selection states. Both selected and fixed-final results are reported for standard LIBERO and Plus (Appendix B.5/Table 17 and Appendix D.1/Table 28); complete per-seed ten-checkpoint scans are in Appendix C.3, Tables 24–25.

The independent-state test and training repeats address different sources of variability: we did not evaluate every training run on the new states or retrain ImageWAM. They therefore do not establish that the original GenPA–ImageWAM margin is stable across independently trained methods.

AuthorsOpenReview ↗

Response to Reviewer irNp (Part 2/2: Q2, Q3)

This continues our response to Reviewer irNp with Q2 and Q3. Evaluation settings follow Part 1; section and table references refer to the revised manuscript.

Q2. Depth supervision: performance and training cost

We compared a DPT-style depth-regression head with depth supervision through the generator’s native interface across all four LIBERO suites and three training seeds, using the same backbone, training data, depth labels, augmentation, and action-update budget. Native supervision achieves similar standard LIBERO performance and higher mean Plus success, at approximately 2.50× the training time per update.

Success rates (%) are mean ± sample SD across training runs. Selected and Final follow the definitions in Q1; Plus never enters selection. Clean denotes the original paper’s main Plus setting, which removes appended perturbation metadata from instructions for non-language perturbations; Official uses released instructions (Appendices C.1–C.2).

EvaluationNative depthDirect depth regression
Selected, standard LIBERO98.87 ± 0.2598.85 ± 0.10
Final, standard LIBERO98.53 ± 0.0898.67 ± 0.24
Selected, Plus Clean82.33 ± 1.2780.61 ± 0.35
Final, Plus Clean82.35 ± 0.8580.27 ± 0.59
Selected, Plus Official78.67 ± 3.9776.72 ± 2.95
Final, Plus Official79.27 ± 3.6276.39 ± 1.96

Native depth supervision and direct depth regression take 1.294 and 0.517 seconds/update, respectively, on the same eight H100 GPUs. The comparison matches action updates, not compute, and compares supervision routes rather than the loss function alone. Results are in Section 3.5 and Appendix B.6/Table 18; implementation details are in Appendix A.6.

Q3. Memory usage and training GPU-hours

The table reports measured update times and memory on eight H100 80GB GPUs, together with estimated GPU-hours for one full training run. Memory is peak PyTorch allocation per GPU during the training loop.

Training variantSeconds/updatePeak allocated (GiB/GPU)Estimated GPU-hours
Action-only0.50135.4340.9
Native depth1.29454.11105.7

The GPU-hours estimates extrapolate measured update times over the full training budget, excluding initialization and checkpoint saving. Depth co-training requires approximately 2.58× the training compute (Appendix A.5, Table 13).

For batch-one compiled inference on a single H100, peak PyTorch allocated memory is 7.63 GiB with cached instructions and the text encoder unloaded, or 15.18 GiB with Qwen3-4B resident, including new-instruction encoding. The corresponding model sizes are 3.98B and 8.00B parameters. Action-only and depth-co-trained GenPA share this inference footprint.

These inference peaks cover calls after warm-up, excluding loading, compilation, and driver allocations (Appendix D.6).

Thank you for the thorough responses and the substantial additional experiments. The new evaluations, multi-seed training, and the direct depth-regression control effectively address my concerns. I am revising my rating to 4.

AuthorsOpenReview ↗

Thank you for your feedback

Thank you for your careful review and constructive feedback. We appreciate your time in considering our responses and additional experiments.

Reviewer_c1uPOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents GenPA, which adapts a pretrained FLUX.2 image-generation backbone into a robot policy. A lightweight regression adapter reads instruction-conditioned observation features and predicts an action chunk in parallel. During training, the shared backbone additionally learns image-form perception targets, primarily current-scene depth, through its native image-editing interface and flow objective. Deployment requires one backbone forward pass and direct action regression, without visual generation or iterative action denoising. The contributions are this simplified policy formulation and an empirical study of current-scene supervision as an alternative to future-image prediction. Validation covers LIBERO, LIBERO-Plus, RoboTwin 2.0, one real-world USB manipulation task, supervision and action-head ablations, and latency profiling.

Critical strengths

1.A useful simplification of image-generator-based policies. GenPA combines direct action regression with native current-scene perception co-training, showing that future-image prediction is not required for a strong policy in the evaluated setting. The perception objective reuses the generator’s existing interface, and its branch adds no inference-time work. This offers a practical route to using a pretrained image-generation backbone while simplifying action decoding.

2.Substantial empirical support for the practical recipe. Evaluation spans two simulation benchmarks, a perturbation benchmark, and one contact-rich real-world task. The within-framework comparisons are particularly useful: regression reduces compiled H100 latency from 29.22 to 20.94 ms per action chunk while obtaining higher reported success than the flow adapter (Table 4); current-depth supervision raises LIBERO-Plus success from 79.07% to 83.79% under the main configuration (Table 15). Detailed timing and configuration reporting strengthen the practical value of these findings.

Critical weaknesses

1.The methodological novelty is modest relative to close prior work. ImageWAM already uses an image-editing backbone for action prediction without decoding images at deployment, while parallel continuous-action regression and action chunking are established in OpenVLA-OFT. Current-depth auxiliary supervision also has prior examples, as acknowledged in Section 4. GenPA’s main distinction is therefore the particular combination of a lightweight regression adapter and native current-scene perception co-training. This is useful, but limits the conceptual advance; the case for acceptance rests primarily on the empirical value of the resulting recipe rather than a fundamentally new policy architecture.

2.The source of the gains from native perception co-training is not sufficiently isolated. Tables 6 and 15 support current-depth co-training as an effective recipe, but the supervised variants compare target types within the same generative flow formulation. They do not include a direct depth-regression auxiliary task using the same depth labels. It therefore remains unclear how much of the improvement comes from additional geometric supervision versus the choice to reuse the generator’s native perception objective. Section 3.5 appropriately presents the geometric explanation as a hypothesis, but does not directly examine whether the observation features used by the action adapter become more informative about geometry. This limits the mechanistic insight and the evidence for the specific supervision design, without negating the observed control gains.

Questions for authors

1.Do you have controlled evidence separating the contribution of depth supervision from that of the native flow objective, for example through a direct depth-regression auxiliary head using the same backbone and labels under a comparable training budget?

2.Clarification of checkpoint sensitivity: Could you provide the already-computed standard LIBERO scores across the ten candidate checkpoints, particularly for regression/flow and None/D/F? Do the main comparisons remain consistent across late-training checkpoints, or do they depend strongly on the selected optimum? This would help assess the influence of the selection protocol without requiring new training runs.

3.Additional analysis of existing evaluation results: Could you use the existing LIBERO-Plus per-instance outcomes to quantify uncertainty in the regression–flow and depth–future-RGB differences, accounting for correlations among instances of the same task? Such analysis would help distinguish evidence for an effective alternative from evidence for superiority. It would characterize evaluation uncertainty for the selected checkpoints, but would not replace an estimate of training-seed variability.

5 comments
AuthorsOpenReview ↗

Response to Reviewer c1uP (Part 1/2: W1, W2, Q1)

We thank the reviewer for the careful assessment and constructive suggestions. Section and table references below refer to the revised manuscript.

W1. Contribution relative to prior work

We agree that the individual components have precedents. Our contribution addresses a question they leave unresolved: do the control benefits of visual co-training require future prediction, or can current-scene supervision provide an effective alternative?

ImageWAM trains with future target images. Fast-WAM removes future generation at inference and also studies removing video co-training. However, removing visual supervision does not test whether current-scene supervision can replace future prediction. OpenVLA-OFT establishes efficient action regression without addressing this supervision choice for image-generation backbones.

GenPA directly learns actions while retaining the generator’s native visual objective for current-scene perception. This enables the missing replacement comparison: current targets, future targets, and their combinations within the same policy architecture, using common training data and action-update budgets. Comparing current depth with future depth additionally holds the target representation constant while changing its temporal position.

The results show that current-scene supervision can improve control robustness without future prediction, supported by repeated training and fixed-final-checkpoint evaluations. They also show that combining individually useful targets does not consistently improve control. The added depth-regression control characterizes the performance–cost trade-off of the native interface (Q1). Together, these findings establish a current-perception route for adapting image generators to control and provide controlled evidence on how to choose its visual supervision.

W2. Sources of improvement and the geometric explanation

The added DPT-style depth-regression control uses the same depth labels and finds similar standard LIBERO performance but higher mean Plus success with native supervision, at greater training cost (Q1). This compares the supervision interfaces as a whole, rather than isolating the loss function. The geometric explanation remains a hypothesis because we have not directly probed geometric information in the action features (Section 3.6).

Q1. Direct depth-regression control

We compared a DPT-style depth-regression head with depth supervision through the generator’s native interface, using the same pretrained backbone, training data, depth labels, augmentation, and action-update budget across three training seeds. Native supervision achieves similar standard LIBERO performance and higher mean Plus success, at approximately 2.50× the training time per update on the same hardware.

Success rates (%) are mean ± sample SD across training runs. Selected is the checkpoint with the highest standard LIBERO success; Final is the fixed training endpoint. Plus never enters selection. Clean denotes the original paper’s main Plus setting, which removes appended perturbation metadata from instructions for non-language perturbations; Official uses released instructions (Appendices C.1–C.2).

EvaluationNative depthDirect depth regression
Selected, LIBERO98.87 ± 0.2598.85 ± 0.10
Final, LIBERO98.53 ± 0.0898.67 ± 0.24
Selected, Plus Official78.67 ± 3.9776.72 ± 2.95
Final, Plus Official79.27 ± 3.6276.39 ± 1.96
Selected, Plus Clean82.33 ± 1.2780.61 ± 0.35
Final, Plus Clean82.35 ± 0.8580.27 ± 0.59

The budgets match action updates, not compute. Results and implementation details are in Section 3.5/Table 6 and Appendices A.6 and B.6/Table 18.

AuthorsOpenReview ↗

Response to Reviewer c1uP (Part 2/2: Q2, Q3)

This continues our response to Reviewer c1uP with Q2 and Q3. Evaluation settings follow Part 1; section and table references refer to the revised manuscript.

Q2. Checkpoint sensitivity

With depth supervision, regression exceeds flow on standard LIBERO at all ten checkpoints in the three-run mean, while the ordering of action-only, depth, and future-RGB supervision varies.

We now report three training seeds for the main comparisons, complete ten-checkpoint LIBERO scans, and both selected and fixed-final results. We also added fixed-checkpoint evaluation on three new initial-state sets disjoint from selection. Below are mean success rates (%) across training runs for variants using the same backbone and action-update budget; per-seed scans are in Appendix C.3, Tables 24–25.

UpdatesNo visual loss / regression actionsDepth / regression actionsFuture RGB / regression actionsDepth / flow actions
20,00097.8898.0798.0290.50
22,00097.4097.6797.6793.60
24,00097.7398.6097.8794.33
26,00098.2298.5098.2895.63
28,00098.6098.5598.4597.22
30,00098.0898.4898.0397.70
32,00098.6098.5298.5798.07
34,00098.6898.5298.2397.92
36,00098.5598.5298.5397.80
36,75098.2898.5398.1898.00

The scans and selected/final standard-LIBERO scores use the original selection states; the fixed endpoint is not an independent test.

On Plus, both depth and future-RGB supervision improve mean success over action-only training at the fixed endpoint. The action-head ranking is less stable: regression leads under Clean instructions at selected checkpoints, but flow leads at the endpoint and under Official instructions. Both selected and fixed-final results are reported in Appendix B.5/Table 17 and Appendix D.1/Table 28.

Separately, fixed checkpoints from training seed 42 achieve 97.63 ± 0.24% (action-only), 98.33 ± 0.29% (depth), 97.80 ± 0.05% (future RGB), and 97.03 ± 0.33% (depth + flow) on the three new initial-state sets (Appendix C.3, Table 26). These are means ± sample SD across evaluation sets for fixed policies; not all training runs were tested on the new states.

Q3. Uncertainty in LIBERO-Plus performance differences

We added paired task-cluster bootstrap confidence intervals, resampling original tasks within each suite to account for correlated perturbation instances (Appendix C.2).

The table reports percentage-point differences and 95% confidence intervals for selected checkpoints under Clean instructions.

ComparisonOriginal training seed 42Mean of training seeds 42/142/242
Regression − flow, augmented 224×224 inputs, depth supervision+1.27 [−0.06, +2.59]+0.84 [−0.49, +2.24]
Depth − future RGB, augmented 224×224 inputs, action regression+1.02 [−0.37, +2.39]+0.53 [−0.43, +1.54]
Depth − future RGB, unaugmented 256×256 inputs, action regression−0.41 [−1.97, +1.19]— (single training seed only)

All intervals include zero, so these results do not establish superiority. They quantify task-level evaluation uncertainty conditional on the trained, selected policies; training-run variability is reported separately as sample SDs. The revised paper reports the intervals in Appendices B.4, B.5, and D.1.

AuthorsOpenReview ↗

Follow-up on our response

Thank you again for your thoughtful review. We would like to briefly follow up on our November 11 response, which includes the three analyses you requested:

  1. A controlled comparison between DPT-style depth regression and native generative depth supervision, using the same backbone and depth labels.
  2. Complete ten-checkpoint LIBERO scans across three training seeds, together with results for both selected and fixed-final checkpoints.
  3. Paired, task-clustered confidence intervals for the regression–flow and depth–future-RGB differences, with conclusions qualified accordingly.

We would appreciate your feedback on whether these additions and our clarification of the contribution address your concerns. We are happy to clarify any remaining questions.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses and Revisions

We thank all three reviewers for their constructive feedback.

GenPA adapts image generators into efficient robot policies through direct action learning and native current-scene perception co-training, and establishes current perception as an effective alternative to future-prediction co-training.

To facilitate the final assessment, we summarize this contribution and the main evidence added during revision.

  • Current perception as a training route for image-generator policies (all reviewers). GenPA combines direct action learning with current-scene perception co-training through the generator's native visual interface. This enables controlled comparisons of current targets, future targets, and their combinations within a common policy architecture. The resulting evidence establishes that current perception can improve control robustness without requiring future prediction, and that combining individually useful targets does not consistently improve performance. This provides both an efficient policy and concrete guidance on choosing visual supervision for control.

  • Repeated training and evaluation beyond checkpoint selection (all reviewers). We added three-seed training for the main action-head and supervision comparisons, complete ten-checkpoint LIBERO scans, results for both selected and fixed-final checkpoints, and separate fixed-policy evaluation on three new initial-state sets disjoint from selection. Current-depth co-training improves mean LIBERO-Plus success over action-only training under both instruction protocols and at both selected and final checkpoints. Independent LIBERO evaluation further supports the benefit of current-depth supervision. Paired task-cluster bootstrap analyses complement these results by quantifying uncertainty in the main comparisons.

  • Direct depth-regression control (c1uP, irNp). We added the requested DPT-style depth-regression baseline using the same backbone, depth labels, training data, and action-update budget. Across three training seeds, native depth supervision achieves similar standard LIBERO performance and higher mean LIBERO-Plus success at both selected and final checkpoints. The revised paper reports the corresponding training costs, providing a direct characterization of the performance–cost trade-off of the native supervision interface.

  • Efficiency across the full policy pipeline (irNp). We extended the resource analysis with training and inference memory and estimated training GPU-hours, complementing the component-level latency study of observation processing, instruction reuse, and action decoding. GenPA achieves 20.9 ms per action chunk on an H100 with compilation and cached instructions, while perception co-training adds no inference-time computation. These measurements make the practical deployment benefits and resource requirements explicit.

We are encouraged that all three reviewers now recommend acceptance: c1uP and irNp raised their ratings following the additional experiments and clarifications, and ykMf maintained their positive recommendation. We appreciate their engagement and the improvements that this discussion has brought to the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.