acceptodds
Under review as a conference paper at ICLR 2027

RAP: Recursive Action-Preview Flow Matching for Fast and Robust Visuomotor Control

Abstract

Recent diffusion and flow-matching methods enable few-step action generation by initializing from temporally grounded sources derived from historical actions or previous action forecasts. However, source construction can differ between offline training and deployment: training sources rely on expert or ground-truth actions, whereas deployment sources come from the policy's own executed or predicted actions. This train–deployment source mismatch can make the learned source-to-action mapping unreliable as deployment sources deviate from the training source distribution, degrading policy performance and robustness. Adding Gaussian noise to temporal sources can improve robustness, but requires manual tuning of the noise scale. We propose Recursive Action-Preview Flow Matching (RAP) to align source construction between training and deployment. RAP carries forward the unexecuted part of the previous action forecast as an action preview to construct the source for the next action-generation cycle, while the latest observation guides its refinement. During training, RAP recursively constructs sources from detached policy-generated action previews, matching the source formation used at deployment without additional Gaussian noise injection. Across 8 simulated tasks from 4 benchmarks and 2 real-world tasks, RAP achieves strong performance with only 1 or 3 inference steps. Under large initial-pose perturbations, RAP improves success by 22–50 percentage points over noise-free baselines while remaining competitive with methods using Gaussian source perturbations. RAP combines fast action generation with robustness to both initial-pose changes and visual uncertainty.

Primary area:
applications to robotics, autonomy, planning
Keywords:
flow matching, visuomotor policy, robot learning, source distribution alignment, inference efficiency, robust control

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_vLJTOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes Recursive Action-Preview Flow Matching (RAP), a few-step visuomotor policy that carries the unexecuted portion of each predicted action chunk forward as the source for the next prediction. Self-Generated Recursive Source Training (SG-RST) constructs these previews by recursively running the policy on consecutive observations from offline demonstrations, then detaches them and supervises their refinement toward expert actions. Temporal Visual Fusion (TVF) combines the cached visual feature associated with the preview with the current observation feature. The paper evaluates RAP on eight simulated tasks and two Franka-arm tasks, reporting competitive success with one or three function evaluations, improved success under initial-pose perturbations, and a 64% smaller paired source gap than WP-Preview on recorded Close Box observation sequences. The supplementary guide describes the accompanying real-robot video and anonymous code repository.

Critical strengths
  1. The paper identifies a concrete failure mode in temporal-source policies: training sources built from demonstrations differ from the policy-generated previews used at deployment. SG-RST directly exposes the learner to its own recursively generated source errors, and TVF provides a simple way to combine the preview's visual context with current feedback without re-encoding the earlier window. The mechanism is clearly specified and directly connected to the stated problem.
  2. The evaluation covers several simulation benchmarks, initial-pose and visual perturbations, and two real-robot tasks. RAP is competitive at one or three function evaluations, and the latency comparison shows that its online cost is similar to WP-Preview at matched inference budgets. The source-gap analysis and implementation appendix make the central training-alignment claim more inspectable than task success alone.
Critical weaknesses
  1. The paper makes the difference between expert-derived training previews and policy-generated inference previews central to its motivation, but does not establish that this difference is a harmful mismatch rather than the ordinary consequence of supervised learning from expert demonstrations.
  2. The 64% source-gap reduction is measured on shared, replayed observations. It isolates source construction but does not show that the gap causes closed-loop failures or that reducing it improves closed-loop task outcomes. The initial-pose experiments do not directly establish this causal link.
  3. The methodological increment over prior action-chunk and warm-start approaches is unclear. RTI-DP, SDP, and RNR-DP also carry forward or shift previously predicted actions, while A2A and WarmPrior are related temporal-source methods. The paper does not include a direct comparison that isolates recursive policy-generated previews from these approaches.
  4. The experiments do not include a no-training baseline that predicts a longer action chunk and shifts it, or a matched comparison using additional history frames (e.g., J=1, 2, 3, 4). Without these controls, it is difficult to attribute the gain specifically to SG-RST.
  5. The added training computation is not compared with the baseline methods, leaving the cost of recursive source generation unclear.
  6. The simulation results provide limited seed-level uncertainty, making the reliability of the reported differences across independently trained policies difficult to assess.
  7. The real-robot evaluation uses only a small number of trials, limiting the strength of the task-success evidence.
  8. Publicly accessible code and evaluation artifacts would be needed to independently reproduce the experiments.
Questions for authors
  1. What evidence shows that the difference between expert-derived training previews and policy-generated inference previews causes deployment failures, rather than reflecting the standard supervised-learning setup?
  2. Could you measure the source gap and task success in fully closed-loop evaluations, where each policy's actions produce its subsequent observations, and test whether reducing the gap improves outcomes?
  3. How does RAP compare with RTI-DP, SDP, and RNR-DP at matched data and compute budgets?
  4. What is the gain over a no-training baseline that predicts a longer action chunk and shifts it?
  5. How does performance vary with the number of history frames (e.g., J=1, 2, 3, 4)?
  6. How much additional training computation does recursive source generation require relative to the baselines?
  7. Could you report per-training-seed results and uncertainty for the main simulation and perturbation comparisons?
  8. How are the 100 simulation rollouts distributed across training seeds?
  9. Could you provide trial-level outcomes for the real-robot tasks?
  10. Could you release the code and evaluation artifacts needed to reproduce the experiments?
  11. Could you report a full-protocol ablation separating SG-RST from TVF?
  12. How does TVF compare with feature concatenation and current-frame-only conditioning?

I am open to raise my score if author solve my concerns.

6 comments
AuthorsOpenReview ↗

Response to reviewer vLJT (1/4)

We sincerely thank you for the careful review and constructive feedback. Below, we address your concerns point by point, grouping closely related weaknesses and questions where appropriate.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Reply to W1 & Q1: Source mismatch harms closed-loop performance despite identical expert supervision.

Thank you for raising this important concern. Our original Introduction and Fig. 1 already identify the potential harm of train-deployment source mismatch in temporal-source policies. This motivation is supported by A2A (Jia et al., 2026, Sec. 4.3.2), which reports being “more sensitive to uncertainties within the action history” than noise-initialized baselines, as also reflected in our Fig. 6. To more directly validate the impact of train-deployment source mismatch beyond general supervised-learning effects, we conducted additional controlled experiments.

Controlled source-construction ablation. Both variants use the same architecture, TVF, NFE=3, repeat-last source padding, and expert-action supervision. The difference lies in training-source construction: w/o SG-RST (σ=0) uses expert action chunks, whereas RAP uses recursively generated policy previews. Here, σ denotes the standard deviation of Gaussian source noise, with σ=0 indicating no noise injection. RAP likewise uses no additional Gaussian source noise.

Table R1: Success rates (%) under nominal and perturbed conditions.

EvaluationMethodClose BoxPick CubeStack Cube
Nominal (30 epochs)w/o SG-RST (σ=0)88.67 ± 0.5879.00 ± 1.000.00 ± 0.00
Nominal (30 epochs)RAP (Ours)91.67 ± 4.7397.67 ± 0.5896.00 ± 2.00
Perturbed (50 epochs)w/o SG-RST (σ=0)46.67 ± 3.0618.67 ± 5.034.00 ± 2.00
Perturbed (50 epochs)RAP (Ours)98.67 ± 1.1589.33 ± 1.1584.67 ± 5.03

Nominal denotes the standard initial robot pose; perturbed denotes initial pose perturbations within ±0.08 rad. Results are mean ± sample standard deviation over 3 seeds, with 100 nominal or 50 perturbed rollouts per seed.

These controlled results show that training-source construction substantially affects closed-loop performance despite identical expert supervision. RAP improves success by 52.00, 70.66, and 80.67 percentage points under ±0.08 rad initial-pose perturbations on Close Box, Pick Cube, and Stack Cube, respectively, while also improving nominal performance. These findings support our motivation that train-deployment source mismatch has practical consequences beyond general supervised-learning effects.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Reply to W2 & Q2: Smaller source gaps accompany higher success in closed-loop evaluations.

Thank you for this suggestion. To directly address this concern, we conducted additional fully closed-loop evaluations to examine the relationship between source consistency and task success. We compare RAP trained with recursive depths J=1 and J=2, using NFE=3 and three seeds. Each policy executes its own actions and generates its subsequent observations. For each seed, we evaluate 50 episodes with paired initial states and perturbations across J=1 and J=2. At each valid decision point, we reconstruct the training-style source using the corresponding recursive depth on the policy's own observation history and compare it with the continuously carried deployment source using normalized-action RMSE.

Table R2: Closed-loop success rates (%) and source gaps under nominal and perturbed conditions.

TaskSettingSuccess J=1Success J=2Gap J=1 ↓Gap J=2 ↓
Close BoxNominal69.33 ± 3.0692.00 ± 5.290.05384 ± 0.000230.01967 ± 0.00030
Close BoxPerturbed78.67 ± 3.0698.67 ± 1.150.06337 ± 0.000330.01401 ± 0.00024
Pick CubeNominal82.00 ± 12.4998.00 ± 0.000.30330 ± 0.000870.05693 ± 0.00034
Pick CubePerturbed74.67 ± 7.5789.33 ± 1.150.28706 ± 0.003430.07570 ± 0.00446
Stack CubeNominal6.67 ± 8.3396.00 ± 2.000.03317 ± 0.000140.01328 ± 0.00015
Stack CubePerturbed12.67 ± 1.1584.67 ± 5.030.04416 ± 0.001480.01267 ± 0.00031

Nominal and perturbed evaluations use epoch-30 and epoch-50 checkpoints, respectively; perturbed denotes initial pose perturbations within ±0.08 rad. Results are mean ± sample standard deviation over three seeds (50 episodes per seed). Source gaps are averaged over valid decisions within each episode, then equally across episodes.

Across all six settings, J=2 consistently achieves smaller closed-loop source gaps and higher task success than J=1. These results provide direct closed-loop evidence that improved source-construction consistency accompanies better task outcomes.

AuthorsOpenReview ↗

Response to reviewer vLJT (2/4)

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Reply to W3 & Q3: Methodological Distinctions and Additional Baseline Comparisons.

Thank you for suggesting these comparisons. We have expanded the Related Work section in our revised manuscript to discuss RTI-DP, SDP, and RNR-DP, clarify their distinctions from RAP, and conducted additional comparisons on all eight simulation tasks.

Methodological distinctions. RTI-DP shifts completed action predictions to warm-start diffusion inference, while SDP and RNR-DP carry partially denoised action buffers across control cycles to reuse denoising progress. RAP also reuses previous predictions, but carries forward the unexecuted portion of a completed action forecast as the next flow-matching source. SG-RST further trains the policy to refine these recursively generated sources under expert supervision, aligning training with their deployment-time construction.

Additional baseline comparisons. We evaluated the official open-source implementations of RTI-DP, SDP, and RNR-DP on the same eight simulation tasks, matching the inference budget at NFE=3. RAP (1) is additionally reported as a lower-budget reference.

Table R3: Simulation success rates (%) of RAP, RTI-DP, SDP, and RNR-DP.

TaskRAP (1)RAP (3)RTI-DP (3)SDP (3)RNR-DP (3)
Close Box89.33 ± 1.1592.00 ± 5.2974.67 ± 1.1576.67 ± 3.0652.00 ± 18.33
Pick Cube98.00 ± 2.0098.00 ± 0.0088.67 ± 3.0684.00 ± 9.1715.33 ± 10.26
Stack Cube94.67 ± 3.0696.00 ± 2.0084.00 ± 5.2980.00 ± 10.009.33 ± 12.70
Pull Cube89.33 ± 1.1591.33 ± 2.3186.00 ± 4.0092.67 ± 3.0673.33 ± 11.72
Open Drawer86.00 ± 6.0085.33 ± 3.0620.00 ± 10.3928.67 ± 5.035.33 ± 3.06
Pick-Place Bowl74.00 ± 4.0085.33 ± 3.0670.67 ± 17.0189.33 ± 1.150.00 ± 0.00
Square82.00 ± 3.4686.00 ± 2.006.00 ± 4.0055.33 ± 9.870.00 ± 0.00
Tool Hang28.00 ± 3.4620.67 ± 1.158.67 ± 7.5718.00 ± 5.290.00 ± 0.00
Average80.1781.8354.8465.5819.42

Results are mean ± SD over 3 seeds, with 50 rollouts per seed. The average is computed across eight tasks with equal weighting. Parentheses indicate NFE. Bold indicates the highest mean success rate per task.

At NFE=3, RAP achieves the highest average success rate across all eight tasks (81.83%), exceeding the strongest compared baseline, SDP (65.58%), by 16.25 percentage points. These results further support RAP's effectiveness and the value of learning from recursively generated policy previews.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Reply to W4 & Q4 & Q5: Ablation Studies Attribute Performance Gains to SG-RST.

Thank you for these constructive suggestions. We understand the central concern to be whether RAP's gains come specifically from SG-RST rather than longer action chunks, action shifting, or additional history. To isolate SG-RST's contribution, we compare RAP against an otherwise identical variant without recursive source training and further evaluate recursive depths J=1-4.

Controlled comparison without SG-RST (Q4). Our w/o SG-RST (σ=0) baseline retains the same 2H-action prediction, execution of the first H actions, and reuse of the remaining H actions as the next source. It also uses the same architecture, TVF, NFE=3, and repeat-last padding as RAP. The key difference is that training sources are constructed from expert actions rather than recursively generated policy previews. This isolates the benefit of SG-RST while keeping inference-time action reuse unchanged. As shown in Table R1 (W1 & Q1), RAP substantially improves success under both nominal and perturbed conditions, demonstrating gains beyond simply predicting and shifting longer action chunks.

Recursive-depth ablation (Q5). Here, J denotes the number of preceding policy calls, rather than the number of observation frames, which remains fixed at eight. We extended the original ablation to J=4 and evaluated all four depths across 3 training seeds.

AuthorsOpenReview ↗

Response to reviewer vLJT (3/4)

Table R4: Success rates (%) across recursive training depths.

TaskNFEJ=1J=2J=3J=4
Close Box156.33 ± 10.7989.00 ± 1.0073.00 ± 5.5796.00 ± 2.65
Close Box369.33 ± 3.0691.67 ± 4.7378.33 ± 5.6995.00 ± 4.36
Pick Cube164.67 ± 8.0898.00 ± 2.0097.33 ± 1.1597.00 ± 1.73
Pick Cube382.00 ± 12.4997.67 ± 0.5898.00 ± 1.7394.00 ± 4.58
Stack Cube14.00 ± 5.2994.33 ± 2.5289.67 ± 7.0295.00 ± 4.36
Stack Cube36.33 ± 8.5096.00 ± 2.0093.00 ± 8.8996.00 ± 2.65

Results are mean ± sample standard deviation over 3 seeds, with 100 rollouts per seed. Bold indicates the highest mean success rate for each setting.

Increasing J from 1 to 2 consistently improves success across all six settings. Together, these controlled comparisons support that SG-RST contributes beyond inference-time action reuse, with J=2 offering a favorable balance between performance and recursive training cost.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Reply to W5 & Q6: Explicit computational analysis quantifies the additional cost of recursive training.

Thank you for suggesting a direct training-cost comparison. Appendix A.3 already quantifies the additional vector-field evaluations required by SG-RST. To further clarify the practical overhead, we measured training-step time against two baselines under identical settings: Stack Cube, RTX 5090, batch size 32, NFE=3, and FP32.

Table R5: Measured training cost of RAP and baselines.

MethodTime/step (ms) ↓Peak GPU memory (GiB) ↓
w/o SG-RST + TVF (σ=0)267.12 ± 1.8917.46
WP-Preview (σ=1)266.94 ± 0.6017.46
RAP (J=2)315.04 ± 1.2018.04

Each method was measured in three independent timing runs, each with 50 warm-up steps and 200 timed steps. Timings include data loading, transfer, source generation, loss computation, backward propagation, optimizer updates, and EMA. Values are mean ± sample SD across timing runs; memory denotes peak allocated GPU memory.

SG-RST increases training-step time by only 17.94% relative to the non-recursive control and 18.02% relative to WP-Preview, with an additional 0.58 GiB of GPU memory. This modest overhead arises from two detached recursive policy calls, which require no backward propagation. Importantly, the additional computation is confined to training, while deployment retains the same NFE and comparable inference latency, as reported in Appendix A.4.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Reply to W6 & Q7 & Q8: Multi-seed evaluations quantify uncertainty.

Thank you for raising this concern. The A2A baseline reported results from a single training seed, and we followed this evaluation setup in our original submission. We have now extended the main simulation and perturbation evaluations to 3 training seeds, with 50 rollouts per seed. The corresponding mean ± sample SD results are reported in our response to Reviewer nVq5 (W2, Tables R2 and R3). We additionally provide the per-training-seed results below.

Table R6: Per-training-seed success rates (%) for the main simulation comparisons.

TaskRAP (1)RAP (3)WP (1)WP (3)A2A (6)VITA (6)FM-UNet (10)DDIM-UNet (40)
Close Box90/88/9098/90/8890/88/8890/92/8890/90/9880/74/7262/70/7054/70/64
Pick Cube100/96/9898/98/9898/98/96100/98/9692/90/9084/92/9484/92/8660/72/62
Stack Cube98/94/9294/96/9884/10/6496/80/8884/96/8668/78/8024/20/2428/40/36
Pull Cube90/88/9094/90/9088/82/9294/92/9090/94/9066/94/9276/86/8476/92/84
Open Drawer92/80/8688/86/8284/80/8288/76/7696/84/8682/80/7852/42/4264/62/68
Pick-Place Bowl74/70/7886/82/8880/70/7278/66/7882/84/9080/80/8056/42/6060/70/56
Square80/80/8684/88/8670/80/8038/64/5468/56/5416/38/4038/28/1424/2/0
Tool Hang32/26/2620/22/2022/28/2618/24/2024/26/3418/14/108/20/820/22/32

Each cell lists success rates for 3 seeds, with 50 rollouts per seed. Parentheses in column headers indicate NFE.

We further repeated the strongest initial-pose perturbation (±0.08 rad) using RAP, A2A, and VITA, with 3 training seeds and 50 rollouts per seed.

Table R7: Per-training-seed success rates (%) under ±0.08 rad initial-pose perturbations.

MethodClose BoxPick CubeStack Cube
RAP (1)96/94/9688/72/8476/86/78
RAP (3)100/98/9890/90/8884/90/80
A2A (6)66/72/7046/50/3834/40/30
VITA (6)78/68/7650/64/8044/50/42
Reviewer_nVq5OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

RAP learns to refine recursively generated action previews through flow matching, conditioned on cached and current visual features. Training reuses the policy’s own predictions along recorded demonstrations. Experiments cover eight simulated tasks and two real-robot tasks, mainly using one or three flow evaluations.

Critical strengths

Recursive preview generation addresses a concrete mismatch in temporal-source policies. The consistent improvement from one to two preceding calls in Table 3 supports the value of training on reused predictions.

The WP-Preview comparison matches visual conditioning and inference budget. The measured latency in Table 7 also supports the efficiency claim beyond simply counting flow evaluations.

Critical weaknesses

The source audit holds expert observations fixed. Its paired-gap result is useful, but the unpaired distribution metrics favor WP-Preview (Appendix B.5). The connection between this diagnostic and closed-loop robustness is therefore only partly established.

Variation across independent training seeds is not reported. On the physical tasks, the gains over WP-Preview amount to one or two additional successes out of 20 trials, which is limited evidence of a reliable advantage.

Questions for authors

Do the main gains, particularly the large improvement from recursion depth one to two, persist across independent training seeds?

Is the paired source-gap reduction preserved on observations collected from RAP’s own rollouts under initial-pose perturbations?

2 comments
AuthorsOpenReview ↗

Response to reviewer nVq5 (1/2)

We sincerely thank you for the thorough review and constructive feedback. We appreciate your valuable suggestions and address the concerns and questions point by point below.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Reply to W1&Q2: Clarifying source diagnostics and linking recursive alignment to closed-loop robustness.

Thank you for this valuable suggestion. Our original source audit held expert observations fixed to enable direct comparison of paired training-deployment source gaps. As explained in Appendix B.5, WP-Preview's favorable unpaired metrics results from its strong Gaussian source noise (σ=1.0), which masks smaller preview discrepancies and makes the training and deployment residual distributions difficult to distinguish.However, Figure 7(c) shows that Gaussian noise is nearly temporally uncorrelated, whereas actual deployment preview errors exhibit strong temporal correlations. Therefore, favorable unpaired metrics do not necessarily indicate better alignment with the structured preview errors encountered at deployment.

Yes, the source-gap reduction from recursive training persists on policy-generated observations under initial-pose perturbations, accompanied by improved closed-loop success. We conducted a controlled J=1 vs. J=2 ablation on Close Box, Pick Cube, and Stack Cube (NFE=3, epoch-50 checkpoints) under ±0.08 rad initial-pose perturbations. Each policy executed its own actions to generate closed-loop observations. At each eligible decision, we reconstructed its training-style source using the corresponding recursion depth on the recorded observations and compared it with the actual continuously maintained deployment source.

Table R1: Closed-loop success rates (%) and source gaps under ±0.08 rad initial-pose perturbations.

TaskSuccess J=1Success J=2Gap J=1 ↓Gap J=2 ↓
Close Box78.67 ± 3.0698.67 ± 1.150.06337 ± 0.000330.01401 ± 0.00024
Pick Cube74.67 ± 7.5789.33 ± 1.150.28706 ± 0.003430.07570 ± 0.00446
Stack Cube12.67 ± 1.1584.67 ± 5.030.04416 ± 0.001480.01267 ± 0.00031

Mean ± sample SD across 3 seeds, with 50 rollouts per seed. Source gaps are RMS differences in normalized action space, averaged within each episode and then equally across episodes.

Across all three tasks, J=2 consistently achieves smaller source gaps and higher success rates, reducing the gaps by 71.3%-77.9% while improving closed-loop success under perturbations. This controlled ablation supports the connection between recursive source alignment and closed-loop robustness.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Reply to W2: Multi-seed validation and competitive performance without noise-scale tuning.

Thank you for raising this concern. The A2A baseline reported results from a single training seed, and we followed this evaluation setup in our original submission. We now provide results across 3 training seeds for both the original Table 1 comparisons (Table R2) and the ±0.08 rad initial-pose perturbation experiments (Table R3). These results further support our main claim that RAP maintains strong nominal performance while consistently achieving robust performance under large initial-pose perturbations. Our 20-trial real-robot evaluation also follows the trial scale of prior work, such as A2A (10 trials) and VITA (20 trials). The modest gains over WP-Preview are understandable, as both methods benefit from informative action previews. Importantly, WP-Preview uses Gaussian source noise (σ=1.0) selected through noise-scale ablations, whereas RAP achieves competitive or superior performance without additional Gaussian source noise or noise-scale tuning.

AuthorsOpenReview ↗

Response to reviewer nVq5 (2/2)

Table R2: Nominal success rates (%) across 3 training seeds (mean ± SD).

TaskRAP (1)RAP (3)WP (1)WP (3)A2A (6)VITA (6)FM-UNet (10)DDIM-UNet (40)
Close Box89.33 ± 1.1592.00±5.29‾\underline{92.00 \pm 5.29}88.67 ± 1.1590.00 ± 2.0092.67 ± 4.6275.33 ± 4.1667.33 ± 4.6262.67 ± 8.08
Pick Cube98.00 ± 2.0098.00 ± 0.0097.33±1.15‾\underline{97.33 \pm 1.15}98.00 ± 2.0090.67 ± 1.1590.00 ± 5.2987.33 ± 4.1664.67 ± 6.43
Stack Cube94.67±3.06‾\underline{94.67 \pm 3.06}96.00 ± 2.0052.67 ± 38.2888.00 ± 8.0088.67 ± 6.4375.33 ± 6.4322.67 ± 2.3134.67 ± 6.11
Pull Cube89.33 ± 1.1591.33±2.31‾\underline{91.33 \pm 2.31}87.33 ± 5.0392.00 ± 2.0091.33±2.31‾\underline{91.33 \pm 2.31}84.00 ± 15.6282.00 ± 5.2984.00 ± 8.00
Open Drawer86.00±6.00‾\underline{86.00 \pm 6.00}85.33 ± 3.0682.00 ± 2.0080.00 ± 6.9388.67 ± 6.4380.00 ± 2.0045.33 ± 5.7764.67 ± 3.06
Pick-Place Bowl74.00 ± 4.0085.33 ± 3.0674.00 ± 5.2974.00 ± 6.9385.33 ± 4.1680.00±0.00‾\underline{80.00 \pm 0.00}52.67 ± 9.4562.00 ± 7.21
Square82.00±3.46‾\underline{82.00 \pm 3.46}86.00 ± 2.0076.67 ± 5.7752.00 ± 13.1159.33 ± 7.5731.33 ± 13.3226.67 ± 12.068.67 ± 13.32
Tool Hang28.00 ± 3.4620.67 ± 1.1525.33±3.06‾\underline{25.33 \pm 3.06}20.67 ± 3.0628.00 ± 5.2914.00 ± 4.0012.00 ± 6.9324.67 ± 6.43

Table R3: Success rates (%) under ±0.08 rad initial-pose perturbations across 3 training seeds (mean ± SD).

MethodClose BoxPick CubeStack Cube
RAP (1)95.33±1.15‾\underline{95.33 \pm 1.15}81.33±8.33‾\underline{81.33 \pm 8.33}80.00±5.29‾\underline{80.00 \pm 5.29}
RAP (3)98.67 ± 1.1589.33 ± 1.1584.67 ± 5.03
A2A (6)69.33 ± 3.0644.67 ± 6.1134.67 ± 5.03
VITA (6)74.00 ± 5.2964.67 ± 15.0145.33 ± 4.16

50 rollouts per seed. Bold and underline indicate the best and second-best distinct mean success rates in each row of Table R2 and each task column of Table R3, respectively. Parentheses following method names indicate NFE.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Reply to Q1: Demonstrating consistent improvements from J=1 to J=2 across independent seeds.

Thank you for this question. Yes, the substantial improvements from J=1 to J=2 consistently persist across 3 training seeds. We repeated the recursion-depth ablation in Table 3 on Close Box, Pick Cube, and Stack Cube at both NFE=1 and NFE=3. As shown below, J=2 consistently outperforms J=1 across all six settings, confirming the effectiveness of recursive source training.

Table R4: Success rates (%) across 3 training seeds, with 100 rollouts per seed (mean ± SD).

TaskNFEJ=1J=2
Close Box156.33 ± 10.7989.00 ± 1.00
Close Box369.33 ± 3.0691.67 ± 4.73
Pick Cube164.67 ± 8.0898.00 ± 2.00
Pick Cube382.00 ± 12.4997.67 ± 0.58
Stack Cube14.00 ± 5.2994.33 ± 2.52
Stack Cube36.33 ± 8.5096.00 ± 2.00
Reviewer_M6UnOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

RAP is an A2A-based latent flow-matching policy that executes the first H of 2H predicted actions, reuses the remaining H padded by repeating the last action as the next source, trains on sources from two detached recursive calls called SG-RST, and linearly fuses cached and current visual features through TVF. With 1 or 3 NFEs it reports best or tied-best success on eight RoboVerse simulation tasks and two Franka tasks, 22–50 pp gains over A2A and VITA under 0.08 rad initial-pose perturbation, and a 64% smaller mean paired source gap than WP-Preview on Close Box.

Critical strengths

The paper makes a convincing case that WarmPrior trains and deploys on different kinds of preview sources, and SG-RST closes this gap with offline data alone and without slowing inference.

RAP stays robust to initial-pose perturbations without any hand-tuned noise, outperforming A2A and VITA by 22 to 50 points at 0.08 rad.

Critical weaknesses
  1. SG-RST's core idea of training on the model's own detached predictions, so that training inputs resemble test-time inputs, also underlies scheduled sampling by Bengio et al. 2015, DAgger and DaD-style training on self-generated inputs, and self-conditioning by Chen et al. 2022, none of which is discussed. Chunk-reuse methods such as Real-Time Chunking and its training-time variant, Bidirectional Decoding, Streaming Diffusion Policy and Falcon are also uncited, and relative to WP-Preview the change reduces to a self-generated preview in place of the expert proxy, no σ = 1.0 noise and linear feature fusion.

  2. SG-RST replaces deployment recursion with two detached calls that restart from the recorded robot state and run on expert observations, yet the paper neither argues that this truncated open-loop unroll matches the sources of unbounded closed-loop recursion nor analyzes how preview error compounds with depth. Its diagnostics suggest residual mismatch, as RAP's training and deployment residuals are more separable than WP-Preview's at logistic AUC 0.8597 against 0.5349, and J = 3, which should be closer to deployment, lowers 1-NFE and 3-NFE Close Box success in Table 3 from 89% and 97% to 68% and 80%.

  3. In Table 1, the A2A, VITA, FM-UNet and DDIM-UNet entries for Close Box, Pick Cube, Stack Cube and Open Drawer are identical to Table 1 of the A2A paper, use unmatched NFEs of 6, 10 and 40 with different backbones, and come without seed counts or error bars. Against WP-Preview, the only baseline with matched architecture and NFE, 3-NFE RAP is within 3 pp, about one binomial standard error at n = 100, or behind on five of eight tasks, and its real-robot margins amount to one and two extra successes in 20 trials.

Questions for authors
  1. Were the Table 1 baselines retrained in your pipeline, and how many training seeds underlie each cell? Please report mean and standard deviation over at least three seeds for RAP and WP-Preview, and add A2A at 1 and 3 NFEs as matched-budget baselines.

  2. How does the same architecture perform with expert-chunk sources and σ set to 0, 0.1, 0.3 or 1.0 but without SG-RST, a control that would separate the effect of removing the train–deploy mismatch from that of dropping WarmPrior's σ = 1.0 noise? Please also state how your WP-Preview fills the last H source positions.

3 comments
AuthorsOpenReview ↗

Response to reviewer M6Un (1/3)

We sincerely thank you for the thorough review and valuable feedback. We appreciate your recognition of RAP's strengths and have carefully addressed the concerns through additional experiments, analyses, and clarifications. Our point-by-point responses are provided below.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Reply to W1: Clarifying RAP's Distinctions from Prior Methods and Its Contributions

Thank you for highlighting these relevant connections. We have expanded Related Work to discuss the suggested methods and clarify their distinctions from RAP. Our original discussion primarily focused on few-step generative policies and temporal-source methods, which are most directly related to RAP's problem setting. The suggested works provide complementary perspectives on exposure bias in sequence learning, action-chunk continuity, and streaming execution.

Relation to self-generated-input training. Scheduled Sampling and DaD primarily address exposure bias in sequence prediction, while DAgger addresses learner-induced state distribution shift through expert labeling of learner-visited states. Self-conditioning reuses model predictions as additional conditioning signals rather than flow sources. SG-RST addresses a related but distinct challenge: the train-deployment mismatch in flow-matching sources for few-step visuomotor control. It recursively constructs policy-generated action-preview sources for expert-supervised flow refinement using existing offline demonstrations, without additional environment interaction or expert relabeling.

Relation to action-chunk reuse. RTC and its training-time variant address execution latency and temporal consistency, BID improves coherence and reactivity through test-time selection, while SDP and Falcon accelerate action generation by reusing partially denoised predictions. RAP focuses on reducing the train-deployment source mismatch arising from recursive preview reuse. Accordingly, SG-RST incorporates policy-generated previews into offline training, enabling the flow model to learn their refinement for fast and robust visuomotor control.

Relation to WarmPrior. WarmPrior-Preview constructs training sources from expert proxies and relies on Gaussian perturbations to improve performance. Building on preview-based initialization, RAP learns directly from recursively generated policy previews, while TVF combines cached and current visual features to guide their refinement. Together, SG-RST and TVF reduce the train-deployment source mismatch and enable effective few-step control without additional Gaussian source-noise tuning.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Reply to W2: Validating Two-Call Approximation and Analyzing Recursive Preview Errors, Residual AUC, and Depth Effects

Thank you for raising this valuable concern. As described in Section 4.2, SG-RST uses a short offline recursion to efficiently approximate deployment-time source construction, without assuming an exact match with unbounded closed-loop recursion. To directly evaluate the effectiveness of this approximation and analyze how preview errors evolve during continuous recursion, we conducted additional experiments.

1. Two-call reconstruction provides an effective balance between approximation accuracy and training cost.

Our new experiments show that two-call reconstruction closely approximates continuously maintained previews, with normalized RMSEs of only 0.01502, 0.01434, and 0.01269 under expert, nominal deployment, and perturbed deployment observations, respectively.

We compare previews reconstructed with different numbers of preceding policy calls against continuous recursion, using the same frozen policy and observation sequence at identical decision points.

Observation streamJ=1J=1J=2J=2J=3J=3J=4J=4
Expert observations0.085570.015020.005070.00176
Nominal deployment0.080720.014340.004620.00141
Perturbed deployment0.070030.012690.004160.00131

Normalized preview RMSE against continuous recursion on Stack Cube, using NFE=3 and the epoch-30 EMA checkpoint. Expert, nominal deployment, and perturbed deployment observations comprise 99, 100, and 50 trajectories, respectively. Perturbed rollouts use initial joint offsets within ±0.08 rad. Means are equally weighted across trajectories.

Increasing reconstruction depth from one to two calls reduces the discrepancy by approximately 82% across all three settings, while further calls provide diminishing marginal benefits in reducing the remaining discrepancy. Moreover, the second call already refines a policy-generated preview, capturing the recursive reuse mechanism at deployment. Thus, J=2J=2 achieves a favorable balance between approximation accuracy and training cost, avoiding the additional computation of deeper recursion for relatively limited gains.

AuthorsOpenReview ↗

Response to reviewer M6Un (2/3)

2. Recursive preview errors evolve non-monotonically rather than accumulating uniformly.

Our additional analysis shows that preview errors can both decrease and increase during continuous recursion, with task-dependent fluctuations and larger errors at later stages.

We measure the normalized RMSE between recursively generated previews and their corresponding future expert actions along recorded observation trajectories.

Recursive call (J)Stack CubePick CubeClose Box
10.018660.016470.03732
20.018170.013130.01628
30.013110.008670.00779
40.014070.007850.01328
50.015090.139250.01072
60.018710.124560.01244
70.123050.064170.01541
80.097230.029070.10717
90.080830.021550.07756
Early (1-3)0.016650.012760.02046
Late (7-9)0.100370.038260.06671

Mean normalized preview-to-expert RMSE over 99 trajectories per task, using NFE=3, epoch-30 EMA policies. Early and late values average the corresponding calls within each trajectory before averaging across trajectories.

The errors do not increase monotonically: for example, Pick Cube exhibits a pronounced peak at calls 5-6 followed by a decrease, while Close Box shows a local minimum at call 3. Nevertheless, all three tasks have higher average errors in calls 7-9 than in calls 1-3.

This behavior is consistent with recursive preview refinement: each new prediction incorporates updated observations, so previous errors may be reduced or amplified rather than simply added. These results characterize how preview errors evolve during sustained recursion.

3. WP-Preview's lower AUC Results from strong Gaussian noise masking smaller preview discrepancies.

As explained in Appendix B.5, WP-Preview's unpaired training and deployment residuals both contain Gaussian noise with standard deviation 1.0, which is substantially larger than typical preview discrepancies. This strong noise can mask the underlying differences, making the residual distributions difficult for the classifier to distinguish. Figure 7(c) shows that Gaussian noise is nearly temporally uncorrelated, whereas actual preview errors exhibit strong temporal correlations. Therefore, a lower AUC does not imply better alignment with actual deployment preview errors.

4. Non-monotonic training-depth performance may arise from different exposure to preview errors.

Notably, the Close Box diagnostic in Section 2 shows a clear decrease in preview error at the third recursive call (0.00779), compared with the second (0.01628) and fourth (0.01328). This suggests a possible explanation for the J=3 performance decrease: relatively accurate previews at certain depths may provide less challenging correction examples, limiting the learned ability to refine larger preview errors. At greater depths, exposure to larger discrepancies may strengthen this refinement capability.

Our additional three-seed ablation supports the non-monotonic depth-performance relationship:

NFEJ=1J=2J=3J=4
156.33 ± 10.7989.00 ± 1.0073.00 ± 5.5796.00 ± 2.65
369.33 ± 3.0691.67 ± 4.7378.33 ± 5.6995.00 ± 4.36

Close Box success rates (%), mean ± sample standard deviation across 3 training seeds, with 100 rollouts per seed.

The J=3 decrease is consistently observed across seeds, while J=4 recovers performance.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Reply to W3 & Q1: Clarifying Baseline Settings and Providing Matched-Budget, Multi-Seed Comparisons

Thank you for the careful examination of our experimental comparisons. The identical baseline results in Table 1 were intentionally adopted from the A2A paper where the experimental settings were identical, while the remaining baselines were independently trained and evaluated in our pipeline. Following the single-seed reporting format of the original A2A baseline, we initially reported one training seed per configuration. Moreover, the comparisons against baselines using 6/10/40 NFEs were designed to demonstrate that RAP achieves strong performance with substantially fewer inference steps.

Regarding the relatively modest gains over WP-Preview in both simulation and real-robot experiments, both methods already benefit from informative action previews, making their performance naturally competitive. However, WP-Preview requires selecting an appropriate Gaussian source-noise scale to achieve strong performance. The original WarmPrior paper identified (σ = 1.0) as a favorable setting through noise-scale ablations. While RAP achieves comparable or superior performance without additional Gaussian source noise or manual noise-scale tuning. This advantage applies to both simulation and real-robot settings.

AuthorsOpenReview ↗

Response to reviewer M6Un (3/3)

To further address Q1, we conducted additional experiments with 3 training seeds for RAP, WP-Preview, and A2A at matched inference budgets of 1 and 3 NFEs across all eight simulation tasks. The results are summarized below.

Table R1: Multi-seed success rates (%) under matched inference budgets.

TaskRAP (1)RAP (3)WP-Preview (1)WP-Preview (3)A2A (1)A2A (3)
Close Box89.33 ± 1.1592.00 ± 5.2988.67 ± 1.1590.00 ± 2.0089.33 ± 5.0391.33±4.16‾\underline{91.33 \pm 4.16}
Pick Cube98.00 ± 2.0098.00 ± 0.0097.33±1.15‾\underline{97.33 \pm 1.15}98.00 ± 2.0091.33 ± 2.3190.00 ± 2.00
Stack Cube94.67±3.06‾\underline{94.67 \pm 3.06}96.00 ± 2.0052.67 ± 38.2888.00 ± 8.0081.33 ± 7.0283.33 ± 4.16
Pull Cube89.33 ± 1.1591.33±2.31‾\underline{91.33 \pm 2.31}87.33 ± 5.0392.00 ± 2.0091.33±2.31‾\underline{91.33 \pm 2.31}90.00 ± 2.00
Open Drawer86.00 ± 6.0085.33±3.06‾\underline{85.33 \pm 3.06}82.00 ± 2.0080.00 ± 6.9384.67 ± 4.1685.33±1.15‾\underline{85.33 \pm 1.15}
Pick-Place Bowl74.00 ± 4.0085.33 ± 3.0674.00 ± 5.2974.00 ± 6.9377.33 ± 3.0684.00±5.29‾\underline{84.00 \pm 5.29}
Square82.00±3.46‾\underline{82.00 \pm 3.46}86.00 ± 2.0076.67 ± 5.7752.00 ± 13.1152.67 ± 6.4355.33 ± 8.33
Tool Hang28.00 ± 3.4620.67 ± 1.1525.33±3.06‾\underline{25.33 \pm 3.06}20.67 ± 3.0625.33±11.72‾\underline{25.33 \pm 11.72}23.33 ± 6.43

Values represent mean ± sample standard deviation over 3 seeds, with 50 rollouts per seed. Numbers in parentheses indicate NFEs. Bold and underline denote the best and second-best distinct mean success rates in each row, including ties.

The additional matched-budget, multi-seed results further support RAP's effectiveness.

━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━

Reply to Q2: Clarifying WP-Preview Source Construction and Isolating SG-RST's Contribution through Noise Ablations

Thank you for suggesting this more fine-grained ablation. Our original WP-Preview baseline follows WarmPrior, sampling the last (H) source positions from standard Gaussian noise (𝒩(0, I)) and adopting its (σ = 1.0) Gaussian source-noise setting. In contrast, RAP uses repeat-last padding without additional Gaussian noise. Therefore, to more precisely isolate SG-RST's contribution, we conducted additional controlled experiments with matched source padding and different Gaussian noise scales.

SG-RST is effective without Gaussian noise tuning, while noise-based alternatives depend on appropriate noise scales.

All variants use the same architecture, TVF, NFE=3, and repeat-last source padding. Without SG-RST, training sources are constructed from expert action chunks with (noise-scale σ ∈ {0, 0.1, 0.3, 1.0}). RAP uses recursively generated policy previews without additional Gaussian noise.

Table R2: Nominal success rates (%) after 30 training epochs.

MethodClose BoxPick CubeStack Cube
w/o SG-RST (σ = 0)88.67 ± 0.5879.00 ± 1.000.00 ± 0.00
w/o SG-RST (σ = 0.1)96.33 ± 1.5387.00 ± 1.730.00 ± 0.00
w/o SG-RST (σ = 0.3)91.33 ± 3.0696.33±0.58‾\underline{96.33 \pm 0.58}0.00 ± 0.00
w/o SG-RST (σ = 1.0)92.33±2.08‾\underline{92.33 \pm 2.08}95.67 ± 0.580.00 ± 0.00
RAP (Ours)91.67 ± 4.7397.67 ± 0.5896.00 ± 2.00

Table R3: Success rates (%) under ±0.08 rad initial-pose perturbations after 50 training epochs.

MethodClose BoxPick CubeStack Cube
w/o SG-RST (σ = 0)46.67 ± 3.0618.67 ± 5.034.00 ± 2.00
w/o SG-RST (σ = 0.1)90.67 ± 4.1628.67 ± 8.080.00 ± 0.00
w/o SG-RST (σ = 0.3)93.33 ± 1.1584.67 ± 1.1570.67 ± 1.15
w/o SG-RST (σ = 1.0)94.00±3.46‾\underline{94.00 \pm 3.46}93.33 ± 5.0383.33±2.31‾\underline{83.33 \pm 2.31}
RAP (Ours)98.67 ± 1.1589.33±1.15‾\underline{89.33 \pm 1.15}84.67 ± 5.03

Nominal denotes the standard initial robot pose. Both tables report mean ± sample standard deviation over 3 seeds, with 100 nominal or 50 perturbed rollouts per seed. Bold and underline indicate the best and second-best distinct mean success rates.

These results validate the effectiveness of SG-RST and its independence from Gaussian noise tuning. RAP consistently outperforms the noise-free expert-source baseline across all evaluated settings, confirming that its gains cannot be attributed simply to removing Gaussian noise. Moreover, the expert-source baselines exhibit substantial sensitivity to (σ), requiring an appropriate noise scale for strong performance. But SG-RST learns to refine policy-generated previews directly, achieving competitive or superior performance without additional Gaussian perturbations or manual noise-scale selection.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.