- 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
- 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.
- 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
- 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.
- 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.
- 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.
- 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.
- The added training computation is not compared with the baseline methods, leaving the cost of recursive source generation unclear.
- The simulation results provide limited seed-level uncertainty, making the reliability of the reported differences across independently trained policies difficult to assess.
- The real-robot evaluation uses only a small number of trials, limiting the strength of the task-success evidence.
- Publicly accessible code and evaluation artifacts would be needed to independently reproduce the experiments.
Questions for authors
- 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?
- 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?
- How does RAP compare with RTI-DP, SDP, and RNR-DP at matched data and compute budgets?
- What is the gain over a no-training baseline that predicts a longer action chunk and shifts it?
- How does performance vary with the number of history frames (e.g., J=1, 2, 3, 4)?
- How much additional training computation does recursive source generation require relative to the baselines?
- Could you report per-training-seed results and uncertainty for the main simulation and perturbation comparisons?
- How are the 100 simulation rollouts distributed across training seeds?
- Could you provide trial-level outcomes for the real-robot tasks?
- Could you release the code and evaluation artifacts needed to reproduce the experiments?
- Could you report a full-protocol ablation separating SG-RST from TVF?
- 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
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.
| Task | RAP (1) | RAP (3) | RTI-DP (3) | SDP (3) | RNR-DP (3) |
|---|---|---|---|---|---|
| Close Box | 89.33 ± 1.15 | 92.00 ± 5.29 | 74.67 ± 1.15 | 76.67 ± 3.06 | 52.00 ± 18.33 |
| Pick Cube | 98.00 ± 2.00 | 98.00 ± 0.00 | 88.67 ± 3.06 | 84.00 ± 9.17 | 15.33 ± 10.26 |
| Stack Cube | 94.67 ± 3.06 | 96.00 ± 2.00 | 84.00 ± 5.29 | 80.00 ± 10.00 | 9.33 ± 12.70 |
| Pull Cube | 89.33 ± 1.15 | 91.33 ± 2.31 | 86.00 ± 4.00 | 92.67 ± 3.06 | 73.33 ± 11.72 |
| Open Drawer | 86.00 ± 6.00 | 85.33 ± 3.06 | 20.00 ± 10.39 | 28.67 ± 5.03 | 5.33 ± 3.06 |
| Pick-Place Bowl | 74.00 ± 4.00 | 85.33 ± 3.06 | 70.67 ± 17.01 | 89.33 ± 1.15 | 0.00 ± 0.00 |
| Square | 82.00 ± 3.46 | 86.00 ± 2.00 | 6.00 ± 4.00 | 55.33 ± 9.87 | 0.00 ± 0.00 |
| Tool Hang | 28.00 ± 3.46 | 20.67 ± 1.15 | 8.67 ± 7.57 | 18.00 ± 5.29 | 0.00 ± 0.00 |
| Average | 80.17 | 81.83 | 54.84 | 65.58 | 19.42 |
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.
Response to reviewer vLJT (3/4)
Table R4: Success rates (%) across recursive training depths.
| Task | NFE | J=1 | J=2 | J=3 | J=4 |
|---|---|---|---|---|---|
| Close Box | 1 | 56.33 ± 10.79 | 89.00 ± 1.00 | 73.00 ± 5.57 | 96.00 ± 2.65 |
| Close Box | 3 | 69.33 ± 3.06 | 91.67 ± 4.73 | 78.33 ± 5.69 | 95.00 ± 4.36 |
| Pick Cube | 1 | 64.67 ± 8.08 | 98.00 ± 2.00 | 97.33 ± 1.15 | 97.00 ± 1.73 |
| Pick Cube | 3 | 82.00 ± 12.49 | 97.67 ± 0.58 | 98.00 ± 1.73 | 94.00 ± 4.58 |
| Stack Cube | 1 | 4.00 ± 5.29 | 94.33 ± 2.52 | 89.67 ± 7.02 | 95.00 ± 4.36 |
| Stack Cube | 3 | 6.33 ± 8.50 | 96.00 ± 2.00 | 93.00 ± 8.89 | 96.00 ± 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.
| Method | Time/step (ms) ↓ | Peak GPU memory (GiB) ↓ |
|---|---|---|
| w/o SG-RST + TVF (σ=0) | 267.12 ± 1.89 | 17.46 |
| WP-Preview (σ=1) | 266.94 ± 0.60 | 17.46 |
| RAP (J=2) | 315.04 ± 1.20 | 18.04 |
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.
| Task | RAP (1) | RAP (3) | WP (1) | WP (3) | A2A (6) | VITA (6) | FM-UNet (10) | DDIM-UNet (40) |
|---|---|---|---|---|---|---|---|---|
| Close Box | 90/88/90 | 98/90/88 | 90/88/88 | 90/92/88 | 90/90/98 | 80/74/72 | 62/70/70 | 54/70/64 |
| Pick Cube | 100/96/98 | 98/98/98 | 98/98/96 | 100/98/96 | 92/90/90 | 84/92/94 | 84/92/86 | 60/72/62 |
| Stack Cube | 98/94/92 | 94/96/98 | 84/10/64 | 96/80/88 | 84/96/86 | 68/78/80 | 24/20/24 | 28/40/36 |
| Pull Cube | 90/88/90 | 94/90/90 | 88/82/92 | 94/92/90 | 90/94/90 | 66/94/92 | 76/86/84 | 76/92/84 |
| Open Drawer | 92/80/86 | 88/86/82 | 84/80/82 | 88/76/76 | 96/84/86 | 82/80/78 | 52/42/42 | 64/62/68 |
| Pick-Place Bowl | 74/70/78 | 86/82/88 | 80/70/72 | 78/66/78 | 82/84/90 | 80/80/80 | 56/42/60 | 60/70/56 |
| Square | 80/80/86 | 84/88/86 | 70/80/80 | 38/64/54 | 68/56/54 | 16/38/40 | 38/28/14 | 24/2/0 |
| Tool Hang | 32/26/26 | 20/22/20 | 22/28/26 | 18/24/20 | 24/26/34 | 18/14/10 | 8/20/8 | 20/22/32 |
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.
| Method | Close Box | Pick Cube | Stack Cube |
|---|---|---|---|
| RAP (1) | 96/94/96 | 88/72/84 | 76/86/78 |
| RAP (3) | 100/98/98 | 90/90/88 | 84/90/80 |
| A2A (6) | 66/72/70 | 46/50/38 | 34/40/30 |
| VITA (6) | 78/68/76 | 50/64/80 | 44/50/42 |
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.
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.
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.