acceptodds
Under review as a conference paper at ICLR 2027

FlowLap: Execution-Aware Inference for Flow-Matching Video-Action World Models

Abstract

Flow-matching video-action models jointly generate visual futures and action chunks, but rollout latency can stall robot execution at chunk boundaries. Although asynchronous action-chunk methods such as Real-Time Chunking (RTC) can hide part of this latency, directly adapting RTC to flow-matching video-action models that commit real context or KV cache only at chunk boundaries can produce stale-context action candidates; VLASH, LingBot-VA 2.0, and DexWorldModel require model-specific training or pretraining to preserve success under overlap. We introduce FlowLap, an execution-aware inference framework with two components: a speculative joint rollout that overlaps inference with action execution while preserving the committed action prefix, and execution-state calibration that conditions the action branch on the state at which the candidate will run. On the 50-task RoboTwin benchmark with 150 episodes per configuration, FlowLap improves success over the asynchronous reference from 88.7% to 91.3% on FlowWAM and, with calibration, from 85.3 % to 92.0 % on LingBot-VA. Relative to matched reduced-work synchronous baselines, it reduces mean transition wait from 896.6 to 39.1 ms on FlowWAM and from 839.9 to 71.8 ms on LingBot-VA, with 0 ms p90 wait for both models.

Primary area:
applications to robotics, autonomy, planning
Keywords:
video-action world models, asynchronous inference, speculative execution, inference efficiency

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_N3rtOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes FlowLap, an execution-aware inference framework for flow-matching video-action world models. It overlaps speculative joint video-action generation with execution of a committed action prefix, then accepts the candidate suffix through timing, identity, and continuity checks or falls back to generation from refreshed real observations. Persistent context is updated only from real observations and executed controls. An optional residual adapter, trained with the backbone frozen, conditions the action branch on the estimated execution-time state. Across 50 RoboTwin tasks, FlowLap reduces mean transition wait from 896.6 to 39.1 ms on FlowWAM and from 839.9 to 232.4 ms on LingBot-VA. Calibration further reduces the latter to 71.8 ms while increasing success from 86.0% to 92.0%. These results demonstrate reduced boundary waiting rather than reduced model computation

Critical strengths
  1. Useful controls for the scheduling benefit.

The efficient synchronous baselines use the same reduced denoising budgets as FlowLap. The substantial additional wait reductions therefore support overlapping inference with execution rather than merely reducing denoising work. Results cover two independently implemented backbones.

  1. Explicit runtime safeguards and detailed reporting.

The framework specifies prefix ownership, real-context updates, and fresh-context fallback. The appendices provide complete task-level outcomes, candidate statistics, and increased model-side cost, making the benefits and limitations more transparent.

Critical weaknesses
  1. The gate’s contribution needs cleaner isolation.

RTC is described as differing from FlowLap only in the validity gate, yet the separate “pure overlap (no gate)” control performs substantially worse than RTC, 80.7% versus 88.7% on FlowWAM. RTC retains native continuation decisions, and FlowLap also projects actions before gating. These distinctions require explicit controls separating candidate rejection from action projection and other safeguards.

  1. Limited statistical support for success improvements.

Evaluation uses only three episodes per task under one fixed initialization seed. Uncalibrated FlowLap improves over RTC by only four episodes on FlowWAM and one on LingBot-VA, without uncertainty estimates or repeated evaluations. These results are promising but insufficient to establish reliable improvements in task success.

  1. Deployment benefits remain narrowly evaluated.

The gate is evaluated before the new boundary observation, so reduced waiting does not establish faster reactions to environmental changes. Experiments cover only clean, easy simulation without external disturbances, while reported model-side time increases by 118–335% over efficient synchronous controls. End-to-end completion time and disturbance-response measurements would better establish the practical value of this trade-off.

Questions for authors

What precisely distinguishes RTC from “pure overlap”? Can you isolate candidate rejection from action projection while keeping scheduling and other safeguards fixed?

Can you repeat calibration on/off with identical model-RNG and attention-backend settings, and include RTC with the same calibration?

Can you provide paired episode outcomes and repeated evaluation seeds with uncertainty estimates for the success-rate differences?

How do total episode time, observation-to-action latency, and success change under shorter execution windows or controlled external disturbances?

9 comments
AuthorsOpenReview ↗

Response to reviewer N3rt (1/3)

We thank the reviewer for the careful review and constructive suggestions. Below, we address each concern with experimental results and mechanism-level explanations.


Q1. What precisely distinguishes RTC from “pure overlap”?

reply: RTC retains native continuation decisions, executor-owned prefix handling, suffix completion, and real-observation boundary refresh. Pure overlap directly consumes an early-ready candidate. Both omit the FlowLap handoff-validity gate.


Q2. Can you isolate candidate rejection from action projection while keeping scheduling and other safeguards fixed?

reply: Candidate rejection and action projection address complementary handoff requirements, and separate controls expose their contributions on both backbones. Rejection checks whether a speculative candidate remains eligible for the current execution state and handoff; projection adjusts the replaceable suffix to satisfy boundary constraints while preserving the executor-owned prefix. We report projection-only and rejection-only controls alongside the complete protocol on the same 50-task suite, with three episodes per task. The table also includes calibrated FlowLap on LingBot-VA so that the effect of improving candidate quality can be assessed separately from the component comparisons. SR denotes the fraction of successfully completed episodes, and mean wait measures blocking at action-chunk transitions.

ModelConfigurationSRMean wait (ms)
FlowWAMProjection-only84.7%45.2
FlowWAMRejection-only84.0%25.2
FlowWAMFlowLap91.3%39.1
LingBot-VAProjection-only87.3%94.1
LingBot-VARejection-only88.7%98.3
LingBot-VAFlowLap86.0%232.4
LingBot-VAFlowLap + calibration92.0%71.8

The complete handoff protocol gives the highest success rate on FlowWAM. FlowLap reaches 91.3% SR; projection-only and rejection-only reach 84.7% and 84.0%, respectively, giving gains of 6.6 and 7.3 percentage points. Rejection removes stale or temporally invalid candidates, and projection aligns the unexecuted suffix with the committed boundary. Applying both operations gives the controller a continuation that satisfies the scheduling and boundary-consistency requirements together. The results support the contribution of the complete handoff protocol on this backbone.

Execution-state calibration improves the candidate supplied to the complete handoff protocol on LingBot-VA. With the full protocol retained, calibration raises SR from 86.0% to 92.0% and reduces mean wait from 232.4 to 71.8 ms, a 69.1% reduction. The resulting configuration also exceeds projection-only and rejection-only, which achieve 87.3% and 88.7% SR with 94.1 and 98.3 ms mean wait. These comparisons establish two distinct findings: the component controls measure the outcomes of suffix adjustment and candidate rejection, while the calibration comparison measures the benefit of conditioning actions on their anticipated execution state before the full gate evaluates commitment. The ablation section reports these controls and their distinct roles.

AuthorsOpenReview ↗

Response to reviewer N3rt (2/3)

Q3. Can you repeat calibration on/off with identical model-RNG and attention-backend settings, and include RTC with the same calibration?

reply: Execution-state calibration improves RTC directly, and FlowLap with the same calibration delivers an additional gain in both success and latency. We evaluated four configurations on the 50-task LingBot-VA suite, with three episodes per task: RTC and FlowLap, each with and without calibration. The configurations use the same execution window and budget of two video-generation steps and five action-generation steps. RTC and FlowLap retain prefix handling and real-context refresh; FlowLap additionally validates candidate commitment. Calibration conditions the action branch on the anticipated execution state through a residual adapter while keeping the backbone frozen.

ConfigurationSRMean wait (ms)
RTC85.3%264.1
RTC + calibration88.0%105.4
FlowLap86.0%232.4
FlowLap + calibration92.0%71.8

Calibration improves execution-state alignment under both scheduling protocols. Adding calibration to RTC raises SR from 85.3% to 88.0% and reduces mean wait from 264.1 to 105.4 ms, a 60.1% reduction. Adding calibration to FlowLap raises SR from 86.0% to 92.0% and reduces mean wait from 232.4 to 71.8 ms, a 69.1% reduction. These comparisons show that future-state conditioning improves the actions proposed for later execution under both RTC and FlowLap.

Validated commitment provides an additional benefit when calibration is present in both methods. With the same calibration, FlowLap reaches 92.0% SR and 71.8 ms mean wait, gaining 4.0 percentage points and reducing wait by 31.9% relative to calibrated RTC. The adapter improves actions for their anticipated execution state, and the gate checks whether the resulting continuation remains compatible when consumed. Together, the four configurations distinguish the benefit of future-state conditioning from the additional benefit of validated admission. The calibration ablation now includes the RTC-calibrated control.


Q4. Can you provide paired episode outcomes and repeated evaluation seeds with uncertainty estimates for the success-rate differences?

reply: Repeated evaluations show higher mean success for FlowLap than RTC on both backbones, with calibration bringing LingBot-VA to 94.00% mean SR. We repeated evaluation with seeds {1,2,4}, using 50 tasks and three episodes per task for each seed: 150 episodes per seed and 450 episodes per configuration. These are evaluation seeds for the reported policies. The table reports mean SR and sample standard deviation across the three evaluation-seed success rates.

ModelConfigurationMean SRSD
FlowWAMEfficient sync95.78%2.69%
FlowWAMRTC83.11%3.67%
FlowWAMFlowLap91.11%3.36%
LingBot-VAEfficient sync94.22%2.78%
LingBot-VARTC87.56%0.77%
LingBot-VAPAINT85.11%2.69%
LingBot-VASEAM90.00%0.67%
LingBot-VAFlowLap90.44%0.77%
LingBot-VAFlowLap + calibration94.00%2.40%

The seed-level outcomes show that the FlowLap–RTC success ordering persists in every repeated LingBot-VA evaluation. Each entry below is the number of successful episodes out of 150.

LingBot-VA configurationSeed 1Seed 2Seed 4
RTC132/150130/150132/150
FlowLap135/150137/150135/150
FlowLap + calibration144/150137/150142/150

The multi-seed results support a success advantage over RTC across two independently implemented backbones. On FlowWAM, FlowLap achieves 91.11% mean SR versus RTC's 83.11%, an 8.00-percentage-point gain. On LingBot-VA, it achieves 90.44% versus 87.56%, a 2.88-point gain, with both methods having a seed-level SD of 0.77 percentage points. FlowLap completes 3, 7, and 3 additional episodes relative to RTC across seeds {1,2,4}. Calibration further raises the mean to 94.00%, compared with efficient synchronous execution's 94.22%, while its seed-level outcomes exceed RTC in all three evaluations. The aggregate statistics and individual seed counts together show that the success improvement persists across repeated evaluations. The appendix and supplementary material provide the corresponding task-level result records.

AuthorsOpenReview ↗

Response to reviewer N3rt (3/3)

Q5. How do total episode time, observation-to-action latency, and success change under shorter execution windows or controlled external disturbances?

reply: Physical-robot results demonstrate lower execution-boundary stalls with improved task success, and the launch-timing ablation shows how available overlap time affects these gains. The SO-ARM101 evaluation uses a 5-DoF arm, a 1-DoF gripper, a top-view RGB camera, and an RTX 4090, with 50 episodes each for StackCube and PickNPlace. The measurements include real camera observations, actuator execution, communication, and host scheduling.

TaskConfigurationSRMean wait (ms)p50 wait (ms)p90 wait (ms)Acceptance
StackCubeRTC86%202.16330.562986.841--
StackCubeFlowLap88%222.33431.9241,204.9480.861
StackCubeFlowLap + calibration92%68.22134.17034.1700.973
PickNPlaceRTC96%292.44033.9101,652.645--
PickNPlaceFlowLap100%69.66932.67932.6790.833
PickNPlaceFlowLap + calibration100%36.84336.84336.8431.000

Transition wait measures the stall visible to the robot at an action-chunk handoff. It is computed as max⁡(0,tready−texec)\max(0,t_{\mathrm{ready}}-t_{\mathrm{exec}}), where texect_{\mathrm{exec}} is the end of the current execution interval and treadyt_{\mathrm{ready}} is when the next action becomes available to the executor. Mean wait summarizes average blocking, and p90 captures long transition stalls. Candidate acceptance is the fraction of speculative candidates accepted for execution. SR is the fraction of successfully completed episodes. The p50 wait is the median transition wait.

Calibrated FlowLap sharply reduces the transition stalls that accumulate during physical task execution. Relative to RTC, it raises StackCube SR from 86% to 92%, reduces mean wait by 66.3% to 68.221 ms, and reduces p90 wait by 96.5% to 34.170 ms. On PickNPlace, it raises SR from 96% to 100%, reduces mean wait by 87.4% to 36.843 ms, and reduces p90 wait by 97.8%. These host control-loop measurements quantify the boundary-wait component of completion time; the task success rates give the corresponding end-to-end outcomes.

High acceptance under real sensing and actuation shows that validated overlap remains effective on the robot. Uncalibrated FlowLap accepts 86.1% of candidates on StackCube and 83.3% on PickNPlace; calibration raises these rates to 97.3% and 100%. On StackCube, calibration simultaneously increases acceptance and reduces p90 wait from 1,204.948 to 34.170 ms. Conditioning on the anticipated execution state improves the continuation supplied to the gate, and real-context refresh anchors subsequent generation to physical observations. The controller preserves its rejection and fresh-context fallback rules throughout deployment.

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

The paper introduces FLOWLAP, a training-free inference framework designed to mitigate boundary-transition latency in flow-matching video-action models. To safely overlap computation with execution without model-specific retraining, FLOWLAP proposes two core components: (1) a speculative joint rollout that runs concurrently with physical execution, guarded by a strict handoff validity gate to reject stale or discontinuous actions, and (2) an optional execution-state calibration that uses zero-initialized residual adapters to condition the action branch on the anticipated future state. The method is evaluated on the RoboTwin benchmark (FlowWAM, LingBot-VA), FLOWLAP drastically reduces transition wait times (e.g., from 896.6 ms to 39.1 ms on FlowWAM) while maintaining task success.

Critical strengths

Reviewer found the following strengths:

Method novelties:

  • The speculative joint rollout with prefix masking is a sound way to ensure prefix immutability while allowing suffix adaptation; the gate that checks identity, timing, shape, and boundary continuity is a reasonable, model-agnostic safety valve.

  • The framework adapts asynchronous scheduling for models that generate both video and actions, moving beyond older methods that were restricted to action-only models. A strict validity gate is also an interesting point that might help ensure these actions are safe to execute, and because the method requires no retraining of the original model, it can be plugged directly into existing systems straight off the shelf.

Experiment Validations:

  • The paper provides an evaluation spanning 50 tasks and 150 episodes per configuration on a standard bimanual benchmark (RoboTwin 2.0), using two distinct VA models (FlowWAM, LingBot-VA), increasing the claim of generality. Important ablation studies are provided such as the roles of overlap, handoff validation, launch timing, and calibration, with consistent trends.

  • Regarding efficiency, the method successfully pushes p90 transition wait times to near-zero. Crucially, it measures this via rigorous host control-loop metrics, avoiding misleading kernel-level micro-optimizations.

Critical weaknesses
  • Missing discussion/Comparing to Important Related Works: The author omits highly relevant concurrent works addressing action-chunk boundary continuity and asynchronous execution in generative flow policies. Specifically, the paper should discuss and compare (predered) against PAINT [1], which enforces smooth asynchronous handoffs via initial noise selection, and SEAM [2], which targets cross-chunk consistency using unexecuted tail references.

  • Simulation-Only Validation: The evaluation relies entirely on a clean, simulated environment (RoboTwin 2.0). For a paper focused on real-time speed, the lack of physical hardware testing is a major flaw. First, the impressive "0.0 ms" delay times might just be a simulation artifact that ignores real-world operating system and hardware lag. Second, normal sensor noise on a physical robot could trigger the strict safety gate too often, forcing the system to frequently fall back to slower methods. Finally, its ability to recover from unexpected physical disturbances remains completely untested.

  • (About presentaiton) Lack of Visual Intuition in Methodology Diagrams: Figure 2 functions purely as a high-level scheduling timeline. It illustrates when execution and inference overlap but fails to provide intuition on how the two core modules: the speculative joint rollout (e.g., prefix-masking mechanics) and execution-state calibration (e.g., residual adapters), actually work under the hood.

  • Lipschitz Assumption in Theoretical Result: The theoretical justification for the execution-state calibration error bound (Theorem A.10) relies heavily on Assumption A.7, which posits that the ideal action target is Lipschitz continuous with respect to execution-state features. In multimodal generative flow-matching, action spaces are rarely globally Lipschitz; a small change in state can cause the optimal trajectory to jump discontinuously to a different topological mode. A discussion on this factor is required.

[1] Start Right, Arrive Right: Asynchronous Execution via Initial Noise Selection, CoRL 2026

[2] SEAM: Smooth Execution of Action-Chunked Motion for Vision-Language-Action Policies,

Questions for authors
  • Can the author provide comparisons to methods such as PAINT or SEAM and provide results in a simulation? Note that methods such as PAINT are ready to use for VLA. In this regard, did you try adding a minimal handoff consistency check to RTC (e.g., identity/timing only) and/or tuning an RTC-specific boundary budget? How sensitive are RTC results to such design choices?

  • Figure 2 currently only conveys timeline scheduling. Could you update this figure (if possible) or add an architectural diagram) to visually explain the internal mechanics of the speculative joint rollout (masking) and the execution-state calibration (residual adapter injection)?

  • Given that flow-matching models frequently capture multimodal distributions where minor state perturbations lead to discontinuous topological choices, does this error bound only hold locally within a single action mode? Also, what precisely constitutes the “uncertainty gate” (feature(s) used, calibration, threshold 0.35 rationale)? How does it interact with the continuity checks when they disagree?

  • Do you have any preliminary on-hardware tests (even small-scale) indicating that the latency and SR trends survive sensor/actuator delays and OS-level contention?

5 comments
AuthorsOpenReview ↗

Response to reviewer DTRr (1/5)

We thank the reviewer for the careful review and constructive suggestions. Below, we address each concern with experimental results and mechanism-level explanations.


Q1. Can the author provide comparisons to methods such as PAINT or SEAM and provide results in a simulation? Note that methods such as PAINT are ready to use for VLA.

reply: FlowLap combines joint video-action overlap with executor-prefix preservation and validated real-context handoff, yielding lower transition wait than PAINT and SEAM; execution-state calibration further improves both success and latency. We evaluated the methods on LingBot-VA using the same 50 tasks, three episodes per task, and a budget of two video-generation steps and five action-generation steps. Success rate (SR) measures task completion, and mean transition wait measures the blocking time between the end of the current action chunk and availability of the next action.

ConfigurationSRMean wait (ms)
FlowLap86.0%232.4
PAINT82.7%252.1
SEAM89.3%987.8
FlowLap + calibration92.0%71.8

FlowLap preserves execution consistency throughout generation and at commitment. PAINT selects the initial noise so that the generated chunk agrees with its predecessor at the boundary. This acts on the starting point of the generative trajectory. FlowLap fixes executor-owned controls at every generation step through prefix masking, shifts the completed candidate to expose only its unexecuted suffix, and validates that suffix using the committed boundary control, timing, and context information available during the execution window. These mechanisms address distinct stages of asynchronous execution: prefix masking preserves controls already under execution, and the gate checks whether the resulting continuation remains admissible when it is consumed. The latest real observation advances the persistent context, and rejected candidates trigger fresh-context regeneration. This directly addresses the state and timing changes that occur between candidate launch and use. In the simulation, FlowLap achieves 3.3 percentage points higher SR than PAINT with lower mean wait; calibrated FlowLap increases that gain to 9.3 points and reduces mean wait by 71.5%.

FlowLap hides joint video-action computation inside physical execution while maintaining a validated continuation. SEAM uses the unexecuted tail of the preceding chunk as a reference to smooth the next chunk. That reference supplies a continuity constraint; the amount of computation completed before the boundary determines the remaining wait. In the evaluated SEAM configuration, joint rollout remains on the boundary critical path. FlowLap starts the joint rollout during execution, preserves the committed prefix, and admits the ready suffix through its validity gate. This is especially relevant to video-action models, whose joint rollout includes both visual and action generation. FlowLap reduces mean wait from SEAM's 987.8 ms to 232.4 ms, a 76.5% reduction. With calibration, FlowLap reaches 92.0% SR versus SEAM's 89.3% and reduces mean wait by 92.7%. Calibration conditions the action branch on the anticipated execution state, improving candidate compatibility before the same gate evaluates admission. These comparisons and mechanism distinctions are included in Related Work and the end-to-end evaluation.

AuthorsOpenReview ↗

Response to reviewer DTRr (2/5)

Q2. In this regard, did you try adding a minimal handoff consistency check to RTC (e.g., identity/timing only) and/or tuning an RTC-specific boundary budget? How sensitive are RTC results to such design choices?

reply: FlowLap's handoff gate validates execution compatibility in addition to identity and readiness, and the component and calibration controls show how these mechanisms affect success and wait. RTC and FlowLap share the joint rollout, prefix handling, and real-context handoff; FlowLap additionally validates candidate commitment. We report the RTC reference, projection-only and rejection-only controls, and both methods with execution-state calibration on the same 50-task LingBot-VA suite, with three episodes per task. Projection-only evaluates suffix adjustment, rejection-only evaluates candidate rejection, and the calibrated RTC comparison evaluates the benefit of validated admission when both methods use future-state conditioning.

LingBot-VA configurationSRMean wait (ms)
RTC85.3%264.1
Projection-only87.3%94.1
Rejection-only88.7%98.3
FlowLap86.0%232.4
RTC + execution-state calibration88.0%105.4
FlowLap + calibration92.0%71.8

Identity/timing checks and the boundary budget govern scheduling eligibility; continuity and uncertainty checks govern execution compatibility. Identity checks bind a result to its episode and target handoff, and timing checks enforce readiness and temporal alignment. The boundary budget sets the deadline by which a candidate must be ready for admission. A candidate that misses this deadline follows the fresh-context fallback path. Continuity checks assess agreement with the committed boundary control, while uncertainty checks assess mismatch using the available context and execution signals. FlowLap requires these conditions to pass together. The component results above quantify the measured success and wait under different admission designs; the budget specifies when those admission decisions remain eligible.

Calibration and validated admission together deliver the strongest measured success–latency combination among the reported controls. Projection-only and rejection-only reach 87.3% and 88.7% SR with 94.1 and 98.3 ms mean wait, respectively. Calibrated FlowLap reaches 92.0% SR and 71.8 ms mean wait, exceeding both partial-gate controls on both metrics. The calibration comparison further separates candidate conditioning from commitment: adding calibration to RTC raises SR from 85.3% to 88.0% and reduces mean wait from 264.1 to 105.4 ms. With the same calibration, FlowLap gains another 4.0 percentage points in SR and reduces mean wait by 31.9%. Future-state conditioning improves the candidate supplied to the gate, and validation governs its commitment. The ablation section now reports these controls and explains the roles of scheduling eligibility, continuity, and context validity.

AuthorsOpenReview ↗

Response to reviewer DTRr (3/5)

Q3. Figure 2 currently only conveys timeline scheduling. Could you update this figure (if possible) or add an architectural diagram to visually explain the internal mechanics of the speculative joint rollout (masking) and the execution-state calibration (residual adapter injection)?

reply: Prefix preservation protects controls already under execution, handoff validation determines whether the remaining suffix can be committed, and calibration conditions that suffix on its anticipated execution state. The added architectural diagram makes these three mechanisms explicit through separate validation and calibration paths.

The speculative rollout preserves the executor-owned prefix and produces a validated continuation. Its inputs are the real-observation context and the controls already committed to the executor. At every generation step, a binary mask restores the committed controls in the protected prefix while allowing the remaining actions to evolve. The completed candidate is shifted to expose its unexecuted suffix; any projection is restricted to this replaceable suffix. Identity, timing, shape/finite, continuity, and uncertainty checks then determine admission using information available during the execution window. A passing candidate supplies the next action continuation. A failing candidate triggers regeneration from refreshed real context. Only real observations and executed controls advance persistent context.

Execution-state calibration corrects the temporal mismatch between candidate generation and consumption. Its inputs are the launch-time robot state and the committed controls expected to execute before the candidate is consumed. A state propagator estimates the future execution state, which is converted into conditioning features for a zero-initialized residual adapter attached to the action branch of the frozen backbone. This conditions candidate actions on where the robot is expected to be when they execute. The visual branch and persistent context remain conditioned on real observations, and the resulting action candidate passes through the same handoff checks. The added diagram, placed after the scheduling description, traces these inputs, transformations, and outputs.

Reviewer_TY4iOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims at finding a high-efficiency inference framework of flow-matching video-action world models. The authors propose FlowLap which introduces speculative joint rollout and execution-state calibration to improve the success rate as well as the transition wait. The authors validated their method on RoboTwin2.0 Benchmark.

Critical strengths
  1. The authors propose two key contributions: One is Speculative Joint Rollout, which introduces an inference-level scheduling protocol that rehosts external scheduling on the joint video-action rollout and overlaps generation with robot physical execution, eliminating boundary-transition latency while validating candidates at each handoff. The other is Execution-State Calibration, with a lightweight calibration component which estimates the anticipated robot state at candidate execution time and injects future-state conditions exclusively into the action branch via zero-initialized residual adapters. It further mitigates temporal mismatch between speculative generation time and future action execution time under asynchronous rollout.
  2. The author evaluated their methods and do a series of ablation study on Robotwin 2.0 benchmark to verify the advantages.
Critical weaknesses
  1. Can the authors clarify why they didn’t conduct comparisons with FAST-WAM (Yuan etal.), FASTER-WAM (Ma, etal and Zhao etal.) and why they didn't do experiments on other benchmarks besides Robotwin?
  2. Can the authors tell whether they did any real-world deployment of their method on the real robot tasks?
Questions for authors
  1. In Table 3, why is the LingBot-VA SR of ‘’FLOWLAP + execution-state calibration’ bolded, since it is not the best SR in the column?
  2. In Line 130, ‘Averaged over FlowWAM FlowLap and calibrated LingBot-VA FlowLap, success improves ..’, the ‘success’ means ‘the success rate’? This sentence is a bit ambiguous.
2 comments
AuthorsOpenReview ↗

Response to reviewer TY4i (1/2)

We thank the reviewer for the careful review and constructive suggestions. Below, we address each concern with experimental results and mechanism-level explanations.


Q1. Can the authors clarify why they didn’t conduct comparisons with FAST-WAM (Yuan etal.), FASTER-WAM (Ma, etal and Zhao etal.)?

reply: FAST-WAM and FASTER-WAM optimize model-level inference speed, while FlowLap converts an existing video-action model to execution-overlapped inference; these optimization dimensions are orthogonal. FAST-WAM and FASTER-WAM improve the architecture or computation of the model to reduce the time required to produce a prediction. FlowLap operates on the execution schedule of an available backbone: it launches the next joint video-action rollout while the robot executes the committed action prefix, preserves that prefix during generation, and validates the candidate suffix before handoff. Its core scheduling protocol applies with the pretrained backbone frozen and requires no model retraining. The distinction follows directly from the latency decomposition: model acceleration reduces the generation time TgenT_{\mathrm{gen}}, and overlap uses the execution interval TexecT_{\mathrm{exec}} to cover that computation, giving an ideal residual wait of max⁡(0,Tgen−Texec)\max(0,T_{\mathrm{gen}}-T_{\mathrm{exec}}) for an admissible candidate. This decomposition identifies two separate optimization targets: the duration of model computation and the portion of that computation exposed as boundary wait. FAST-WAM's standalone inference speed therefore does not directly measure FlowLap's scheduling contribution. Our experiments measure that scheduling contribution on FlowWAM and LingBot-VA with the backbone and inference-step budget held fixed. Related Work now discusses FAST-WAM and FASTER-WAM and makes this orthogonal relationship explicit.


Q2. And why they didn't do experiments on other benchmarks besides Robotwin?

reply: RoboTwin provides a standardized task suite for isolating FlowLap's scheduling contribution, while evaluations on two backbones and a physical robot test its architectural and deployment relevance. RoboTwin 2.0 offers 50 bimanual tasks with a native task-completion signal, allowing success and transition wait to be measured consistently across scheduling methods. For an execution-overlap method, both outcomes matter: reduced waiting establishes the scheduling benefit, and task completion checks whether the generated continuations remain useful for control. FlowWAM and LingBot-VA provide two independently implemented joint video-action backbones, allowing the same execution protocol to be evaluated across architectures. Comparisons with efficient synchronous execution hold the backbone and reduced inference-step budget fixed, isolating the contribution of overlapping generation with execution.

Physical StackCube and PickNPlace evaluations establish that the scheduling benefit carries into real robot execution. On SO-ARM101, calibrated FlowLap reaches 92% and 100% SR, with mean transition waits of 68.221 and 36.843 ms, respectively. StackCube exercises contact alignment, and PickNPlace exercises object transport. These tasks test candidate handoff under physical sensing, actuation, communication, and host scheduling. The evaluation therefore combines standardized task coverage, replication across two video-action backbones, and measured physical deployment outcomes. The Real-World Experiments section now reports the setup and results.

AuthorsOpenReview ↗

Response to reviewer TY4i (2/2)

Q3. Can the authors tell whether they did any real-world deployment of their method on the real robot tasks?

reply: Real-robot deployment confirms that calibrated FlowLap improves task success and sharply reduces transition stalls, reaching 92–100% SR and 34–37 ms p90 wait. On SO-ARM101, equipped with a 5-DoF arm, a 1-DoF gripper, a top-view RGB camera, and an RTX 4090, we ran 50 episodes each for StackCube and PickNPlace.

TaskConfigurationSRMean wait (ms)p90 wait (ms)Candidate acceptance
StackCubeRTC86%202.163986.841--
StackCubeFlowLap88%222.3341,204.9480.861
StackCubeFlowLap + calibration92%68.22134.1700.973
PickNPlaceRTC96%292.4401,652.645--
PickNPlaceFlowLap100%69.66932.6790.833
PickNPlaceFlowLap + calibration100%36.84336.8431.000

FlowLap sustains frequent validated handoffs under physical execution. Uncalibrated FlowLap accepts 86.1% of StackCube candidates and 83.3% of PickNPlace candidates, while raising SR over RTC by 2 and 4 percentage points. On PickNPlace, it also reduces mean wait from 292.440 to 69.669 ms, a 76.2% reduction. These results show that asynchronous candidates remain usable under real camera observations, actuation, communication, and host scheduling. Prefix preservation maintains committed controls, and real-context refresh and candidate rejection govern subsequent handoffs.

Execution-state calibration strengthens both candidate compatibility and latency reduction. On StackCube, calibration raises acceptance from 86.1% to 97.3% and SR from 88% to 92%, while lowering mean wait from 222.334 to 68.221 ms. Against RTC, calibrated FlowLap gains 6 percentage points in SR and reduces mean and p90 wait by 66.3% and 96.5%. On PickNPlace, it reaches 100% SR and acceptance, with mean and p90 waits of 36.843 ms; these correspond to 87.4% and 97.8% reductions relative to RTC. Conditioning the action branch on the anticipated execution state improves the continuation evaluated by the same gate, connecting higher candidate acceptance with fewer long transition stalls. The real-world section now includes the measurements and their mechanism-based interpretation.


Q4. In Table 3, why is the LingBot-VA SR of ‘’FLOWLAP + execution-state calibration’ bolded, since it is not the best SR in the column?

reply: We have removed the incorrect boldface from the LingBot-VA calibrated FlowLap SR entry in Table 3.


Q5. In Line 130, ‘Averaged over FlowWAM FlowLap and calibrated LingBot-VA FlowLap, success improves ..’, the ‘success’ means ‘the success rate’? This sentence is a bit ambiguous.

reply: “Success” means success rate (SR), the fraction of successfully completed episodes. We have revised the ambiguous wording in the original paper to state “success rate” explicitly.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Essay changes

The related-work section now positions FAST-WAM and FASTER-WAM. It also adds PAINT and SEAM references and matched comparisons.

The end-to-end evaluation now includes LingBot-VA results for PAINT, SEAM, FlowLap, and calibrated FlowLap under the shared protocol.

The scheduling section now explains prefix ownership, suffix projection, real-context refresh, handoff checks, commit, and fresh-context fallback. It also describes the projection-only, rejection-only, and RTC-plus-calibration controls for FlowWAM and LingBot-VA.

The calibration discussion now specifies the uncertainty signals, thresholds, threshold-selection protocol, fail-closed behavior, and continuity checks. It also states the local mode-conditioned calibration guarantee.

The essay now reports repeated-seed mean success rates and sample standard deviations for Efficient sync, RTC, PAINT, SEAM, FlowLap, and calibrated FlowLap where multi-seed runs are available. Complete task-level counts are included in the appendix. The new configuration and seed experiment data have also been updated in the supplementary material.

The physical-evaluation section now documents the SO-ARM101 setup, hardware, 50 episodes per task, StackCube and PickNPlace results, wait distributions, candidate-acceptance rates, and success rates under real observations, actuator execution, communication, and host scheduling.

Terminology and tables now define SR consistently as success rate and correct the calibrated FlowLap emphasis in Table 3. Every essay modification described here has been marked in red in the original essay.

Dear Reviewer,

Thank you again for your valuable comments on our manuscript. We have submitted our revised version together with a detailed point-by-point response, and have highlighted the changes in the manuscript for your convenience.

We would greatly appreciate it if you could take some time to review our responses at your earliest convenience. Should you have any remaining questions or concerns, we will be more than happy to address them.

Thank you for your time and consideration.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.