acceptodds
Under review as a conference paper at ICLR 2027

Don't Throw Away the Solution: Training-Free Acceleration of Flow-Based Robot Policies

Abstract

Standard samplers for flow-based robot policies generate each action chunk by sequential denoising and discard the solver trajectory at every replan, even though consecutive control cycles solve closely related problems. We introduce RePIC, a training-free sampler that turns this discarded trajectory into the initialization of the next solve. After an action prefix is executed, RePIC shifts the cached trajectory to the new horizon and re-aligns it to fresh noise with a closed-form correction that preserves the clean-action endpoint and needs no network evaluation. The pretrained velocity network then refines all transferred states under the new observation through batched fixed-point updates on the original integration grid, so an accurate initialization is converted into fewer sequential network calls. On Fast-WAM, RePIC reduces sequential calls per chunk from 10 to 2.2-2.6 and action-generation latency from 578 to 218ms on RoboTwin2.0 () and from 564 to 205ms on LIBERO () at comparable success; with system optimizations, RePIC-Flash reaches . The same sampler accelerates and GR00T N1.5 by and without retraining. Transferred drafts start closer to the serial solution on the carried action steps, and on a physical robot Fast-WAM with RePIC succeeds in versus for the original sampler. These results show that the numerical path of a previous solve is a reusable resource for fast robot control. Code will be made publicly available.

Primary area:
applications to robotics, autonomy, planning
Keywords:
flow matching, efficient inference, robot manipulation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_znUXOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes RePIC, a training-free inference acceleration method for flow-based robot policies. It reuses complete denoising trajectories across consecutive replans through noise-aligned trajectory transfer and parallel Picard refinement, reducing sequential network evaluations. Experiments demonstrate 2.66× and 2.75× speedups on RoboTwin 2.0 and LIBERO with comparable success rates, along with evaluations on other VLA models and a real robot.

Critical strengths
  1. Clear motivation and practical relevance. The paper addresses an important inference bottleneck in flow-based robot policies by exploiting computational redundancy across consecutive replans.
  2. Simple and effective design. RePIC integrates trajectory reuse with parallel refinement without requiring additional training or modifications to pretrained models.
  3. Comprehensive experimental evaluation. The paper includes multiple benchmarks, different policy architectures, real-world experiments, and ablation studies that provide evidence for the effectiveness of the proposed framework.
Critical weaknesses
  1. The novelty and contribution of the framework could be better clarified. ParaDiGMS has explored parallel Picard sampling, C3ache has investigated cross-chunk denoising reuse, and DVAC has studied adaptive execution horizons. In my view, the key contribution of RePIC is the observation that intermediate denoising states contain reusable information beyond final action predictions, improving subsequent parallel refinement. The authors should better highlight and substantiate this insight. (Please feel free to point out if my understanding is incorrect.)
  2. The necessity of full-trajectory reuse remains unclear. The current ablations mainly compare endpoint-only and full-trajectory reuse, without exploring intermediate alternatives. Reusing a sparse subset of denoising states might achieve comparable efficiency with lower storage overhead. Further analysis would help establish whether retaining the entire trajectory is necessary.
Questions for authors
  1. Initialization of newly exposed action positions. RePIC directly reuses the same-index denoising states from the previous replan, despite the lack of temporal correspondence. What motivates this choice? Have the authors considered alternatives such as fresh-noise initialization or trajectory extrapolation?
  2. Inconsistency in the noise-alignment ablation. Section 4.2 states that noise alignment has no effect without trajectory shifting due to the fixed noise seed. However, Table 9 reports different network calls (3.36 vs. 3.18) and latency (233.4 vs. 290.9 ms) for these theoretically equivalent initialization settings. Could the authors clarify this discrepancy?
Reviewer_Kx2YOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes RePIC, a training-free approach for accelerating inference in flow-based robot policies. The key idea is to exploit the temporal redundancy between consecutive replanning steps by reusing the full denoising trajectory from the previous solve. Specifically, RePIC shifts the cached trajectory to the current action horizon, aligns it with the new noise initialization through a closed-form correction, and employs parallel fixed-point refinement to reduce the number of sequential network evaluations.

The method is evaluated on Fast-WAM, π0.5, and GR00T N1.5 across LIBERO and RoboTwin 2.0, with additional real-robot experiments. The results demonstrate substantial reductions in action-generation latency while largely preserving task success rates. The paper further provides ablation studies to investigate the contributions of trajectory reuse, noise alignment, and parallel refinement.

Critical strengths

1. Clear motivation and effective trajectory reuse.
RePIC exploits the redundancy between consecutive replanning steps through noise-aligned trajectory reuse and parallel refinement. The ablations show that full-trajectory reuse outperforms endpoint-only initialization and significantly reduces sequential network calls, providing direct evidence for the effectiveness of the core idea.

2. Substantial speedup with largely preserved task performance.
RePIC achieves 2.75× acceleration on LIBERO and 2.25× on RoboTwin 2.0 under fixed execution lengths, while maintaining comparable success rates. The improvement also persists on an optimized backend, supporting its practical value beyond implementation-level optimizations.

Critical weaknesses

The comparison with equally optimized few-step samplers remains incomplete.
Table 3 shows that 5-step Euler achieves 96.90% success on LIBERO, close to RePIC-Flash (97.00%). However, Table 6 only evaluates the 10-step Euler baseline with system optimizations. It remains unclear whether RePIC offers a substantial latency–success advantage over simpler 3-step or 5-step Euler samplers with the same optimizations. This limits the assessment of its practical advantage, although it does not undermine the core findings.

Questions for authors

On LIBERO, Table 3 shows that 3-step and 5-step Euler sampling retain relatively high task success rates despite substantially reducing the number of sampling steps. In particular, the success rate of 5-step Euler (96.90%) is very close to that of RePIC-Flash (97.00%). Meanwhile, Table 6 demonstrates that system-level optimizations, including feature caching and CUDA Graph replay, substantially reduce the latency of the sequential Euler sampler, but the optimized comparison only includes 10-step Euler.

Could the authors report the inference latency and task success rates of 3-step and 5-step Euler when equipped with the same system-level optimizations as RePIC-Flash, under the same action execution length?

This comparison would help clarify the practical latency advantage of RePIC over simply reducing the number of sampling steps at comparable task success rates, and to what extent this advantage comes from preserving the original integration grid through trajectory reuse and parallel refinement.

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

This paper introduces RePIC, a training-free inference method for flow-based robot policies operating under receding-horizon control. Instead of discarding intermediate denoising states after generating an action chunk, RePIC caches the complete solver trajectory and transfers it to the next replan. The transfer shifts overlapping action positions and applies a noise-dependent correction that matches the new initial noise while preserving the transferred clean endpoint.

The transferred trajectory initializes a sliding-window parallel Picard solver. Batched velocity evaluations and residual-guided window advancement reduce sequential network invocations while retaining the original Euler integration grid. An optional mechanism selects the next execution length using residuals from the current solve. RePIC-Flash additionally applies feature caching and compiled inference optimizations.

Experiments on Fast-WAM report approximately (2.7\times) lower action-generation latency with similar aggregate success on RoboTwin 2.0 and LIBERO. Additional evaluations demonstrate smaller but positive speedups on (\pi_{0.5}) and GR00T N1.5, as well as preliminary real-robot validation. Ablations investigate parallelism, trajectory initialization, execution length, and system optimizations.

Critical strengths
  1. A meaningful controlled comparison supports the full-trajectory reuse contribution. On LIBERO, adding reuse to parallel refinement reduces latency from (298.10) to (204.77) ms. More specifically, full-trajectory reuse improves over endpoint-only initialization, reducing sequential calls from (2.88) to (2.22) and latency by (12.8%), with similar success. This directly supports the central claim that retaining intermediate solver states provides useful information beyond the previous action prediction.

  2. The acceleration remains useful under stronger controls and across policy transfers. Fixed-execution-length RePIC achieves approximately (2.25\times) acceleration on RoboTwin, while the matched compiled-backend comparison on LIBERO retains approximately (1.60\times) acceleration. These results show that the benefit does not depend entirely on adaptive execution length or comparison against an unoptimized implementation. Additional results on (\pi_{0.5}), GR00T N1.5, and a physical robot strengthen the practical significance.

Critical weaknesses
  1. The alignment ablation is not fully consistent with the stated protocol. Table 7 states that Fast-WAM regenerates its initial noise from the same seed at every replan. Without a horizon shift, Equation (7) should therefore leave the cached initialization unchanged, as the authors themselves explain. Nevertheless, Table 9 reports (3.36) versus (3.18) calls and (233.4) versus (290.9) ms for the unaligned and aligned no-shift variants. Table 4 also reports smaller differences. I need clarification on these comparisons, as they underpin the attribution of gains to alignment. Furthermore, repeated initial noise establishes a narrower empirical setting than independently resampled “fresh noise.”

  2. The headline speedups exceed what can be attributed specifically to full-trajectory reuse, and the implications of their deployment remain incompletely quantified. The (4.42\times) LIBERO Flash result combines sampler and system optimizations; the matched compiled comparison gives (1.60\times). The optimized RoboTwin serial baseline is padded to a batch size of 10, limiting its value for comparison with an optimized serial implementation. Adaptive execution length also changes the number of replans per executed action, so latency per chunk does not establish overall control-loop efficiency. These qualifications are largely disclosed, but peak memory, total state-wise evaluations, and aggregate inference cost per executed action are missing. They matter because the method trades additional batched computation for shorter sequential latency.

Questions for authors
  1. Can you reconcile the no-shift alignment ablation with the repeated-noise protocol? Are the initial noise tensors identical across replans? If so, what causes the differences between aligned and unaligned no-shift variants in Tables 4 and 9? Please identify any evaluation randomness or implementation differences beyond the stated toggle.

  2. What does adaptive commitment achieve after accounting for execution length? From the existing RoboTwin runs, please report the mean (R_n) and aggregate inference time per executed action, or equivalent totals. Does adaptive commitment improve this measure over fixed (R=24) RePIC?

  3. What resource costs accompany the latency gains? Are peak memory, total state-wise evaluations, and latency percentiles available from existing profiling? Was an unpadded batch-one TensorRT serial baseline measured? These details would materially clarify the practical advantage over optimized serial inference.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.