acceptodds
Under review as a conference paper at ICLR 2027

BATON: Passing Denoising Trajectories Across Replans for Parallel Sampling of Robot Policies

Abstract

Diffusion and flow policies built on large action transformers generate each action chunk through many narrow, dependent model calls. We present BATON, a training-free sampler that evaluates the sampling steps of a frozen receding-horizon policy in parallel. BATON is based on two designs. First, it transports the previous denoising trajectory into the next replan by shifting the action horizon, completing the tail, and restoring the current noise boundary. Second, it starts two parallel Picard rounds over the full ten-step schedule from this trajectory and combines them with a depth-one Anderson extrapolation that adds no model calls. Ten dependent transformer calls thus become two wider ones. Without assuming contraction, we bound the error of the raw parallel rounds linearly in the initializer error, which is bounded by the change of the policy's plan between replans and the drift of its clean predictions, and explain the Anderson step as averaging an oscillation of the raw rounds. On 450 paired LIBERO-Plus instances at a fixed execution stride, BATON reduces Fast-WAM's median action-expert latency by 2.80× and backend latency by 1.73×, and solves 311 instances versus 304 for serial sampling, a difference that is not statistically significant. Paired ablations show that removing either transport or Anderson extrapolation lowers success. The speedups carry over to Fast-WAM on RoboTwin 2.0 (2.74× action expert, 1.68× backend), and with π0.5 on LIBERO-Plus the compiled full policy call becomes 1.30× faster, again without a significant change in success.

Primary area:
applications to robotics, autonomy, planning
Keywords:
diffusion policy, flow matching, inference acceleration, parallel sampling, receding-horizon control, real-time robot inference, Picard iteration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_ExGMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The core idea behind this paper is pretty straight forward. When using diffusion or flow-based planners, the entire denoising trajectory contains information that is useful for replanning at subsequent time steps.

The authors argue that reusing this history allows a new plan to be generated with fewer sequentially dependent model calls (denoising steps), resulting in faster inference while retaining the original denoising schedule.

BATON reuses the entire denoising trajectory from the previous planning step. It transports this trajectory by shifting the action horizon, repeating the last action to complete the tail, and restoring the current noise boundary according to the denoising schedule. It then performs two successive parallel-denoising rounds, each using one batched model call, and combines the two refined trajectory estimates using Anderson mixing.

Critical strengths
  1. The method appears straightforward to implement within compatible diffusion- and flow-based planners.

  2. It is training-free and empirically demonstrates substantial inference speedups over replanning from scratch and the observed speedups are roughly in the ballpark for reduced planning steps of a diffusion based planner.

  3. The evaluation covers several policy–benchmark combinations that I consider relevant, including π0.5\pi_{0.5} and Fast-WAM on LIBERO-Plus and RoboTwin 2.0.

Critical weaknesses
  1. Novelty : Based on my understanding, novelty is limited. The main contribution is the transport-to-replanning pipeline. While reasonable the contribution is narrow because most of the underlying components already exist . To be more specific, many of the elements (and not limited to the following) of the proposed method are available in different prior works based on my reading.

ParaDiGMS[1] introduced Picard-parallel denoising and its application to receding-horizon Diffusion Policy. ParaTAA initializes parallel sampling from previous correlated conditioning input’s denoising steps [2], with anderson mixing. RTI-DP’s main contribution as far as I understand is it shifts previous action predictions and repeats the final action for subsequent refinement [3].

My concern is that many of its distinct algorithmic ingredients exist. The contribution is mainly the particular cross-replan full-trajectory transport and its integration into a fixed two-round solver.

  1. Clarity and presentation. The paper is very difficult to read in its current form because of its heavy reliance on jargon and paper-specific terminology. A more but not limited to accessible abstract, introduction, and preliminaries would help communicate the relatively straightforward idea it presents. I strongly recommend a substantial rewrite to explain the central intuition before introducing the detailed notation and analysis.

  2. Limited guarantees for error accumulation. I could not find any theory that would support that approximation error remains uniformly bounded over long episodes. The cross-replan in Appendix C.4 provides a long-run bound iff cR<1c_R < 1 and the transport mismatch remains bounded. However, I could not find in the paper if the authors establish that the former condition holds for the evaluated policies.

References:

[1] Parallel Sampling of Diffusion Models, shih et al

[2] Accelerating Parallel Sampling of Diffusion Models, Tang et al

[3] Real-Time Iteration Scheme for Diffusion Policy, Duan et al

Questions for authors
  1. Error accumulation over long episodes. How does approximation error evolve over many consecutive replans when each new plan is initialized from BATON's own cached approximate trajectory? Appendix D.4 includes LIBERO-Long tasks, but that diagnostic uses serial-generated history and therefore does not measure the accumulation of errors from repeated approximate replanning. Could the authors provide a long-horizon closed-loop evaluation that measures error as a function of the number of replans? This would help clarify the potential tradeoff between inference speed and error accumulation.

  2. Choice of Anderson mixing. The oscillation analysis motivates mixing successive iterates, but is an adaptive mixing needed? Would a constant mixing ratio achieve comparable performance? A fixed-weight mixing ablation would help distinguish the benefit of combining the two estimates from the benefit of selecting the coefficient adaptively. I am asking this since the anderson mixing is posed as another contribution of this work and to what extent does it make a difference?

  3. Conditions for the recursive bound. Can the authors establish, or provide empirical evidence, that the coefficient cRc_R in Appendix C.4 remains below 1 across successive replans?

  4. Utility of denoising history Do we really need to remember every intermediate denoising state? Please compare BATON against keeping only the previous final action chunk, shifting it, and constructing noisy versions to initialize the same two parallel rounds and Anderson mixing.

Reviewer_eAXvOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes BATON, a training-free method for accelerating robot policy inference, aiming to reduce the inference latency caused by multi-step sequential denoising in diffusion and flow-matching policies. The method reuses the complete denoising trajectory from the previous replan for initialization and combines two rounds of parallel Picard iteration with Anderson extrapolation to replace the conventional ten sequential denoising steps with two parallel rounds. The authors also provide a theoretical analysis of the method's approximation error. Experiments are conducted with Fast-WAM and π0.5 on the LIBERO-Plus and RoboTwin 2.0 simulation benchmarks.

Critical strengths

The method demonstrates a degree of novelty. By exploiting the temporal overlap between consecutive replanning steps, the paper reuses the complete denoising trajectory from the previous plan and combines parallel Picard iterations with Anderson extrapolation to accelerate inference without additional training or modifications to the model parameters. Ablation studies further validate the effectiveness of the two key components.

Critical weaknesses
  1. Strong dependence on Action Chunk overlap and limited validation under low-overlap settings. BATON relies on reusing the complete denoising trajectory from the previous replan. The main experiments use H=32 and S=10, resulting in approximately 68.75% overlap between consecutive action chunks. Although the authors also evaluate S=5 and S=20, the method has not been tested when S approaches H and the overlap becomes very small or disappears entirely. In such cases, the historical trajectory may provide less useful initialization information, potentially affecting task performance. Therefore, the applicability of BATON under low-overlap execution settings remains insufficiently validated.

  2. Unclear whether per-chunk inference acceleration translates into improved overall task execution efficiency. The paper mainly evaluates acceleration by comparing the inference latency of individual action chunks at a fixed execution stride. However, faster inference per chunk does not necessarily lead to higher control frequency or shorter task completion time. Under the main experimental setting, the original serial method already completes inference well within the replanning time window. Although the authors provide exploratory results for different execution strides, the actual benefits in terms of cumulative inference cost, task completion time, and achieved control frequency remain unclear.

  3. Lack of real-world robotic experiments. All experiments are conducted in simulation, where physics is paused during model inference. Although the results demonstrate that BATON reduces inference latency, they do not establish whether the method can improve actual control frequency, task completion time, or closed-loop execution performance when inference and robot control operate in real time. Given that the paper focuses on accelerating robot policy inference, the absence of real-time deployment validation limits the demonstrated practical value of the proposed method.

Questions for authors
  1. Regarding the impact of Action Chunk overlap: BATON's core design relies on the overlap between consecutive action chunks. Although the authors evaluate S=5, 10, and 20, could they further investigate the performance when S approaches H (e.g., S=28 or S=32)? In particular, how do the quality of the transported initialization and task success rate change when little or no overlap remains?

  2. Regarding practical inference efficiency: The paper mainly reports inference acceleration per action chunk, while different execution strides can change the number of replanning steps required to complete a task. Could the authors further evaluate cumulative inference time, task completion time, and achieved control frequency under different execution strides to clarify whether the per-chunk acceleration translates into improved overall execution efficiency?

  3. Regarding real-world robot deployment: All experiments are conducted in simulation. Could the authors provide additional real-world robot experiments to further validate the practical effectiveness of BATON?

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

This paper proposes BATON, a training-free method for accelerating diffusion- and flow-based robot policies. It reuses the denoising trajectory from the previous replan to initialize two parallel Picard iterations, followed by Anderson extrapolation. This replaces ten sequential denoising calls with two batched calls. Experiments on Fast-WAM and pi0.5 show reduced inference latency with comparable task success. The paper also provides ablation studies and a finite-round error analysis.

Critical strengths
  1. The paper addresses a practical bottleneck in robot policy inference. Reducing sequential denoising calls without retraining or modifying the policy is useful, and the reported speedups are substantial.

  2. Reusing the full denoising trajectory across replans is a reasonable way to exploit temporal overlap in receding-horizon control. The method is relatively simple and can be applied to existing policies.

  3. The theoretical analysis provides useful insights into finite-round parallel sampling, although it does not fully characterize the deployed algorithm.

Critical weaknesses
  1. The algorithmic novelty is somewhat limited. Parallel denoising, Anderson acceleration, and trajectory-based initialization have been explored in ParaDiGMS and ParaTAA, while RTI-DP and FALCON investigate historical action reuse. BATON mainly combines these ideas with a transport operation tailored to receding-horizon control. The distinctions from these works are not sufficiently discussed in the main text, with most related work deferred to the appendix.

  2. The benefit of the complete design over simpler alternatives is not fully convincing. In Appendix Table 6, directly copying the previous trajectory and running two raw Picard rounds achieves 306 out of 450 successes, matching the serial baseline, while BATON achieves 311. Both approaches require two batched model calls. Although their control-step counts differ, the additional benefit of transport and Anderson over this simpler baseline remains unclear.

  3. The numerical results do not consistently agree with closed-loop performance. Figure 3 shows that transport substantially reduces approximation error, yet transport without Anderson decreases success from 306 to 264 out of 450 compared with unaligned copying. Anderson improves the transported variant but hurts the unaligned one. The paper does not adequately explain this behavior, which raises questions about the motivation for combining these components.

  4. The theoretical guarantees concern raw Picard iterations rather than the Anderson-extrapolated trajectories actually cached and executed. Appendix C.6 also reports promising approximation errors from a single clean-form iteration, but its closed-loop performance is not evaluated. A stronger justification for the chosen two-round solver would be helpful.

  5. The practical real-time benefit remains uncertain. Since simulation physics is paused during inference, the experiments demonstrate latency reduction rather than improved feedback frequency. Both samplers already meet the nominal replanning deadline at the main execution stride.

Questions for authors
  1. Could the authors clarify the main algorithmic differences from ParaTAA and existing trajectory-reuse methods? It would be helpful to discuss these distinctions more explicitly in the main paper.

  2. How does BATON compare directly with unaligned trajectory copying and two raw Picard rounds in terms of paired success, execution steps, and task completion time? Given their similar success counts and identical number of batched calls, what justifies the additional components?

  3. Why does transport significantly improve approximation accuracy but reduce closed-loop success without Anderson? What explains the different effects of Anderson on transported and unaligned trajectories?

  4. Could the authors evaluate the single-round clean-form alternative from Appendix C.6 in closed-loop experiments? How stable are the Anderson-extrapolated trajectories when repeatedly reused across replans?

  5. Can the authors demonstrate improved feedback frequency or task performance under asynchronous execution, where inference latency affects the robot's control loop?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.