acceptodds
Under review as a conference paper at ICLR 2027

SEAM: Smooth Execution of Action-Chunked Motion for Vision-Language-Action Policies

Abstract

Vision-Language-Action (VLA) policies generate locally coherent action chunks, yet consecutive predictions can disagree in their overlapping future actions, introducing abrupt changes at execution boundaries. We propose SEAM (Smooth Execution of Action-chunked Motion), a training-free inference method for flow matching VLAs. SEAM uses the previous chunk's unexecuted tail as a temporally aligned consistency reference. Its core mechanism, Velocity-guided Loss Steering (VLS), alternates learned velocity-field steps with closed-form corrections toward a time-dependent mean reference. The corrected state feeds subsequent velocity evaluations, coupling cross-chunk consistency with observation-conditioned generation without policy-network backpropagation. In a 1,300-episode comparison on LIBERO-10 with π₀.₅, SEAM reduces boundary action second differences by 28% and chunk transition discontinuity by 27% at near-baseline denoising-loop cost. Matched-state controls show that the aligned reference improves the continuity–success trade-off over zero or shuffled references. With SmolVLA on LIBERO-Spatial, the same guidance setting reaches 72.0% success versus 65.6% for baseline and 60.0% for RTC, with respective realized end-effector jerk reductions of 58.8% and 40.0%, at approximately 1% additional full-policy execution cost.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Action Chunking, Inference-Time Guidance, Trajectory Smoothing

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_5Z9eOpenReview ↗
Rating
3(+1)
Confidence
4
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SEAM, a retraining-free inference-time method aimed at mitigating inconsistencies between consecutive action chunks in Flow Matching-based VLA policies. SEAM leverages the unexecuted portion of the previous action chunk to construct a temporally aligned reference and applies closed-form consistency corrections during ODE integration via Velocity-guided Loss Steering (VLS), without requiring backpropagation through the policy network. The authors validate the proposed method across four simulation task suites in LIBERO using π0.5\pi_{0.5} (or pi0.5) and SmolVLA, comparing it against methods such as RTC and ACT-TE. Experimental results demonstrate that SEAM reduces discontinuities at action chunk boundaries and end-effector motion jerk with inference overhead close to that of the original policy, while maintaining or improving success rates in several tasks, although its performance gains are not consistent across all tasks.

Critical strengths

None

Critical weaknesses

Both the task success rate and end-effector motion smoothness evaluations in the paper are conducted solely within the LIBERO simulation environment, lacking validation through real-world robot deployment. Although the authors report realized end-effector jerk, this metric is computed from end-effector positions in the simulator and cannot directly reflect vibrations, contact stability, or control performance during real-world robot execution.

Furthermore, RTC is already a well-established solution for addressing cross-chunk inconsistencies and achieves superior boundary continuity metrics in the paper's main experiments. While the primary advantage of SEAM lies in reducing inference computational overhead, the existing experiments have not yet demonstrated that the additional inference latency of RTC constitutes a significant bottleneck in practical robot control. Therefore, the lack of a direct comparison between SEAM and RTC on the same real-world robot platform makes it difficult to determine whether SEAM's computational efficiency advantage is sufficient to translate into tangible gains in actual control performance.

Questions for authors

In robotic tasks genuinely affected by cross-chunk inconsistencies, does SEAM offer a better trade-off among success rate, actual motion smoothness, and execution efficiency compared to RTC?

Could the authors provide real-world robot experiments on manipulation tasks where cross-chunk inconsistencies are pronounced or where trajectory discontinuities can directly lead to execution failure, particularly in contact-sensitive (or contact-rich) manipulation?

Such experiments should compare SEAM and RTC under identical policy and control settings, reporting task success rate, actual execution smoothness (e.g., measured end-effector jerk), and task completion time alongside inference latency.

Reviewer_svy8OpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces SEAM, a method designed to smooth out robotic movements generated by VLA policies without needing any extra training. It tackles the problem of jerky motions that happen when a robot strings together different "action chunks" by using the leftover part of the previous action chunk as a guide. The mathematical corrections are applied during the flow-matching process, avoiding the need for complex backpropagation, and experiments show it improves smoothness with very little extra computing cost.

Critical strengths
  • It solves a very real problem in robotics (using VLAs), as jerky movements between action chunks can mess up a robot's ability to complete a task.
  • The algorithm is highly efficient as it appears to me and in tests, it significantly smoothed out the robot's actions (reducing "jerk" from 0.194 to 0.141) and keeps the same success rate with only about 1% more computing power.
  • The experiments are well-designed, featuring good control tests and comparisons against other models to prove the method actually transfers well.
Critical weaknesses

The core idea isn't entirely new. Several other papers (like RTC, Bidirectional Decoding, Legato, and POTR) already use past actions to smooth out future ones. The authors need to do a better job explaining why their specific mathematical approach is a major step forward rather than just another variation. (just a writing modification needs to do to make it differentiable.)

  • Smoother motion doesn't always mean the robot succeeds. In one test, while the movement got smoother, the robot's success rate dropped from 45.6% to 34.0%. This suggests that forcing the robot to stick to its old plan might stop it from reacting and changing course when necessary.
  • Some of the improvements in the robot's success rate are very small, and the paper doesn't use statistical tests to prove those small bumps aren't just random luck.

note a weakness but I want to point out that try this on the hardware if you can during our discussion period, I am curious to see the results.

Questions for authors
  1. What exactly makes SEAM different from existing methods like RTC or Legato? Is the main breakthrough the math, where it happens in the process, or just that it's faster? (I am hoping you would make changes in the manuscript)
  2. Based on your tests, which specific part of your math formula is responsible for balancing smooth movement with the robot actually completing its task?
  3. Why did the success rate drop so much in the SmolVLA LIBERO-10 test even though the motion was smoother? Can you pinpoint what went wrong?

small ask: can you also do statistical tests that minor improvements in task success are actually reliable?

1 comment

Thank you for the detailed response and additional experiments. The clarification of the method, SmolVLA tuning analysis, statistical tests, and real-robot results address my main questions; the hardware study is still limited to one task and 20 trials per method, so I will maintain my score and current rating is positive given the paper state.

Thank you again and good luck!!

Reviewer_2E7YOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes SEAM, a training-free inference method for improving continuity between action chunks generated by flow-matching VLA policies. It uses the previous chunk’s unexecuted tail as a temporally aligned reference. Its core mechanism, Velocity-guided Loss Steering, applies a closed-form correction after each Euler step, blending the candidate toward a time-dependent mean reference. Subsequent velocity evaluations use the corrected state, without backpropagation through the policy network. Experiments with π0.5\pi_{0.5} and SmolVLA across four LIBERO suites evaluate task success, action continuity, realized end-effector jerk, and inference overhead.

Critical strengths
  1. SEAM preserves the pretrained policy and requires neither additional training. Full-policy timing supports its low computational overhead. This makes the method a potentially useful addition to existing action-chunked policies.
  2. The experiments include reference-content controls, different execution lengths, delivery delays, and realized end-effector jerk. Particularly valuable is the evaluation on jointly successful trajectories, which shows lower jerk with similar completion lengths. This strengthens the evidence that the improvements are not simply consequences of unsuccessful or substantially slower execution.
Critical weaknesses
  1. Novalty limited. RTC already uses the previous chunk’s unexecuted actions to constrain subsequent generation, while iterative replacement during generative sampling has precedents in inpainting methods such as RePaint. SEAM’s specific contribution is the mean reference and its scheduled soft correction. Equation (9) makes this update a convex combination, and Appendix A derives its gradient from an expected quadratic loss. These are valid constructions, but the recurrent use of corrected states is also inherent in iterative replacement methods. The paper therefore needs stronger evidence that its particular reference and correction schedule provide a meaningful advance beyond a lightweight adaptation of these existing ideas.
  2. The output-blending baseline is insufficiently evaluated. Table 3 evaluates output blending only at a fixed weight of 0.43. Although Appendix A.1 explains this choice, no weight sweep is reported. It remains unclear whether another blending weight could achieve a success–jerk trade-off comparable to SEAM. This comparison is important for establishing the benefit of applying corrections inside the solver.
Questions for authors
  1. Comparative evidence: Does SEAM retain an advantage over output blending when the blending weight is varied? A success–EEF-jerk comparison under otherwise matched settings would materially affect my assessment of the proposed solver placement.

  2. Technical clarification: Why is the reference evaluated at the pre-update time tit_i but applied to the post-Euler candidate? With the default settings, the final reference is 0.9a0.9a and the blending coefficient is 0.180.18. Even a final candidate equal to the aligned action aa would therefore become 0.982a0.982a. Is this additional contraction intended, and does using the candidate’s time ti+1t_{i+1} change the results? This concerns the implemented correction, rather than a request for a new method.

Comments on the paper

1 comment

Thank you for the authors' efforts and detailed response. Our concerns have been largely addressed, and we will consider increasing our rating.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.