acceptodds
Under review as a conference paper at ICLR 2027

LaDyP: Phase-Aware Action Chunking

Abstract

Action-chunking policies predict several future actions from the current observation, so a single chunk often spans a grasp, a release, or the switch to the next sub-task. On long-horizon instructions these transitions are where most rollouts fail. We introduce LaDyP, a phase-conditioning interface for Flow-based chunk policies. An offline teacher labels demonstrations with recurring interaction roles, within-role progress, and control-boundary evidence; a causal estimator predicts a soft phase belief, local progress, and current and upcoming boundary likelihoods from observation history. These signals condition the action generator through a global token, chunk-position embeddings, and a boundary-gated residual velocity. We attach LaDyP to a Flow policy trained from scratch and to the pretrained SmolVLA, and evaluate both on all four LIBERO suites with 1,500 paired trials per suite and method. On LIBERO-Long, LaDyP raises success from 49.2% to 57.5% for the Flow policy and from 71.8% to 78.7% for SmolVLA, while gains on the short suites are small. Delaying the signals by 32 steps or replacing all of them with signals from another episode removes most or all of the gain, and failures during engagement, release, and sub-task switches decrease by 31% and 30% for the two backbones while approach and transport failures stay near their Base counts.

Primary area:
applications to robotics, autonomy, planning
Keywords:
robotic manipulation, action chunking, phase conditioning, flow-based policies, vision-language-action models, LIBERO

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_sdPNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a specific mechanism for encoding task progress or in state information into a conditional policy (VLA/diffusion policy). They show results on the LIBERO benchmark.

Critical strengths

Overall this paper is pretty well written. It's very nice that the authors included variant statistics on their experiments, and the figure 5 ablations are good. Similarly figure 7 is a nice analysis to show the speed improvements.

Critical weaknesses

The motivation in this paper is slightly strange. In the introduction, the authors mentioned that if a policy goes OOD then performance degrades, which is true. Then they immediately discuss that the action generator does not know that a transition is imminent. It’s not clear how these two points are connected. Why does the action generator need to know the phase transition explicitly? Why not let the policy learn that implicitly? And how does that cause the OOD behavior to be better? Certainly even if the policy knows where the phase transition is, if it goes OOD performance will degrade.

As such, I'm not sure if these experiments are entirely fair. The LaDyP results are incrementally better than the clock results and within the margin of error for clock (and several null adapter results). The LaDyP network has access to a longer history (32 steps) than the baselines and one might ask what the performance would be like if instead of this specific structure for how that history is used, one simply gave the policy the 32 frame history (or a smaller recurrent head to compress that 32 frame history if it's intractable to consume it in the diffusion body). If LaDyP is truly a generally valuable addition presumably the authors could add it to OpenVLA-OFT to set SOTA score on this benchmark?

Finally this paper would be much stronger with some real world experiments showing that the improvements are transferable.

Questions for authors

“no shuffled, clock, discrete-option, or clustered-segment signal reaches LaDyP, and a 32-step delay removes most or all of the gain. Failure attribution locates the change at engagement, release, and sub-task switches.” This is unclear and should be re-written.

Reviewer_wkE4OpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes adding “task phase”-conditioning information to improve the performance of flow-matching policies for long-horizon tasks. The hypothesis is that by being aware of these critical events the policy can be more precise at transitions. The authors propose a rule-based annotation of these phase boundaries based proprioceptive + visual feature changes. The results demonstrate gains on long-horizon simulated manipulation tasks for training from scratch and fine-tuning SmolVLA and reduced failures at task phase transitions.

Critical strengths
  • The paper has a reasonable high-level hypothesis to address shortcomings of flow-matching policies in long-horizon tasks. The empirical evidence demonstrates improvements that come from having a belief over impending task phase transitions.
Critical weaknesses
  • Several sentences in the paper are hard to read and parse (potentially due to use of an LLM to rephrase to be concise). Examples:
    • Introduction sentences (L049-053, L093-096) are hard to parse.
    • “rule-based” labelling statement is hinted at in related work rather than explicitly acknowledged in the introduction.
    • I find the “execution timing” section of the related work a stretch to include in this paper.
    • There are many straggler sentences (L156, L161, L196-199) that do not make much sense and hinder clarity.
    • A key implementation detail is the determination of these phases – this is largely relegated to the appendix; making it harder to grasp the main contributions. I feel the Equation 8 + justifications – should be in the main paper, otherwise the added commentary (L161) don’t make much sense.
  • Several introduced concepts don’t have sufficient backing for its design choices:
    • How the rule-based control event detection was arrived at (why those weights, thresholds etc.,)
      • Why should these thresholds be universally applicable across tasks that we want to scale this approach to?
      • This also limits scalability of the approach especially with the rule based threshold + LLM atomic phase annotation.
    • Section 3.4, what was the need for these different pathways to condition the policy? Why were the residual corrections needed near transitions? The proposed chunk position conditioning is not very clear from its description.
  • As the paper tackles the problem of improving long-horizon manipulation, the paper should compare and contrast it with suitable, potentially simpler, baselines like HI Robot that do explicit sub-task prediction. Stating that it is orthogonal (L115-117) does not serve as convincing evidence of synergy, especially given the added moving components of the proposed approach.
Questions for authors

See points under weaknesses that felt missing from a read of the paper. Additionally, had the following questions:

  • How does the paper compare with recent work (see [1-2]) that tackle similar long-horizon manipulation problems? It would be good to see why these proposed design choices are better than relevant prior work.
  • Can the authors provide more clarity on the chunk position embedding (Equation 5) -- the description and justification for this was confusing to me.

References
[1] Liu, Yuanzhe, et al. "PALM: Progress-aware policy learning via affordance reasoning for long-horizon robotic manipulation." CVPR 2026
[2] Yan, Hongyu, et al. "ProgressVLA: Progress-guided diffusion policy for vision-language robotic manipulation." arXiv preprint arXiv:2603.27670 (2026).

Reviewer_1PvCOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies timing errors in action-chunking policies during grasping, release, and subtask transitions. LaDyP uses an offline teacher to generate interaction-phase, within-phase progress, and control-boundary labels, then trains a causal estimator to predict these signals from observation history. The predictions condition a Flow policy through a global token, chunk-position embeddings, and a boundary-gated residual. Experiments cover two backbones and all four LIBERO suites. Success on LIBERO-Long increases from 49.2% to 57.5% for Flow and from 71.8% to 78.7% for SmolVLA, with smaller gains on short-horizon tasks.

Critical strengths
  1. The improvements on long-horizon tasks are consistent across the two backbones. Table 1 reports gains of 8.3 and 6.9 percentage points for Flow and SmolVLA, respectively. Both paired confidence intervals exclude zero, and the gains are positive across all three training seeds. The method preserves the chunk length and the number of actions executed per replan, giving it practical value.

  2. The experiments include targeted controls and ablations. Sections 4.3–4.6 examine model capacity, clock signals, signal shuffling, temporal delays, and component removal, and report failure categories and inference overhead. These analyses support the contribution of temporally aligned auxiliary signals to the overall improvement.

Critical weaknesses
  1. The contribution of phase labels has not been adequately separated from that of additional history. In Section 3.1, the LaDyP estimator uses the most recent 32 replanning observations, while the Flow baseline uses four and the SmolVLA baseline uses only the current observation. BiggerBase controls only for parameter count, while NullAdapter, Clock, and the shuffled-signal controls do not retain a useful history summary aligned with the current rollout. The current results therefore do not rule out additional history as the main source of improvement. Section 4.3 also does not specify whether the learned-segments control uses the same history window, estimator, and training budget. This directly affects the assessment of the independent contribution of phase labels and boundary design.

  2. The independent contributions of the key design choices remain unclear. The discrete-option control in lines 375–377 changes both the phase representation and the action heads while removing progress, boundary, and residual inputs. The performance difference therefore cannot be attributed specifically to soft phase conditioning. In Figure 5, the confidence intervals for the differences resulting from removing progress, the global token, or the local embeddings all include zero. The current experiments support the effectiveness of the complete method, but do not establish which parts of the more complex interface provide substantial benefits. The authors should clarify the scope of the existing comparisons and narrow the contribution claims according to the ablation results.

Questions for authors
  1. Does the learned-segments control use the same 32-observation replanning history, SigLIP 2 features, estimator architecture, and training budget as LaDyP? Please clarify the specific differences.

  2. Are there existing results that retain progress, boundary, and residual inputs while replacing only the soft phase belief with a one-hot representation? If not, can the discrete-option comparison support an advantage of the soft phase representation itself?

  3. Are there existing results for a simplified conditioning interface, such as jointly removing progress, the global token, and the local embeddings, whose individual contributions remain unresolved? Please provide any existing results to clarify the benefit of the full interface over a simpler implementation.

Comments on the paper

1 comment

Hi everyone, please do acknowledge the authors’ rebuttal with a brief reply, even if your assessment is unchanged.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.