acceptodds
Under review as a conference paper at ICLR 2027

When Is a Subtask Complete? Consistency-Aware Subtask Annotation for Hierarchical Robot Policies

Abstract

Hierarchical robot policies decompose long-horizon manipulation into a sequence of subtask to conduct a difficult task by high-level policy plans subtask instructions that a low-level policy executes. Training such policies requires decomposing long-horizon demonstrations into subtask segments, each labeled with an instruction and terminal state where the subtask is complete. In prior works, automated decomposition methods cannot ensure that the same subtask ends at the same state across demonstrations; one annotation may end "place the cup" at object release, while another ends it after arm retraction. Without a shared convention, the policy receives conflicting supervision about when to procede to the next subtask. In our pilot study, mixing two individually valid annotation sources lowers hierarchical policy success below that of either source alone (41.4% vs. 49.5% and 50.8%). To address this, we propose Consistency-Aware Subtask Decomposition (CASD), which learns an annotation convention from a small set of reference annotations and applies learned convention across the dataset. CASD over-proposes candidate boundaries from motion cues and scores every interval for the two properties a convention must fix: (1) which skill it shows and (2) whether it covers one complete occurrence of that skill. Because this completion criterion is learned from the references rather than inferred from per-episode visual change, it applies uniformly across demonstrations; a structured loss and dynamic programming then select a consistent labeled partition of each episode. On RoboCasa365 composite tasks, hierarchical policies trained on CASD annotations improve success rate by 10.0%p over the strongest automated annotation baseline and nearly matching simulator ground-truth annotations.

Primary area:
applications to robotics, autonomy, planning
Keywords:
subtask decomposition, data annotation, hierarchical policy

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

This paper addresses the problem of inconsistent subtask completion boundary annotations in hierarchical robot policy learning and proposes a Consistency-Aware Subtask Decomposition (CASD) method. Specifically, the method first generates candidate subtask boundaries using robot kinematic information and constructs segment representations by integrating pretrained vision-language-action (VLA) features, robot states, and action information. It then learns skill identity and completeness from a small set of human-annotated reference demonstrations, categorizing candidate segments as full, partial, or mixed skill executions. Finally, a structured loss and dynamic programming are employed to generate subtask labels and temporal boundaries that cover entire demonstration trajectories. Extensive experiments across multiple datasets validate the effectiveness of the proposed method, demonstrating competitive performance.

Critical strengths
  1. The paper presents a relatively clear research motivation by addressing the inconsistency in subtask completion criteria across demonstrations. It proposes a specific method to establish a unified annotation convention, which can lead to significant improvements in subsequent policy training.
  2. The proposed method is well-motivated and relatively comprehensive. It enables the model not only to recognize the skill performed within a segment but also to determine whether the segment represents a complete execution of that skill.
Critical weaknesses

The paper presents a relatively clear motivation, but its technical novelty is somewhat limited. Although the paper addresses the practically meaningful problem of inconsistent subtask completion criteria, the main components of CASD, including kinematics-based candidate boundary generation, MLP-based skill classification, structured loss, and dynamic programming, are largely based on established techniques. The overall framework appears to be more of a combination of existing techniques rather than a method with a clearly distinguished technical innovation. Furthermore, given that the paper primarily addresses a data annotation consistency problem, it is worth questioning whether a simpler and more efficient approach could achieve the same objective without relying on such a complex combination of technical components. 2. Lack of comparison with simpler rule-based segmentation methods. To address inconsistent subtask completion boundaries, could unified skill completion rules be directly defined based on observable physical events, such as gripper closure, object release, or end-effector retraction, and applied across the entire dataset? Such an approach might be simpler and more efficient, without requiring an additional completeness scoring model. The authors are encouraged to include comparisons with such rule-based methods to justify the necessity and practical advantages of the proposed complex learning framework.

Questions for authors
  1. The proposed method relies on a small set of human-annotated reference demonstrations to learn a unified subtask completion criterion. However, it remains unclear whether these reference annotations are sufficiently consistent. The paper mentions that two annotators contributed to the reference dataset. If different annotators disagree on the completion boundaries of the same skill, how does CASD ensure that the learned completion criteria are reliable and consistent?
  2. Does complete skill execution necessarily imply successful skill completion? If a skill execution ultimately fails, how does CASD distinguish between the completeness of the action execution and the actual success of the skill?
  3. Is a unified subtask completion criterion always appropriate? Inconsistent completion criteria across different demonstrations do not necessarily imply that all demonstrations should follow the same completion convention. The appropriate completion timing may depend on the specific task context and subsequent actions. How does CASD account for such context-dependent variations?
Reviewer_nLz7OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies consistency-aware subtask annotation. Authors train a hierarchical model to first detect task identity, then detect the task finish. In this way the model can split a long horizon task into several subtasks which are easier to learn for robotics policy.

Critical strengths
  1. Task annotation is getting more and more important in robotics, but the problem has not been systematically learned by the community enough. Thus, I think this is a well-motivated paper.

  2. The hierarchical method is simple and makes great sense.

Critical weaknesses
  1. Biggest issue of the paper is comparison fairness, in table 4, where are the GT label comes from? If you are using labels from open-sourced dataset, what if these labels are wrong and noisy (which is very likely to happen). Also for the baselines, author only disclose tune both models with the same data but what are their backbones? Do they utilize pretrained QWen VLLM? If there is a gap compared to proposed method, how you separate the gain from method and gain from larger pretrained model?

  2. Second issue is the clarity, (1) for real robot task, how do you annotate? If you use 1 annotator how you make sure they are correct, if there are multiple ones, how you deal with conflict? (2) do you have any protocol or candidate subtask/verbal set to standardize the language of annotators?

  3. The paper has lots of careless mistakes, even only at abstract, there are 2 at 018, 032 lines.

Overall the task and method are interesting, but the paper seems not ready. If authors can resolve the mentioned issues I will increase the rate.

Questions for authors

same as weakness

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

The research question in this paper is how to mitigate the inherent ambiguity in subtask completion annotations, e.g., the exact stage at which a subtask is considered complete can differ even given the same task instruction. Such ambiguity can lead to conflicting supervision for robot learning and degrade the model’s performance when annotations based on different conventions are mixed together. In this work, the authors present CASD, a method that learns a shared completion convention from human-annotated references. CASD predicts which skill is performed in the input interval and whether it is mixed, full, or partial, and uses these predictions with dynamic programming to produce episode partitions.

Critical strengths

The paper identifies an realistic problem

individually reasonable subtask annotations can still follow different completion conventions across a dataset. Mixing two annotation sources reduces the success rate from 49.5% and 50.8% when used separately to 41.4%, providing clear empirical motivation for addressing this inconsistency (Table 1).

Proof of Effectiveness

On RoboCasa365, CASD improves the average success rate by 10 percentage points over the strongest automated annotation baseline, supporting the practical value of its annotations for downstream policy learning (Tables 2–4).

Critical weaknesses

Generalizability and scalability need stronger evidence.

Honestly the overall statement "Consistency-Aware Subtask Annotation " is aimed quite high, and very easy to be over-claimed and is over-claimed in this work. The author use 640 RoboCasa365 human references to infer a consistent convention, with a dependence on fixed skilled vocabulary, to match the statement of "Consistency-Aware Subtask Annotation", the author may consider conduct experiment on other simulation environment, and larger data-scale with more diverse skill vocabulary and larger quantitive real-world experiment to prove the generalizability and scalability.

Stronger VLM-based baselines.

Intuitively, such segmentation and annotation can be done also with a general VLM. Although, possibly, by the time of submission, some strong LLM\VLM agent e.g gpt6-astra is not available, but due to its especially strong performance on SOTA robotic benchmark [2], it's better to have a VLM baseline for SOTA performance comparison, e.g. GPT6-Astra, FLORA [1].

[1] https://research.nvidia.com/labs/gear/flora/assets/paper/flora.pdf

[2] https://arxiv.org/abs/2609.24170

Questions for authors

Given the capabilities of modern agents and large models, could the authors justify CASD’s reliance on a fixed skill vocabulary, dataset-specific annotators, and human-annotated references, and provide evidence that it can scale and adapt to larger and more diverse datasets while maintaining annotation quality and downstream policy gains without a proportional increase in human effort?

Comments on the paper

1 comment
Area_Chair_4nHXOpenReview ↗

Reminder: Please Review Authors’ Rebuttal and Provide Feedback

Dear Reviewers,

Thank you for your reviews. If you have not already done so, please read the authors' rebuttal and share your feedback in the discussion thread at your earliest convenience. Your timely responses would be greatly appreciated.

Best,
AC

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.