acceptodds
Under review as a conference paper at ICLR 2027

Cross-Realization Consistency: Reducing Run-to-Run Variation in Further Training of Robot Manipulation Policies

Abstract

Even with the same data, objective, and training recipe, repeated further training of a robot policy can produce different functional changes because of training randomness. Stabilization approaches based on keeping the updated model close to the original policy can limit how much each run changes, but do not directly encourage similar functional changes across runs. We introduce Cross-Realization Consistency (CRC), which aligns the functional changes of two branches further trained from the same original policy. For independent runs, the shared change cancels in their functional difference, allowing CRC to target run-specific variation without requiring either branch to remain close to the original policy. Across eight MimicGen manipulation tasks using ACT-lite, CRC reduces functional difference by 21% on average and improves directional alignment on all tasks while maintaining average task success. Runs trained with CRC also behave more similarly during execution, in both their actions and end-effector trajectories.

Primary area:
applications to robotics, autonomy, planning
Keywords:
imitation learning, behavior cloning, robot learning, robot manipulation, policy learning

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_EkfMOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Cross-Realization Consistency (CRC), a method aimed at reducing run-to-run functional variation caused by training randomness during the further training of robot manipulation policies. Starting from the same original policy, the method trains two LoRA branches and introduces a cross-branch output consistency loss alongside the behavior cloning objective to encourage more consistent functional changes across training runs. The authors evaluate CRC on eight MimicGen manipulation tasks using Cross-run Functional Difference (CFD), Cross-run Directional Alignment (CDA), and closed-loop trajectory variation metrics. The results show that CRC reduces functional differences across training runs while largely preserving task success rates.

Critical strengths
  1. The paper addresses run-to-run variation during the further training of robot manipulation policies, rather than focusing solely on task success rates. Improving consistency across training runs is a meaningful research direction, particularly for robotic systems that require predictable and repeatable behavior.

  2. The experimental evaluation and ablation studies are thorough. The authors compare CRC with methods such as Standard BC, Dropout Consistency, and Mutual Learning across multiple robot manipulation tasks. They also investigate the effects of consistency constraints through controls involving Shared Batch, Batch Size, Shrinkage, and L2 Regularization. In addition, closed-loop trajectory evaluations and experiments with additional training runs provide further validation.

Critical weaknesses
  1. The methodological novelty is relatively limited. CRC primarily relies on a consistency loss between the outputs of two branches, which is closely related to existing approaches such as Mutual Learning and Codistillation. Although independent minibatches lead to further improvements, the gains over Mutual Learning and Shared Batch are relatively modest, while the method incurs additional training costs.

  2. The practical value of run-to-run consistency has not been fully demonstrated. Although the authors show that CRC reduces differences in action predictions and execution trajectories across training runs, more consistent behavior does not necessarily imply better robot policies. The paper does not provide direct evidence that such consistency improves deployment reliability, robustness, or generalization. Moreover, stronger proximity regularization settings can also achieve competitive CFD and CDA results, suggesting a potential trade-off between consistency and the magnitude of functional changes. However, the practical significance of this trade-off remains unclear.

Questions for authors
  1. CRC and Mutual Learning employ closely related consistency objectives. Given that Shared Batch already accounts for a substantial portion of the observed improvements, could the authors further clarify the unique benefits of using independent minibatches and whether the additional training cost is justified?

  2. For policies that successfully complete the same task but follow different action trajectories, why is it necessary to reduce these differences? Could the authors further explain how improvements in run-to-run consistency translate into practical benefits for robot deployment?

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

The paper proposes to use a new technique called Cross-Realization Consistency (CRC) that aims at aligning two independent random draws of the functional changes resulted from training randomness. This helps ensure seed-insensitivity, which allows attribution of changes to the training configuration / parameter changes rather than randomness. The method works by minimizing the functional inconsistency between two LoRA fine-tuned models when subject to different random training batches. The technique is evaluated on 8 MimicGen tasks and is shown to reduce the Cross-run Functional Difference (CFD) and improve Cross-run Direction Alignment (CDA), which quantify the run-to-run variation.

Critical strengths

The proposed method is technically sound, and it seems to improve the seed-insensitivity.

Critical weaknesses
  • While I can see sometimes having seed-insensitivity is useful, the paper did not extensively explore concretely how we can leverage seed-insensitivity. It would be much more convincing to me if this actually helps with some downstream tasks (e.g., faster experimentation).

  • The improvements in CFD and CDA seem marginal and it is hard to interpret how much improvement is considered significant.

Questions for authors

N/A

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

This paper studies run-to-run variation while "fine tuning" a policy on its original training data (not a typical or useful definition of fine tuning). It proposes Cross-Realization Consistency (CRC), which trains two LoRA branches off of a shared frozen policy. Each branch is further trained on its original training data, with its behavior cloning loss and an additional L2 consistency term that encourages the outputs of the branches to be similar. The authors show that for eight MimicGen tasks, ACT policies trained with CRC remain more aligned to each other while retaining similar task success rates.

Critical strengths

The paper is clearly written and transparent about its implementation. The simple testbed setting is well-designed for isolating randomness, and the experiments are carefully controlled; however, these simplifications reduce the relevance of the analysis substantially.

Critical weaknesses

My main concern with this paper is that the problem is not well-motivated, and the experimental setting has no clear use case. The paper motivates the problem with related work that addresses refinement from new data and fine-tuning generalist policies. However here, "further training" means continuing to train on the same training split used for the original policy, through a small LoRA adapter, with no new data, task, or distribution shifts. That is continued training, not fine-tuning or refinement. The to run-to-run variation we expect to see in this case, particularly since these are BC policies (i.e. continued supervised training on a fixed dataset), should be very low. This is consistent with the paper's own Figure 1, which shows very little run to run action trajectory variation, and no difference in task outcome. Furthermore, the paper treats agreement between runs as a goal in and of itself. However in practical workflows, only one policy is deployed, so whether a different seed would have produced a slightly different trajectory doesn't affect anyone using the policy. The paper provides no evidence that the run-to-run variation impacts the success rate of the further trained policies, nor does it provide evidence that CRC acts as a regularizer that improves task success.

Questions for authors
  1. Does the effect hold in typical fine-tuning settings, with new demonstrations or a new task?
  2. Why is agreement across training runs desirable in itself, given that only one policy is deployed?
  3. Can you show that CRC reduces something with downstream consequences (like variance of success rate across seeds, or that it improves success rate by serving as a regularizer)?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.