acceptodds
Under review as a conference paper at ICLR 2027

T2Align: Tempo and Transition Alignment for Asynchronous Robot Manipulation

Abstract

Chunked imitation learning policies, especially vision-language-action (VLA) models, have shown impressive performance in real-world robot manipulation by predicting temporally extended action chunks. However, deploying chunked policies in time-critical dynamic tasks remains challenging: task states change during execution, while success depends on reacting to these changes and completing motions within narrow temporal windows. We present Align, a robot manipulation system that provides hierarchical tempo and transition alignment for asynchronous execution of chunked policies. At the execution level, Align preserves the policy-intended execution tempo under safety constraints. At the action chunk level, Align aligns returned chunks to the current execution timeline and stabilizes their transition through continuity-consistency blending and similarity-gated transition control. Align is plug-and-play with off-the-shelf policies and improves their applicability to time-critical dynamic tasks. Experiments with widely used policies show that Align substantially improves success rate and stability.

Primary area:
infrastructure, software libraries, hardware, systems, etc.
Keywords:
Robotics, Asynchronous Inference, Vision-Language-Action Models, Time-Critical Manipulation

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_7ihoOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies asynchronous deployment of chunked imitation-learning policies in time-critical robot manipulation, where the intra-chunk tempo dilation and inter-chunk discontinuities can affect the stability of the policy. It introduces a training-free framework that combines two levels of alignment to mitigate these issues. The experiments shows that the method can improve the performance of the policy in both simulation and real-world tasks.

Critical strengths
  • The formulation of the temporal alignment is clear and well-motivated, and the decomposition of the problem into two levels of alignment is reasonable.
  • The method is simple and training-free, and can be easily applied to existing chunked policies. The experiments also show its effectiveness on multiple base models and tasks. The ablation study provides detailed analysis of the contribution of each component.
  • The writing is clear and the paper is well-structured, and the contribution is easy to follow.
Critical weaknesses
  • The compensation mechanism of the method is limited. For example, if the action scale of the base model is too large to finish in a tick, the accumlated B is always negative.
  • The chunk level blending is linear, hence the reliablity and smoothness is questionable for complex rotation movements.
Questions for authors
  • What is the consideration of not comparing with other async inference methods in real world tasks?
  • The design of BrefB_{\text{ref}} encourages the model to constantly execute the action 0.2s faster, how to ensure that this gap does not cause catastrophic failure in some tasks?
Reviewer_nG9aOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents T²Align, a plug-and-play system for asynchronous execution of action-chunked robot policies in time-critical dynamic tasks. It addresses the need to react promptly to changing task states while completing motions within narrow temporal windows. The system combines execution-level tempo alignment, which aims to preserve the policy-intended execution pace under safety constraints, with chunk-level transition alignment, which synchronizes incoming action chunks with ongoing execution. Continuity-consistency blending and similarity-gated transition control further stabilize transitions between chunks. Experiments with widely used imitation learning policies, including VLA models, show improvements in task success and execution stability.

Critical strengths

1.The paper addresses an important problem in deploying action-chunked policies for time-critical dynamic manipulation. Coordinating execution tempo and asynchronous chunk transitions is practically relevant, as successful manipulation requires both timely responses to changing task states and stable execution within narrow temporal windows. 2.The proposed method is supported by experiments across multiple policies, tasks, and robot platforms. Evaluations with Diffusion Policy, ACT, and π0.5, together with component ablations, demonstrate improvements in task performance and motion stability. These results support the practical usefulness of the framework across the tested settings.

Critical weaknesses

The evaluation provides limited evidence of robustness in complex, long-horizon dynamic tasks. The real-world tasks mainly involve relatively simple and repetitive interactions, leaving performance in sustained, multi-stage manipulation unclear.

Questions for authors

Could the authors evaluate the framework on more complex, long-horizon dynamic tasks that require recovery from execution errors and timely responses to unexpected external disturbances, to better assess its robustness under more demanding conditions?

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

T2Align is a deployment framework for chunked robot policies in time-critical dynamic manipulation. It aims to preserve the policy's intended execution tempo under safety constraints, aligns returned chunks to the current execution timeline, and uses continuity-consistency blending and similarity-gated switching to manage transitions. The paper evaluates the framework with multiple policies on real-world dynamic tasks.

Critical strengths
  1. The paper addresses a practical deployment problem in which timing and responsiveness matter as much as spatial accuracy.
  2. Real-robot dynamic tasks and multiple underlying policies provide useful evidence beyond static simulation benchmarks.
  3. Component analyses and comparisons with synchronous and asynchronous execution help explain the roles of the proposed system elements.
Critical weaknesses
  1. The contribution primarily combines execution-time alignment and switching mechanisms. The conceptual and technical distinction from existing asynchronous inference and trajectory blending methods needs to be sharpened through carefully matched strong baselines.
  2. The method treats demonstration sampling intervals as encoding the desired physical tempo. Demonstrator speed, robot dynamics, task conditions, and control frequency can vary; the consequences of these assumptions are not yet sufficiently established.
  3. Similarity gating and blending may postpone a necessary reaction or smooth away a crucial contact action. Reaction-latency distributions, threshold sensitivity, and failure analysis would make this trade-off clearer.
  4. The scope of real-robot evaluation limits generalization. More trials with uncertainty estimates and tests across robots, inference-delay distributions, and unseen disturbances would strengthen the claims.
Questions for authors

See the weaknesses.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.