acceptodds
Under review as a conference paper at ICLR 2027

TAMGen: Time-Aware Motion Generation for Goal-Conditioned Humanoid Control

Abstract

Generative humanoid motion models can reach a spatial goal while still producing an inappropriate motion for the requested time. For the same spatial goal, simply speeding up or slowing down a motion to meet different time commands may compromise its physical quality, even if the endpoint remains accurate. This is because contact patterns and whole-body poses may also need to adapt to the time requirements. To address this challenge, we introduce TAMGen, a framework for time-aware motion generation and evaluation. Its generator uses flow matching with a root–pose architecture to capture the coupled effects of spatial goals and explicit time budgets, so that changing the time budget can change motion content rather than merely rescale playback speed. However, generated candidates may differ in target accuracy and physical quality. We therefore train a Motion Compatibility Evaluator (MCE) to assess complete candidates and select motions with better target accuracy and physical quality. When no duration is specified, MCE selects both a duration and a corresponding motion. Experiments show that TAMGen adjusts contact patterns and whole-body poses across different time commands. Compared with MotionBricks, TAMGen reduces root endpoint error by 81.6% on LAFAN and 84.9% on BoneSeed. Mean foot-sliding speed decreases by 48.6% on LAFAN and 31.6% on BoneSeed. Real-world experiments demonstrate real-time tracking of changing root and hand targets and recovery from external disturbances.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Robotics, Humanoid Robot, Motion Generation, Robot Learning

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

This paper presents TAMGen, a framework for time-aware, goal-conditioned humanoid motion generation. The main motivation is that satisfying the same spatial target under different time budgets may require qualitatively different motion patterns, rather than simply temporally rescaling the same trajectory. TAMGen addresses this with a flow-matching-based generator that decomposes generation into root trajectory and articulated pose generation while explicitly conditioning on duration, physical time, and normalized motion progress. The framework further introduces a Motion Compatibility Evaluator (MCE), which scores generated candidates based on both goal satisfaction and physical motion quality and can also jointly select a duration when the completion time is unspecified. The method is evaluated on G1-retargeted LAFAN and BoneSeed motions, sparse-target tasks, and a physical Unitree G1 robot. The results demonstrate strong target accuracy, improved physical-quality metrics after MCE selection, online target updates, and recovery from external disturbances.

Critical strengths
  1. The paper addresses an important and somewhat underexplored distinction between spatial goal satisfaction and temporal appropriateness. Simply reaching a desired state at a requested timestamp is not sufficient if the resulting contact sequence or whole-body coordination is physically inappropriate. Explicitly modeling duration, physical time, and normalized progress is therefore well motivated, especially for humanoid control. The root-to-pose decomposition is also a sensible architectural choice for this setting.
  2. The work provides a fairly complete generation-to-execution pipeline rather than only evaluating generated motions offline. MCE provides a practical mechanism for selecting among stochastic candidates according to both target compatibility and physical quality, and the quantitative results show that this selection consistently improves the raw generator. The physical G1 experiments further demonstrate online target updates and disturbance recovery, making the practical relevance of the method substantially stronger. The paper reports that MCE improves all seven metrics on both datasets, although the best method still varies for some physical metrics.
Critical weaknesses
  1. Although the paper demonstrates promising results for time-aware motion generation, the evaluation of duration-dependent adaptation remains relatively limited. Most demonstrations focus on locomotion, with some simple object-interaction examples, leaving it unclear whether the framework generalizes to more diverse actions, such as more complex object manipulation under different time constraints.
  2. While the paper motivates its approach by highlighting the limitations of simple temporal rescaling, a direct comparison with such a baseline is missing. A more systematic evaluation across motion types and duration constraints would further strengthen the paper's central claim.
Questions for authors

Q1. Could the authors clarify how well TAMGen generalizes to different motion types and more extreme duration constraints? For instance, can it generate appropriate motions for fast and slow object manipulation? Additionally, how does the framework handle unusually short or long duration constraints (e.g., traveling 3 m in 1 second or 2 minutes)? Does it adapt the motion accordingly, or are there limitations on the supported duration range? Q2. Since the paper emphasizes adapting motion content rather than merely rescaling motion speed, could the authors provide a comparison with a simple temporal-rescaling baseline to better demonstrate this advantage?

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

This paper studies an interesting point: goal-conditioned humanoid motion generation with an explicit arrival time. It combines a time-conditioned root-and-pose flow generator with a Motion Compatibility Evaluator (MCE) that selects candidates based on goal satisfaction and motion quality. The selected motions are tracked on a physical G1 robot, including settings with changing targets and disturbance recovery.

Critical strengths

The paper addresses an important gap: reaching the right pose at the right time may require a different gait or contact pattern, rather than simply changing playback speed. The system is coherent, and the comparisons on LAFAN and BoneSeed show strong endpoint accuracy and reduced foot sliding. The real-robot demonstrations are a substantial strength, particularly the online target updates and replanning after a disturbance.

Critical weaknesses

My main concern is whether the experiments isolate the effect of time-aware generation. An ablation removing the explicit duration and physical-time embeddings from Root Flow and Pose Flow would clarify their contribution. Since trajectory length also conveys duration, a fixed-length generation followed by time warping would provide a complementary baseline. Reporting results before and after MCE selection would further separate generation from reranking. In Figure 3, within-duration contact-pattern similarity is lower than cross-duration similarity for TAMGen; this makes it difficult to separate a systematic response to duration from ordinary generation diversity. The root boundary construction also favors endpoint accuracy, so that metric alone does not establish the benefit of time conditioning. Finally, physical-quality gains are mixed across metrics, and repeated hardware measurements of arrival-time error and tracking success would strengthen the execution claim. I'm willing to raise my score if the authors properly address my concerns.

Questions for authors

see weaknesses

Reviewer_Er99OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TAMGen generates goal-conditioned humanoid motions under explicit time budgets. It combines a time-conditioned motion generator with a learned Motion Compatibility Evaluator (MCE) for candidate selection. The paper evaluates goal accuracy and physical quality on two motion datasets and demonstrates online target updates and disturbance recovery on a Unitree G1.

Critical strengths

The paper addresses a useful problem: different time budgets for the same goal may require different steps and whole-body coordination. The system shows promising results, including strong endpoint accuracy and improvements in all seven reported metrics when MCE is added to the Generator on both datasets (Table 1). The hardware demonstrations further support the feasibility of the overall pipeline.

Critical weaknesses

The experiments do not yet clearly establish the benefits of the two main designs. Figure 3(a) shows differences between motion segments, but does not clearly separate the effect of the requested duration from general motion variation. The benefit over simply changing playback speed also remains unclear. Similarly, MCE improves the results, but its advantage over simpler motion-scoring methods is not established. These missing comparisons limit the evidence supporting the paper’s main contributions. If additional experiments clearly demonstrate the benefits of both designs, I would consider raising my rating to 3 — Weak acceptance.

Questions for authors
  1. Given prior work on time-conditioned motion such as Robust Motion In-betweening, could the authors better demonstrate the benefits of generating motions for the requested duration over simply changing playback speed?

  2. Drawing on the cost-based matching described in Learned Motion Matching, could the authors compare MCE with simpler selection methods on the same candidate motions to clarify its added value?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.