acceptodds
Withdrawn from ICLR 2027

SURE: Spatiotemporal Uncertainty Representation for Efficient Vision-Language-Action Learning

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-language-action (VLA) policies have become a promising paradigm for end-to-end robot control, but their effectiveness still depends heavily on large volumes of demonstration data. In manipulation tasks, this dependence is compounded by structured spatiotemporal uncertainty in the action space: multiple nearby actions or trajectories may successfully solve the same task, while semantically equivalent demonstrations often unfold at different speeds. Standard pointwise supervision treats each demonstrated action as an exact target, imposing an unnecessary learning burden under such uncertainty. We present SURE, a data-efficient VLA training framework that explicitly models both spatial and temporal uncertainty. SURE augments a VLA backbone with a lightweight transposed-convolutional probabilistic action decoder that captures spatial uncertainty through local correlations among neighboring action bins. It further introduces Soft-DTW alignment supervision to accommodate local timing variations between predicted and demonstrated action chunks. Together, these two components reduce sensitivity to demonstrator-specific poses and execution timing, enabling more data-efficient action learning. We evaluate SURE on LIBERO, RoboTwin 2.0, and real-world manipulation tasks. Even with only 10% of the demonstration data or 20% of the training steps, SURE achieves manipulation success rates that match or surpass baseline VLAs.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Vision-Language-Action Models, Robot Learning, Uncertainty Modeling, Structured Action Representation, Temporal Alignment, Imitation Learning

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

This paper proposes SURE, a data-efficient VLA framework that addresses spatial and temporal uncertainty in robot demonstrations. It combines a probabilistic action decoder using transposed convolutions with Soft-DTW alignment supervision to accommodate variations in action trajectories and execution timing. Experiments on LIBERO, RoboTwin 2.0, and six real-world tasks demonstrate improvements in manipulation performance and data efficiency.

Critical strengths
  1. Addressing spatial ambiguity and temporal variations in robot demonstrations is meaningful for data-efficient VLA learning.
Critical weaknesses
  1. A central motivation of the paper is to capture multiple valid actions under spatial uncertainty. However, the probabilistic decoder ultimately collapses its predicted distributions into deterministic actions through expectation-based decoding. This design may average distinct action modes and does not demonstrate how multimodal uncertainty is preserved during execution. Without direct evaluations of multimodality or comparisons with alternative decoding strategies, the central methodological claim remains insufficiently supported.

  2. The reported OpenVLA-OFT baseline is substantially below its originally published performance, making the claimed absolute performance advantage difficult to establish without a clearer explanation of the evaluation settings.

Questions for authors
  1. ActionMap [2] already explores structured probabilistic action representations, while CLASS [3] uses DTW-based alignment for robot representation learning. Could the authors clarify the methodological novelty of SURE beyond combining probabilistic action decoding with existing temporal alignment techniques?

  2. Although the decoder predicts action distributions, the final action is obtained through expectation-based decoding, which may average multiple valid action modes. Could the authors provide direct evidence that the proposed decoder effectively models spatial uncertainty and multimodal action distributions?

  3. The paper demonstrates performance improvements from Soft-DTW but lacks direct experiments on execution-speed variations. Could the authors evaluate controlled temporal perturbations and clarify how Soft-DTW interacts with pointwise L1 supervision? Additional comparisons with alternative trajectory-level losses would strengthen this contribution.

[1] Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. RSS 2025.

[2] ActionMap: Robot Policy Learning via Voxel Action Heatmap. arXiv 2026.

[3] CLASS: Contrastive Learning via Action Sequence Supervision for Robot Manipulation. CoRL 2025.

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

This paper proposes SURE, an action decoding and training framework for Vision-Language-Action (VLA) models built on OpenVLA-OFT. It replaces the continuous MLP action head with a transposed-convolutional head that predicts a 256-bin distribution regularized toward a Gaussian prior and decodes actions via expectation, while adding a Soft-DTW divergence loss on cumulative action sums alongside L1 regression. Experiments with OpenVLA-7B on LIBERO, five RoboTwin 2.0 tasks, and six real-world tasks report gains over baselines.

Critical strengths
  1. Compressing OpenVLA-OFT's oversized action head (151.12M parameters) to 16.93M parameters via a grouped 1×11 \times 1 convolution (Table 4) is practical and speeds up early convergence under low-step (10k) and low-data (10%) budgets.
  2. The evaluation covers single-arm simulation (LIBERO), bimanual simulation (RoboTwin 2.0), and six real-world manipulation tasks.
Critical weaknesses
  1. The method is a straightforward combination of existing components whose design contradicts its "spatiotemporal uncertainty and multimodality" motivation. Decoding via global expectation an,j=∑r=1RPn,j,rcra _ {n,j} = \sum _ {r=1}^R P _ {n,j,r} c _ r (Eq. 4) under a narrow unimodal Gaussian prior (σ=1.5\sigma = 1.5 bins, Eq. 9) and primary LL1\mathcal{L} _ {\mathrm{L1}} supervision collapses multimodal distributions to the mean just like L1 regression, and OpenVLA-OFT [1] (Table I) already showed that continuous L1 regression beats 256-bin discretization on LIBERO.
  2. The ablation in Table 5 confounds 256-bin probabilistic decoding with a 9×9\times reduction in action-head parameters (16.93M vs. 151.12M in OpenVLA-OFT). There is no comparison against a parameter-matched (~17M) continuous conv head without 256-bin upsampling, even though upsampling to 256 bins increases head FLOPs by 4.5×4.5\times (Table 4) and learns diffuse, noisy bin stripes on LIBERO (Appendix Fig. 4).
  3. The baseline comparisons on LIBERO and RoboTwin 2.0 are not convincing. In Table 1, OpenVLA-OFT [1] is stopped at 50k steps (94.3%, 85.0% on Long) before its official 100k-step LR decay and 150k-step schedule (97.1% average and 94.5% on Long in [1]), which SURE (97.0% average, 91.0% on Long, single seed) does not surpass. In Table 2, the OpenVLA-OFT baseline (59.2%) is copied from JEPA-VLA [4] under a weaker setup (no proprioception, batch size 8 vs. SURE's proprioceptive input and batch size 16), while omitting JEPA-VLA itself from that same table (76.2% average on those five tasks, above SURE's 72.2%).
  4. All experiments use a single backbone (OpenVLA-7B), and the temporal module directly applies off-the-shelf Soft-DTW divergence [2, 3] (Eq. 7, without citing [3]) to cumulative action sums xi=∑ℓ=1iaℓ,1:da−1x _ i = \sum _ {\ell=1}^i a _ {\ell, 1:d _ a - 1} (Eq. 5)—a formulation that only makes physical sense for translational deltas (LIBERO), not for RoboTwin 2.0 where actions are absolute joint angles (Appendix A).

[1] Kim et al. 2025. Fine-Tuning Vision-Language-Action Models: Optimizing Speed and Success. RSS 2025.

[2] Cuturi and Blondel. 2017. Soft-DTW: a Differentiable Loss Function for Time-Series. ICML 2017.

[3] Blondel et al. 2021. Differentiable Divergences Between Time Series. AISTATS 2021.

[4] Miao et al. 2026. JEPA-VLA: Video Predictive Embedding is Needed for VLA Models. arXiv:2602.11832.

Questions for authors
  1. In Table 5, how does a parameter-matched (~17M) continuous head—retaining the grouped 1×11 \times 1 conv bottleneck of SURE but directly regressing scalar actions (R=1R = 1) with LL1\mathcal{L} _ {\mathrm{L1}}—perform compared to the 256-bin head?
  2. How does SURE perform against the official converged 150k-step OpenVLA-OFT on LIBERO (97.1% average, 94.5% on Long in [1]), and how does OpenVLA-OFT perform on RoboTwin 2.0 when trained with the same proprioceptive input and batch size 16 as SURE?
  3. Why is the cumulative sum in Eq. (5) applied to RoboTwin 2.0 where actions are already absolute joint angles (Appendix A), and does SURE generalize to other VLA backbones (e.g., π0\pi _ 0 or GR00T)?
Reviewer_TuDBOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes SURE, a training framework that aims to improve the data efficiency of VLA policies by accounting for spatial and temporal variability in demonstrated actions. On top of OpenVLA-7B, it adds a transposed-convolutional probabilistic decoder that predicts a 256-bin distribution per action dimension and outputs its expectation, together with a Soft-DTW loss that compares the cumulative trajectories of predicted and demonstrated action chunks while tolerating local timing shifts. On LIBERO, RoboTwin 2.0, and six real-world tasks, SURE reports success rates that match or exceed OpenVLA-OFT and other baselines with fewer training steps or less data.

Critical strengths

Simple, lightweight, and consistent gains across settings. The method only adds a lightweight head and auxiliary losses to an existing VLA and requires no iterative inference such as diffusion-based decoding. Improvements under reduced training steps or data are reported consistently across LIBERO, RoboTwin 2.0, and real-world experiments. The motivation, that treating every demonstrated action as an exact target is inefficient, is intuitive, and the paper is clearly written.

Critical weaknesses
  1. The claim of modeling spatiotemporal uncertainty does not match what the method actually does. The spatial decoder outputs the expectation of the predicted distribution, and the main loss is an L1 loss on this expectation. When the prediction is multimodal, the expectation falls between modes, which is precisely the averaging failure that multimodal action representations aim to avoid; this contradicts the claim that the decoder captures multimodality (Figure 1). The ingredients, a Gaussian-smoothed histogram target and expectation decoding, are essentially those of HL-Gauss (Imani & White, 2018; Farebrother et al., 2024), and the paper does not disentangle whether the gains come from these known effects (label smoothing, robustness to noisy targets) or from uncertainty modeling.

    Soft-DTW has long been used as a training loss for sequence prediction (Cuturi & Blondel, 2017; DILATE, Le Guen & Thome, 2019). Moreover, the DTW cost is computed from the expected actions and does not use the predicted distributions, so spatial and temporal uncertainty are not modeled jointly. Finally, the framewise L1 loss (weight 1), which enforces exactly the per-step, exact-timing supervision that the paper argues is an unnecessary burden, remains the primary objective, with the tolerant terms (KL at 0.05, Soft-DTW at 0.5) added on top of it. It is therefore unclear whether the gains reflect relaxed supervision or auxiliary regularization.

  2. The comparisons are not on equal footing, and the statistical support is limited. The OpenVLA-OFT paper itself reports 97.1% average success on LIBERO, on par with SURE's 97.0%; the claim of outperforming OFT relies on the authors' own lower reproduction (94.3% at 50k steps). Most baseline numbers are taken from other papers with differing protocols, and SURE and OFT are evaluated with a single seed. The ablation is limited to one suite (LIBERO-Spatial), one seed, and small differences (1.6–2.2 pts), and no experiment directly tests the central claim that the transposed-convolutional structure captures correlations among neighboring bins. The data-efficiency study also covers only one suite with a fixed number of training steps, so smaller data fractions mean many more passes over the same data; this setting favors any regularizer and cannot separate the claimed mechanism from generic regularization.

Questions for authors
  1. Are the predicted distributions actually multimodal?

  2. Why is the framewise L1 loss kept as the primary objective, with the tolerant terms added on top? What happens if L1 is replaced by, rather than supplemented with, the Soft-DTW and KL terms?

  3. On RoboTwin 2.0 the actions are absolute joint angles, and Appendix A states that Soft-DTW is applied directly to joint-action chunks. How does this relate to the formulation in Section 3.2, which builds trajectories as cumulative sums of actions? Please also specify the action representation used in the real-world (SO-101) experiments.

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

Related papers

Discussion (0)

Sign in to comment.