acceptodds
Withdrawn from ICLR 2027

UATL: Uncertainty-Aware Temporal Learning for Open-Loop Vision-Language-Action Control

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Vision-language-action (VLA) models map image observations and language instructions directly to robot actions, but the need for a full forward pass for each action prediction hinders the action rates required for control. Action chunk addresses this by predicting and executing multiple actions open-loop from a single policy query, reducing policy query frequency and increasing VLA action throughput. However, existing action chunk VLA methods based on deterministic continuous regression treat all timesteps and action dimensions within a chunk uniformly, capturing neither differences in predictive confidence nor temporal variation between consecutive actions. Consequently, local prediction errors accumulate during open-loop execution, degrading policy stability in fine-grained manipulation and long-horizon sequential tasks. To address this issue, we propose an Uncertainty-Aware Temporal Learning (UATL) framework for open-loop vision-language-action control. Specifically, we introduce Heteroscedastic Action-chunk Learning (HAL) to model continuous action prediction as a conditional Laplace distribution and jointly learn the action mean together with per-timestep, per-dimension scale parameters via negative log-likelihood, thereby explicitly characterizing aleatoric uncertainty in action prediction. Furthermore, we design a second-order Uncertainty-Weighted Temporal Consistency Regularizer (UW-TCR) that uses the reciprocal of the predicted scale as a per-timestep, per-dimension regularization weight, suppressing irregular jitter during open-loop execution while preserving the rapid action changes required at contact and direction reversals. Finally, evaluations on LIBERO, LIBERO-Plus, SimplerEnv, and CALVIN show improved performance in fine-grained manipulation, distribution-shift robustness, and long-horizon control, with negligible additional inference overhead.

Primary area:
applications to robotics, autonomy, planning
Keywords:
Visual-Language-Action, Action Chunk, Heteroscedastic Action-chunk Learning, Temporal Consistency Regularizer

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_BS4uOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

UATL keeps the original VLA backbone, action-chunking protocol, query frequency, and open-loop execution. It adds (i) Heteroscedastic Action Learning (HAL), which predicts a Laplace location and a per-time-step/per-dimension scale, and (ii) an uncertainty-weighted second-order temporal-consistency regularizer (UW-TCR). The inverse predicted scales weight the smoothness penalty, so confident regions are smoothed more strongly while uncertain regions can change rapidly. Experiments use OpenVLA-7B with LoRA. UATL reports 98.3% average success on LIBERO, 71.2% on LIBERO-Plus, 69.3%/54.3% on the two Google Robot settings of SimplerEnv, 63.5% on WidowX, and an average CALVIN sequence length of 4.05 (96.0/89.6/81.0/73.8/65.0% success for lengths 1--5). On LIBERO, the ablation improves 95.8% (baseline) to 96.6% (HAL) and 98.3% (HAL+UW-TCR), with only a small inference overhead.

Critical strengths
  1. The method is simple to integrate: it changes the action head and training loss without changing the backbone, policy-query schedule, or execution pipeline.
  2. HAL and UW-TCR are conceptually complementary and address both nonuniform prediction reliability and within-chunk temporal jitter.
  3. The evaluation is broad, covering standard manipulation, perturbation robustness, cross-platform transfer, and long-horizon control. The ablation and throughput measurements help isolate the proposed components.
Critical weaknesses
  1. The paper treats the predicted Laplace scales as aleatoric uncertainty, but does not report calibration metrics, reliability diagrams, scale-error correlations, or comparisons with calibrated uncertainty baselines. It is therefore unclear whether the scales are meaningful uncertainty estimates or simply adaptive loss weights.
  2. The methodological novelty is incremental relative to heteroscedastic regression, jerk/acceleration regularization, and action-chunk continuity or aggregation methods. Strong matched comparisons with RTC, overlapping-chunk aggregation, and recent temporal-smoothing methods are missing.
  3. LIBERO is close to saturation, and the absolute gain over the baseline is 2.5 points. Results are less consistent under harder transfer conditions, especially Variant Aggregation (54.3%) and drawer tasks. The paper should avoid attributing all gains to uncertainty modeling without stronger transfer evidence.
  4. There is no real-robot evaluation. A second-order penalty could suppress necessary abrupt contact transitions or gripper-adjacent changes, even if the current simulation tasks do not expose this failure mode.
  5. Results rely on OpenVLA-7B, LoRA, and a selected regularization weight (lambda_TCR=0.05). More seeds, confidence intervals, and robustness across backbones and training choices are needed to establish that the gains are not setup-specific.
Questions for authors
  1. Please report NLL, calibration error/reliability plots, rank correlation between predicted scales and realized action errors, and results with shuffled or detached random scales. This would test whether HAL learns uncertainty rather than only a reweighting heuristic.
  2. Please compare against strong temporal baselines (first/second-order smoothness, RTC or action aggregation) using the same OpenVLA checkpoint, data, training steps, and hyperparameter-selection budget.
Reviewer_hk8BOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

UATL trains an action-chunk VLA with a heteroscedastic Laplace loss and a temporal regularizer weighted by predicted uncertainty. The regularizer penalizes second-order differences between actions, with stronger penalties where predicted uncertainty is low. Evaluation on LIBERO, LIBERO-Plus, SimplerEnv and CALVIN reports improved performance with little inference overhead.

Critical strengths
  1. The method is simple to integrate. It changes the action head and training objective while preserving the existing inference and chunk-execution protocol.

  2. Evaluation covers several complementary settings, including visual perturbations, different robot embodiments and long-horizon instruction following. This provides useful evidence beyond standard LIBERO success rates.

  3. Table 6 provides a matched component comparison: success increases from 95.8% to 98.3%, while inference latency changes only from 0.148 to 0.150 seconds.

Critical weaknesses
  1. The main ablation does not establish the value of uncertainty weighting. Table 6 compares the baseline, HAL and HAL plus UW-TCR, but lacks HAL with a uniformly weighted temporal penalty. The gains could therefore come largely from ordinary smoothing.

  2. The interpretation of the predicted scales is not validated. There is no quantitative analysis showing that they track prediction errors or identify necessary rapid transitions. Figure 3 alone does not establish that the method reduces jitter while preserving contact-related motion.

  3. Important experimental details are missing, including benchmark-specific training data, chunk lengths, scale-clipping bounds and evaluation counts. Variability across runs is also unreported. The relationship between the internal baseline and OpenVLA-OFT, a closely related parallel action-chunk model, needs clarification.

Questions for authors
  1. Does uncertainty weighting outperform a uniformly weighted second-order penalty when both regularization strengths are tuned?
  2. Do predicted scales correlate with held-out action errors and necessary rapid transitions? Can the claimed smoothing effect be quantified?
  3. What are the exact training and evaluation settings for each benchmark, and how does the Table 6 baseline differ from OpenVLA-OFT?
Reviewer_RsZXOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

UATL proposes a training objective for open-loop VLA action-chunk prediction. Starting from OpenVLA-7B, it predicts both the means and scales of continuous actions, using a heteroscedastic Laplace negative log-likelihood. The predicted scales are then used to weight a second-order temporal regularizer, with gradients detached through the weights. The method aims to suppress action jitter while preserving necessary rapid changes, without modifying the inference-time execution protocol. Experiments on LIBERO, LIBERO-Plus, SimplerEnv, and CALVIN report improved performance with negligible additional inference overhead.

Critical strengths
  1. The approach is simple and inexpensive at deployment. It modifies the action head and training objective while retaining deterministic action prediction and the original execution protocol. Table 6 reports latency of 0.150 s compared with 0.148 s for the baseline. Improving control without additional policy queries or auxiliary inference-time models is a useful practical property.

  2. The reported empirical results are promising across multiple simulation benchmarks. The LIBERO ablation improves success from 95.8% to 98.3%, while experiments on LIBERO-Plus, SimplerEnv, and CALVIN provide additional evidence across different evaluation settings. These results suggest that the proposed training objective can be beneficial.

Critical weaknesses
  1. The benefit of the distinctive uncertainty-dependent mechanism is not yet convincingly established. Heteroscedastic regression is an established approach to learning predictions and input-dependent uncertainty; for example, Kendall and Gal (2017) discuss learned loss attenuation and use a Laplacian regression likelihood. Temporal smoothness penalties are also established, as the submission acknowledges. The potentially distinctive contribution is therefore the coupling of predicted scales with temporal regularization. However, Table 6 does not compare this coupling with a comparably tuned constant-weight temporal regularizer under the same heteroscedastic regression objective. The reported improvement consequently does not isolate the value of uncertainty-dependent weighting from the benefit of adding temporal smoothing. Furthermore, prediction uncertainty does not necessarily indicate where rapid action changes should be preserved: a sharp transition may be predictable, while uncertain motion may still benefit from smoothing. Direct evidence supporting this relationship would help establish why the proposed weighting is preferable to simpler alternatives.

  2. The evidence for reduced execution drift and the scope of applicability remains limited. The paper motivates its method through error accumulation during open-loop execution, but the reported success rates and qualitative trajectories do not directly establish that it reduces accumulated execution error. All evaluations are in simulation, including SimplerEnv; real-robot trials or controlled execution-perturbation experiments measuring drift would provide more direct evidence for this motivation. In addition, all UATL results use OpenVLA-7B. The other VLAs in the comparison tables do not establish whether adding UATL improves other compatible continuous-regression policies. These limitations do not invalidate the reported gains, but they leave the physical deployment benefits and transferability of the training objective unresolved.

Questions for authors
  1. Uncertainty-dependent weighting: Under the same heteroscedastic regression objective and tuning budget, how does UATL compare with a constant-weight temporal regularizer? Is there evidence that the predicted scales identify regions where relaxing smoothing improves control, rather than simply reflecting prediction difficulty?

  2. Scope across backbones: Is there evidence that the gains persist on another compatible continuous-regression VLA backbone? If such evidence is unavailable, please clarify the intended scope of applicability beyond the evaluated OpenVLA configuration.

  3. Execution drift: Beyond task success and qualitative trajectory smoothness, is there quantitative evidence that UATL reduces accumulated execution error? Real-robot trials or controlled execution perturbations measuring drift would directly address the stated motivation.

  4. CALVIN comparison: Table 5 reports an average length of 4.05 under ABC→D, while FLOWER and Unified Vision-Language-Action Model report 4.53 and 4.41, respectively. Could the authors explain relevant differences in training data, inputs, and evaluation settings, and clarify whether the claimed best performance is restricted to the methods selected in Table 5?

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

Related papers

Discussion (0)

Sign in to comment.