acceptodds
Under review as a conference paper at ICLR 2027

Vela: Scaling Vision-Language-Action Models with Compact Action Curve Parametrization

Abstract

Most vision-language-action models represent future motion as fixed-rate action chunks, tying temporal resolution and prediction horizon to a fixed output budget. This pointwise representation wastes capacity on highly correlated neighboring actions, leaves temporal continuity and smoothness to be learned implicitly, and forces a tradeoff between long-horizon coverage and the local precision required for contact-rich manipulation. To address these limitations, we introduce Vela, a vision-language-action foundation model that represents future robot behavior as continuous trajectories. Vela combines a compact spline-based action representation with motion-dependent temporal support and a shared action interface for heterogeneous embodiments, allowing a fixed output budget to adapt its temporal resolution across motions. We pretrain Vela on large-scale multi-embodiment robot data and evaluate it on LIBERO-X, EBench, and two real-world long-horizon tasks, egg-cake cooking and potato shredding, obtaining promising results across simulation and physical manipulation. These results highlight the potential of continuous action representations as a foundation for future embodied foundation models.

Primary area:
applications to robotics, autonomy, planning
Keywords:
embodied intelligence, vision-language-action model

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

This paper presents Vela, a vision-language-action model that represents future actions using a fixed number of cubic B-spline control points and a predicted temporal horizon. A noise-aware fitting criterion constructs adaptive-horizon supervision from demonstrations, while decoded-trajectory flow matching weights control-point errors by their effects on the resulting action trajectories. The method is evaluated on LIBERO-X, EBench, and two real-world manipulation tasks.

Critical strengths
  1. The representation is well motivated. It enables the model to allocate a fixed representation budget adaptively across different action segments.

  2. The experiments are extensive, and the pretrained version demonstrates the scalability of the method.

Critical weaknesses
  1. Table 4 does not specify the fixed horizon used. The authors should also experiment with different fixed horizons to determine whether adaptive temporal support consistently provides an improvement.
Questions for authors
  1. Which robot embodiments are included in the pretraining data?
  2. Could the authors provide uncertainty estimates or statistical significance tests for the small gains? For example, the improvement in Table 5 is only around one percentage point, making it difficult to tell whether it is robust to training and evaluation variability.
3 comments
AuthorsOpenReview ↗

Responses to Reviewer jMxd Weakness 1

Answer to Weakness 1: Fixed-horizon specification and comparison across horizons

Thank you for pointing out this omission. The fixed horizon in the original Table 4 is H=25. To examine whether adaptive temporal support consistently improves over alternative fixed choices, we evaluated H∈{16,20,25,30,35,40} on LIBERO-X. All fixed-H variants are post-trained from the same Vela pretrained checkpoint as the adaptive model, with the same post-training budget and execution settings. Success rates (%):

HL1L2L3L4L5Avg.
1665.553.742.929.623.443.0
2065.854.142.529.123.943.1
2566.454.342.828.124.443.2
3065.753.041.828.625.642.9
3565.353.041.628.924.942.7
4064.552.240.728.123.741.8
Adaptive67.255.944.931.527.145.3
  1. Best fixed horizon. Among the six fixed configurations, H=25 achieves the best average (43.2%). Adaptive H reaches 45.3% (+2.1 points) and exceeds every fixed configuration at each of the five levels; relative to H=25, the gains at L1–L5 are 0.8, 1.6, 2.1, 3.4, and 2.7 points.
  2. No single best fixed horizon. H=25 performs best at L1–L2, H=16 at L3–L4, and H=30 at L5. Adaptive H outperforms all of them, supporting motion-dependent temporal coverage over a single global horizon.
  3. Robustness to the candidate range. Narrowing the adaptive candidate set to {16,…,30} or {16,…,35} gives 43.7% and 44.4%, both still above the best fixed horizon (43.2%).

Consistency under a different initialization. We also ran the same comparison from the original π₀.₅ checkpoint, which was pretrained with pointwise actions and never exposed to adaptive horizons. Both variants use Vela's spline representation (N=12) with the same post-training budget and execution settings, differing only in fixed versus adaptive horizon:

Model (from the π₀.₅ checkpoint)L1L2L3L4L5Avg.
π₀.₅-Vela, fixed H=2564.750.338.825.921.340.2
π₀.₅-Vela, adaptive H66.153.540.727.322.542.0

Adaptive H again improves the average by 1.8 points and is better at all five levels. The improvement therefore holds against the best fixed horizon in the sweep, across candidate ranges, and under both Vela and pointwise π₀.₅ initialization.

We have added the sweep to Sec. 4.4 (“Adaptive versus tuned fixed horizons,” Table 4), with the additional initialization and candidate-range controls in Appendix E.2 (Tables 7–8).

AuthorsOpenReview ↗

Responses to Reviewer jMxd Question 1 and 2

Answer to Question 1: Robot embodiments in the pretraining data

Thank you for the question. The main robot embodiments are AgiBot G1 and G2, Franka/Panda (single- and dual-arm), ARX X5, AgileX PiPER/ALOHA, UR5, and YAM. Smaller contributions, mainly from Open X-Embodiment (OXE), include Google Robot, WidowX, FANUC Mate, and DLR EDAN.

We have added this list to Appendix E.1, in the “Robot embodiments” paragraph.

Answer to Question 2: Uncertainty of the loss-ablation results

1. How we measured variability. For each loss configuration in Table 5, we ran three additional training seeds. Together with the original run, this gives four training runs per configuration (n=4). For each run, we record its LIBERO-X average success rate. The table reports the mean ± standard deviation (SD) of these four values, showing how the results vary across training seeds:

Geometry objectiveAvg. success rate (%, mean ± SD; n=4)
CP-FM44.3 ± 0.3
CP-FM + DT-FM (equal weighting)45.0 ± 0.3
DT-FM45.4 ± 0.2

2. What the results show. DT-FM has the highest mean success rate, 1.1 percentage points above CP-FM. Combining the two losses does not improve the mean over DT-FM alone (45.0% vs. 45.4%). The SDs quantify the observed variation across training seeds; we do not claim statistical significance or equivalence between the losses.

3. Why we use DT-FM. Table 5 tests how much the choice of loss matters for learning our trajectory representation. All three losses produce effective policies. DT-FM gives the highest observed mean, and adding CP-FM does not improve it, so we use DT-FM as the default. These results suggest that DT-FM's decoder-aware weighting is a useful refinement, while the action representation also works well without it.

We have updated Table 5 and Sec. 4.4 (“Geometry supervision and uncertainty”), with the four-run protocol and original-run results documented in Appendix E.7 (“Loss uncertainty,” Table 13).

Thanks for the response. My concerns are addressed and I will keep the current positive score.

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

This paper presents Vela, a VLA model that predicts spline control points and a temporal horizon instead of pointwise action chunks. It uses a noise-aware criterion to construct adaptive-horizon training targets and is pretrained on around 40,000 hours of robot data. The authors report improved overall performance after fine-tuning on two simulation benchmarks, along with higher subtask success rates on two real-world tasks.

Critical strengths
  1. The paper addresses an interesting limitation of fixed-rate action chunks. Using a fixed number of control points to describe motions of different durations is intuitive and offers a useful way to balance temporal coverage and local precision.

  2. The study includes large-scale pretraining and interesting real-world tasks involving bimanual coordination, tool use, and contact-rich manipulation.

Critical weaknesses
  1. Since the predicted horizon determines how the control points map to execution time, I am concerned about sensitivity to horizon selection. The fixed-versus-adaptive comparison is useful, but the fixed horizon used in Table 4 is not specified. How does performance change with different candidate horizon ranges, and does the adaptive version outperform a well-tuned fixed-horizon baseline?

  2. The noise-aware horizon criterion is plausible, but it is unclear how reliably the noise estimate distinguishes incidental demonstration variation from genuine fine-scale motion. Second differences can reflect both noise and real motion, especially during contact or rapid corrections. More analysis of the noise estimator and sensitivity to the fitting thresholds would strengthen this part of the method.

  3. The pretraining data and procedure are not matched between full Vela and π0.5, making it difficult to isolate the gains from trajectory-space modeling. A pointwise model pretrained on the same data with matched initialization and training budget would be an important control. The same-checkpoint π0.5-Vela comparison is helpful, but its gains are not uniform: EBench long-horizon success decreases from 34.1% for π0.5 to 28.5% for π0.5-Vela.

  4. Given the scale of pretraining, I would like to see clearer evidence of its transfer and data-efficiency benefits. The current simulation results follow full-parameter, benchmark-specific fine-tuning: 30K steps on LIBERO-X and 200K steps on EBench. These results show useful downstream performance, but do not establish whether pretraining reduces the need for task-specific adaptation. Few-shot or reduced-budget fine-tuning, or zero-shot evaluations where appropriate, would help clarify this.

  5. Spline-based action representations have already been studied in BEAST and Spline Policy. ABPolicy also uses flow matching in B-spline control-point space, while B-spline Policy includes adaptive fitting and temporal rescaling. Vela acknowledges these works but does not directly compare against them. Post-hoc spline fitting is not a sufficient substitute for a policy trained to predict spline parameters. Comparisons with the closest methods under matched backbones and training settings would help clarify the added value of Vela’s noise-aware horizon selection beyond existing spline representations.

Questions for authors

See weakness

7 comments
AuthorsOpenReview ↗

Responses to Reviewer miGd Weakness 1

Answer to Weakness 1: Horizon-range sensitivity and comparison with tuned fixed horizons

Fixed-horizon comparison. The fixed-horizon setting in the original Table 4 is H=25. To assess whether the adaptive policy outperforms a tuned fixed-horizon baseline, we evaluated six fixed horizons on LIBERO-X, all post-trained from the same Vela pretrained checkpoint with matched training budgets and execution settings:

HL1L2L3L4L5Avg.
1665.553.742.929.623.443.0
2065.854.142.529.123.943.1
2566.454.342.828.124.443.2
3065.753.041.828.625.642.9
3565.353.041.628.924.942.7
4064.552.240.728.123.741.8
Adaptive67.255.944.931.527.145.3

The best fixed horizon tested is H=25, with 43.2% average success. Adaptive H achieves 45.3%, improving by 2.1 percentage points, and outperforms every tested fixed horizon at each evaluation level. The best fixed horizon also varies across levels: H=25 for L1–L2, H=16 for L3–L4, and H=30 for L5, supporting the benefit of adapting temporal coverage rather than imposing a single global horizon.

Additional comparison from the π₀.₅ checkpoint. To test whether the benefit of adaptive H also holds without adaptive-horizon pretraining, we compare fixed H=25 and adaptive H starting from the same original π₀.₅ checkpoint, pretrained with pointwise actions. Both variants use Vela's spline representation (N=12), with identical post-training budgets and execution settings; H=25 is taken from the fixed-horizon sweep above. This comparison controls for pretrained initialization and post-training configuration, isolating the effect of fixed versus adaptive horizon selection:

Model (post-trained from the π₀.₅ checkpoint)L1L2L3L4L5Avg.
π₀.₅-Vela, fixed H=2564.750.338.825.921.340.2
π₀.₅-Vela, adaptive H66.153.540.727.322.542.0

Adaptive H improves average success from 40.2% to 42.0% (+1.8 percentage points), with gains at all five levels.

Candidate-range sensitivity. Offline reconstruction analysis showed reduced fitting fidelity beyond 40 steps with N=12, motivating this upper bound. We evaluated three candidate ranges by regenerating training targets and post-training from the same Vela checkpoint under identical settings:

Candidate set ℋL1L2L3L4L5Avg.
{16,…,30}65.454.543.130.025.743.7
{16,…,35}66.354.843.830.626.344.4
{16,…,40}67.255.944.931.527.145.3

The default range {16,…,40} achieves the highest average success (45.3%). All three ranges outperform the best tested fixed horizon (43.2%), supporting the benefit of adaptive H across the tested ranges.

We have added the sweep to Sec. 4.4 (“Adaptive versus tuned fixed horizons,” Table 4), with initialization and candidate-range comparisons in Appendix E.2 (Tables 7–8).

AuthorsOpenReview ↗

Responses to Reviewer miGd Weakness 2

Answer to Weakness 2: Noise-scale reliability and fitting-threshold sensitivity

The key concern is whether our method smooths away useful small movements along with noise. We explain how the noise estimate is used, how we chose the thresholds, and what our reconstruction tests show.

1. How the noise estimate is used. Second differences can contain both noise and real motion. We estimate a reference noise scale once for each data source and action coordinate, and keep it fixed during horizon selection. A rapid correction in the current window therefore does not automatically raise its allowed error. For each candidate H, we fit a spline and check its reconstruction error using the normalized criterion in Eq. (9). If a long horizon fails this check, we examine shorter candidates. Gripper changes are checked separately using η.

2. Why we chose κ=2.5 and η=0.05. We tested all 15 combinations of κ∈{1.2,1.5,2.0,2.5,3.0} and η∈{0.02,0.05,0.10} on 50,000 randomly sampled pretraining trajectories, using N=12 and the sparse candidate subset {16,20,25,30,35,40}.

  1. κ controls how much fitting error is allowed. Increasing κ from 2.0 to 2.5 and 3.0 raises the fraction of windows assigned H=40 from approximately 55% to 66% and 72%. For windows whose starting point is more than 30 frames from any gripper transition, increasing κ from 2.0 to 2.5 reduces short-horizon labels (H<25) from 32% to 21%. We choose 2.5 as a compromise between strict fitting and excessive shortening of horizons.
  2. The tested η values give the same labels. Gripper reconstruction errors cluster near 0 or approximately 0.19, so all three tolerances distinguish these two groups. We use the middle value, η=0.05.

These thresholds are shared across data sources, with each source using its own normalization and noise scale.

3. Whether small corrective movements are preserved. We constructed translation trajectories with known clean references, added millimeter-scale corrections and noise, and measured how well the spline recovered the clean motion. We used κ=2.5, η=0.05, N=12, and ℋ={16,…,40}, with noise set to the calibration floor of one data source. This test uses 5 Hz sampling (0.2 s per frame) and 32 random correction positions per condition. Waveform retention measures the recovered correction amplitude by projection onto the clean correction waveform; 100% means its amplitude is preserved, and values slightly above 100% indicate mild amplification.

  1. Noise decreases: translation MSE relative to the clean reference falls by 69.24% for white noise and 31.93% for temporally correlated noise.
  2. Small position changes are preserved: when the trajectory moves by 1 mm and holds the new position, the recovered amplitude stays close to 100%, even with H=40:
Transition durationWaveform retentionSteady-state position errorTransition RMSE
2 frames (0.4 s)99.65%0.014 mm0.148 mm
4 frames (0.8 s)99.84%0.013 mm0.103 mm
8 frames (1.6 s)99.73%0.014 mm0.040 mm
16 frames (3.2 s)100.70%0.016 mm0.028 mm
  1. Brief out-and-back movements are also largely preserved: a 1 mm return pulse retains 85.16% of its amplitude at 4 frames (0.8 s) and 92.35% at 8 frames (1.6 s).

Adaptive H adds flexibility when a long horizon fails the fitting check: shorter candidates let the same 12 control points describe a shorter interval, helping preserve rapid local changes. Together, these controlled tests show that the default settings reduce noise while retaining the tested millimeter-scale corrective movements.

We have added these analyses to Sec. 4.4 (“Noise calibration and fine-motion retention”) and Appendix E.4 (“Threshold selection” and “Controlled reconstruction,” Table 10 and Figs. 10–11).

AuthorsOpenReview ↗

Responses to Reviewer miGd Weakness 3 and 4

Answer to Weakness 3: Matched pretraining control and uneven post-training gains

This question raises two points: whether the gains come from the action representation when pretraining is matched, and why π₀.₅-Vela performs worse on EBench Long-Horizon tasks. We address them separately.

1. What happens when pretraining is matched. We agree that the original comparison between full Vela and π₀.₅ does not separate the effect of the representation from differences in pretraining data. We started a matched pretraining experiment after submission and have now completed it.

The baseline uses the native π₀.₅ architecture with pointwise actions. It starts from the same pretrained PaliGemma VLM weights as Vela, without loading pretrained π₀.₅ policy weights. Both models use the same approximately 40k hours of data, batch size, learning-rate schedule, and 300k optimization steps. Downstream fine-tuning and evaluation settings are also matched, including an execution cap of 25 actions before replanning. The only variable is the action representation: pointwise action chunks versus Vela's control points and horizon.

Model (LIBERO-X)L1L2L3L4L5Avg.
π₀.₅ (same-data pretraining)65.352.639.426.519.240.6
Vela67.255.944.931.527.145.3
Improvement (percentage points)+1.9+3.3+5.5+5.0+7.9+4.7

Vela improves average success by 4.7 percentage points, with gains at all five levels. The largest gains are at the harder levels, L3–L5. With data, initialization, and training settings matched, these results support the benefit of trajectory-space modeling.

2. How to interpret the EBench results. We agree that π₀.₅-Vela does not improve every metric. The table below reproduces the full breakdown from Table 2 on the official EBench test set of 510 episodes. Each entry gives success rate (%) / task-progress Score:

ModelTable TopSimple PnPLong-HorizonOverall
π₀.₅29.3 / 4556.5 / 6134.1 / 5541.4 / 54
π₀.₅-Vela34.3 / 5459.0 / 6328.5 / 5542.0 / 58
Full Vela44.3 / 6463.0 / 6739.1 / 6549.7 / 66
  1. Success rate decreases only in the Long-Horizon category. π₀.₅-Vela improves both success rate and Score on Table Top (+5.0% / +9 Score), Simple PnP (+2.5 / +2), and overall (+0.6 / +4).
  2. The reported Long-Horizon Score stays at 55. Success rate measures whether the whole task is completed; Score measures progress through the task. Thus, π₀.₅-Vela completes fewer full sequences, while its reported average progress is unchanged. The two metrics show a drop in full-task completion without a corresponding drop in average task progress.
  3. Full Vela improves on every task family, including Long-Horizon tasks, where it reaches 39.1% success and a Score of 65.

The matched control and EBench discussion are now included in Sec. 4.2 (Tables 1–2), with matched pretraining settings in Appendix E.1.

Answer to Weakness 4: Transfer under reduced fine-tuning budgets

To test whether Vela needs less downstream fine-tuning, we evaluated LIBERO-X performance after 5K, 10K, 20K, and 30K training steps. All models use the same downstream dataset and fine-tuning settings, including a global batch size of 256. At each step budget, they have therefore seen the same number of training samples, counting repeated sampling from the dataset.

The pointwise π₀.₅ baseline is the same-data model described in Weakness 3, pretrained on the same approximately 40k hours with matched initialization and training budget. π₀.₅-Vela starts from the original π₀.₅ checkpoint, as in the paper. Average success rates (%):

Model5K10K20K30K
π₀.₅ (same-data pretraining)30.233.637.240.6
π₀.₅-Vela30.834.137.642.0
Vela35.639.142.845.3
  1. Better performance with fewer fine-tuning steps. Vela at 5K already exceeds the pointwise baseline at 10K (35.6% vs. 33.6%), and Vela at 10K exceeds that baseline at 20K . At 20K, Vela reaches 42.8%, exceeding both baselines at 30K (40.6% and 42.0%) with one-third fewer updates.
  2. Better use of the same training sample budget. At the same number of steps, all models have seen the same number of samples, yet Vela performs better at every budget. Its gains over the same-data pointwise baseline range from 4.7 to 5.6 percentage points. This shows that Vela makes more effective use of the samples presented during fine-tuning.

These results show that Vela reaches stronger downstream performance with fewer fine-tuning steps and uses the available training data more efficiently, measured by the number of samples seen during training.

We have added the budget comparison to Sec. 4.2 ( Fig. 2), with exact results and sample-budget details in Appendix E.3 (Table 9).

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

Vela is a vision-language-action model that represents future actions using cubic B-spline control points and a jointly predicted temporal horizon. Training targets are constructed by selecting motion-dependent horizons based on noise-calibrated fitting errors, and the resulting trajectory parameters are learned through flow matching. The model uses a shared action interface for heterogeneous robot data and is pretrained on approximately 40k hours of demonstrations. The authors evaluate Vela on LIBERO-X, EBench, and two real-world bimanual manipulation tasks, reporting improvements in overall performance.

Critical strengths
  1. A well-motivated approach to allocating trajectory capacity.

Predicting control points together with a temporal horizon allows a fixed representation budget to cover longer smooth motions or shorter, more complex motions. The noise-calibrated fitting criterion provides a concrete basis for this allocation, while the separate treatment of gripper transitions accounts for their different characteristics.

  1. The experiments examine several important design choices.

The π0.5\pi_{0.5}-Vela comparison starts from the same checkpoint as the pointwise baseline, providing a useful assessment of the trajectory-based post-training approach. Ablations on post-hoc spline fitting, adaptive horizons, and training objectives further help identify the contributions of individual components. The real-world experiments also extend the evaluation to bimanual tool use and sustained physical contact.

Critical weaknesses
  1. The scope of “scaling” should be clarified.

The paper demonstrates pretraining on approximately 40k hours of heterogeneous robot data, but does not compare spline and pointwise action representations across data scales or examine how a fixed control-point budget accommodates increasing data diversity. The contribution is therefore primarily the use of trajectory representations for large-scale pretraining; whether their advantages grow with scale remains an open question. The title and corresponding claims should reflect this distinction. I do not expect additional scaling experiments during rebuttal.

  1. A pointwise baseline pretrained on the same data is missing.

Without a π0.5\pi_{0.5}-architecture baseline trained to predict raw action sequences on the same 40k-hour dataset under comparable settings, it is difficult to separate the benefits of the trajectory representation from those of the pretraining data and training configuration. The π0.5\pi_{0.5}-Vela comparison provides a useful control at the post-training stage, but does not resolve this question for pretraining. Given the computational cost and rebuttal timeline, I do not request this experiment, but the limitation should be acknowledged when interpreting the results.

Questions for authors
  1. Cross-chunk continuity and asynchronous execution need clarification.

Spline smoothness holds within each predicted segment and does not automatically ensure continuity across chunk boundaries, particularly when the predicted horizon HH varies. It is also unclear how Vela would accommodate methods such as RTC, which constrain action prefixes directly, whereas Vela predicts control points and a horizon. Please clarify whether the current system executes synchronously or asynchronously, how chunk boundaries are handled, and how RTC-style prefix alignment and inference-delay handling could be implemented with this representation. A methodological explanation is sufficient; additional asynchronous execution experiments are not required.

  1. The overhead of online target construction should be quantified.

Each training sample requires evaluating multiple candidate horizons, fitting control points, and computing residuals. The appendix explains that basis matrices and pseudoinverses are precomputed and candidate fits are batched, but does not quantify the computational cost or runtime overhead. Please provide a complexity analysis in terms of the number of candidate horizons, control points, and action dimensions, along with timing or throughput statistics showing the impact on training.

5 comments
AuthorsOpenReview ↗

Response to Reviewer PoMk Weakness 1 and 2

Answer to Weakness 1: Scope of scaling

Thank you for pointing out the need to clarify the scope of “scaling.” Our main claim is that compact action-curve representations are effective for large-scale pretraining on heterogeneous robot data. We clarify our contribution in two respects.

1. A comparison using the same pretraining data and training configuration. After submission, we completed an additional experiment using the native π₀.₅ architecture with its pointwise action representation. Like Vela, this baseline is initialized from the same pretrained PaliGemma VLM weights, without loading any pretrained π₀.₅ policy weights, and is pretrained on the same approximately 40k hours of data with identical batch size, learning-rate, and number of training steps (300k). Downstream fine-tuning and evaluation also use the same settings. The only difference is the action representation: pointwise action chunks versus Vela’s representation. Downstream LIBERO-X success rates (%) are shown below; improvements are in percentage points.

ModelL1L2L3L4L5Avg.
π₀.₅ (same-data pretraining)65.352.639.426.519.240.6
Vela67.255.944.931.527.145.3
Improvement+1.9+3.3+5.5+5.0+7.9+4.7

Vela improves average success by 4.7 percentage points and performs better at every level. This comparison supports the benefit of our action representation for large-scale pretraining: the advantage persists when both methods use the same pretraining data and training configuration. It provides evidence of effectiveness at this scale.

2. A fixed control-point budget with adaptive temporal coverage. Although Vela uses a fixed number of control points (N = 12), the horizon H covered by these control points is selected adaptively according to the motion:

  1. When an action segment is complex, for example during contact or rapid corrections, the fixed control points represent a shorter trajectory, providing finer temporal resolution.
  2. When an action segment is simple and smooth, for example during free-space transport, the same control points represent a longer trajectory, avoiding redundant capacity on highly correlated neighboring actions.

Thus, the representation size stays fixed while its temporal resolution adapts to motion complexity. This makes the representation flexible across motions of widely varying complexity, while remaining compact and low-cost in output size.

Together, these points support our claim of effective large-scale pretraining with compact action curves; whether the relative advantage grows with data scale remains an open question.

To make this scope explicit, we have revised the title to “Vela: Large-Scale Vision-Language-Action Pretraining with Compact Action Curve Parametrization” and aligned the corresponding claims with effectiveness at large scale rather than an increasing advantage across data scales.

We have revised the title and scope statements in the Abstract and Sec. 5, and added the matched pretraining results to Sec. 4.2 (“Simulation Evaluation on LIBERO-X,” Table 1).

Answer to Weakness 2: A pointwise baseline pretrained on the same data

We agree that the original π₀.₅-Vela post-training comparison alone does not resolve this pretraining confound. We initiated the matched pointwise pretraining experiment shortly after submission and have now completed it, as detailed in our response to Weakness 1. With pretraining data, initialization, training budget, and downstream settings matched, Vela achieves 45.3% average success versus 40.6% for the pointwise baseline. This 4.7-point improvement directly addresses the missing control and supports the benefit of trajectory-space modeling under matched pretraining conditions.

This control is now described in Sec. 4.1 (the paragraph beginning “We additionally pretrain”), with results in Sec. 4.2 (“Simulation Evaluation on LIBERO-X,” Table 1) and settings in Appendix E.1.

AuthorsOpenReview ↗

Response to Reviewer PoMk Question 1

Answer to Question 1: Cross-chunk continuity and asynchronous execution

Vela changes how the network represents actions, but keeps the controller interface unchanged. We sample the continuous action curve at the controller's discrete timesteps and send the resulting action sequence to the controller. Existing asynchronous methods that operate on these sequences therefore remain applicable. Methods such as RTC, which also guide action generation, can be adapted through the spline decoder. We explain the execution setup, RTC adaptation, and timing below.

1. How the robot executes the chunks. Our real-robot demonstration videos use synchronous execution. After executing a chunk, the policy replans from the current observation and sends the newly decoded actions directly to the controller. We apply no special smoothing or blending at chunk boundaries, so the videos show the model's output without these adjustments. The spline guarantees smoothness within each chunk; it does not explicitly enforce continuity between chunks.

We also tested RTC-style asynchronous execution on the real-robot potato-shredding task. The mean success rate across its four subtasks was 87.5%, matching the synchronous result in Fig. 9.

2. How we adapt Vela to RTC. We apply RTC-style prefix alignment directly to the decoded actions. This encourages the overlapping part of the new chunk to match the previous chunk:

Lalign=∑i∈Owi∥(BH^C^)i−Aiprev∥22.\mathcal{L}_{\mathrm{align}} = \sum_{i \in \mathcal{O}} w_i \lVert (B_{\widehat{H}} \widehat{C})_i - A_i^{\mathrm{prev}} \rVert_2^2 .

Here, BH^C^B_{\widehat H}\widehat C is the decoded estimate of the trajectory, and O\mathcal O identifies timestamps shared with the previous chunk. The weights wiw_i give priority to the committed prefix. For a fixed horizon, the decoder is linear. The alignment gradients therefore pass directly to the control points, allowing RTC-style guidance to be applied through the decoder.

3. How we handle different horizons and inference delay. All decoded actions use the same controller time grid. We therefore align consecutive chunks by their timestamps, even when their horizons H differ. Changing H changes how many timesteps overlap, but does not change the alignment rule. While the next chunk is being generated, the robot continues executing the previous chunk. These actions form the committed prefix. When the new chunk is ready, we skip its actions corresponding to the elapsed inference time and start from the matching time offset.

We have clarified the controller interface in Sec. 3.2 (the paragraph following Eq. (5)) and added the execution details and RTC success rate to Appendix E.6.

AuthorsOpenReview ↗

Response to Reviewer PoMk Question 2

Answer to Question 2: Computational cost of online target construction

We measured the cost of constructing targets during training. Below, we explain what is computed in advance, report the training overhead, and explain why it differs from the time measured for target construction alone.

1. What is computed in advance and during training. As described in Sec. 3.3 (“On-the-fly target construction”) and Algorithm 1 in Appendix A.1, supervision targets are constructed during training for each sampled window. Only quantities that do not depend on individual samples are computed in advance:

  1. The spline basis matrices and their pseudoinverses depend only on the candidate horizons and fixed spline configuration. We compute them once and reuse them throughout training (Appendix A.1, “Computation”).
  2. Each data source's normalization statistics and noise scales are also computed in advance (Sec. 3.4, “Shared action space”; Appendix B.1).

For each sampled window, we fit control points for all candidate horizons together, using matrix multiplication in one batched pass. We then check the reconstruction errors against the fitting criteria, mask out candidates that fail, and select the longest passing horizon through a masked reduction (Appendix A.1). This requires no matrix inversion for individual samples and no iterative spline optimization. The complexity per GPU and per batch is

O(bND∑H∈HH),O\left( b N D \sum_{H \in \mathcal{H}} H \right),

where bb is the per-GPU batch size, NN the number of control points, DD the action dimensionality, and H\mathcal{H} the candidate-horizon set.

2. How much time this adds to training. We compared two modes: our default mode constructs targets during training; a reference mode constructs targets for the same samples in advance and loads them during training. The reference mode is used only for this timing comparison. Both runs use the same pretraining configuration, are otherwise identical, and produce identical targets. We average the timings over the stable training phase.

Target construction modeMean training-step timeGlobal throughput
On-the-fly construction (default)2000 ms768 samples/s
Pre-constructed targets (reference)1996 ms769.5 samples/s

Online construction adds only 4 ms per step, approximately 0.2% of the total training-step time.

3. How long the individual operations take. In the same profiling run, we also timed the target-construction operations separately. The following times are per GPU and per batch:

ComponentMean time
CPU horizon selection20 ms
GPU control-point construction and span assignment0.25 ms
Total serial target-construction wall time20.4 ms

The 20.4 ms measures how long target construction takes on its own. The 4 ms measures how much longer a complete training step takes because of it. These are different because DataLoader workers select horizons for upcoming batches while the GPU trains on the current batch. Most of the CPU work therefore overlaps with GPU training instead of adding waiting time. GPU control-point construction runs before the forward pass and takes only 0.25 ms.

Thus, the measured training overhead is about 0.2% in our pretraining setup. Fitting and horizon search are needed only during training. At deployment, the policy predicts the control points and horizon directly.

We have added the timing summary to Sec. 4.4 (“Online construction cost”) and the complexity and timing breakdown to Appendix E.5 (Table 11).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Title Revision

Inspired by Reviewer PoMk’s Comment on clarifying the scope of “scaling,” we have revised the title to “Vela: Large-Scale Vision-Language-Action Pretraining with Compact Action Curve Parametrization.” The revised manuscript reflects this change.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.