- Rating
- 3(+1)
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper addresses VLA representations that fail to distinguish different control requirements under the same instruction. CSDR builds near and far relationships from demonstrated action prefixes. It verifies comparisons over a common valid interval and applies a finite-margin ranking loss. Detached weights emphasize action errors and ranking violations, while instruction-based sampling supplies more comparison candidates. The main strength is using existing control supervision without adding an inference module. The contribution lies in constructing and allocating these relationships, rather than introducing a new ranking loss. However, the smallest gains need stronger comparison with matched continued training, and some training settings are missing. SpatialVLA success increases from 38.54% to 56.51% in the reported simulation setting. Real-robot success rises from 650/800 to 692/800 trials. Gains on the stronger LIBERO baselines are smaller, so the empirical benefit varies across policies.
Critical strengths
-
A suitable control objective. Common-prefix verification avoids padded comparisons. Relative rankings and a finite margin provide supervision without forcing exact action distances into representation space.
-
Extensive alternative controls. Task-only continuation, sampling-only, metric-learning, and component ablations help distinguish CSDR from extra optimization or generic representation regularization.
-
Practical deployment and evaluation. Inference remains unchanged. Multiple policies, trajectory-split representation probes, and real-robot trials provide complementary evidence of usefulness.
Critical weaknesses
-
Small gains need matched uncertainty estimates. CSDR training repeats are provided, but equivalent task-only continuation repeats are missing. Gains over the original checkpoints are only 16 and 10 successes out of 2,000 trials.
-
Simulation weighting scales are unspecified. The action-error and ranking-violation scales are given for real robots but not the simulation implementations. Clipping and detachment do not remove their effect on supervision.
-
Coverage and added training cost are unclear. The fraction of anchors retaining valid triplets after all filters is unreported. Training time and peak memory also lack a task-only reference. The authors need to provide the detailed information for further consideration.
Questions for authors
I still have three questions:
-
How stable are the small LIBERO gains over matched task-only continuation across independent training seeds?
-
What values or estimation rules define the two weighting scales for each simulation implementation?
-
What fraction of anchors retains valid triplets, and what training cost is added over task-only continuation?
If the authors can address my concerns well, I will raise my score.
7 comments
Response 2: Weighting Scales and Valid Triplets
Thank you for your valuable feedback. Both weights use fixed scales: W_ERROR = STOP_GRAD(CLAMP(ERROR_i / ERROR_SCALE, 0, 1)); W_CONF = STOP_GRAD(CLAMP(VIOLATION / CONF_SCALE, 0, 1)). The component scales for control distances are estimated separately from the training set offline. A valid anchor is a window that retains at least one valid triplet after all filters; windows without a valid relationship still contribute to the task loss.
| Model / ERROR_i | ERROR_SCALE / CONF_SCALE | Valid-anchor proportion |
|---|---|---|
| OpenVLA-OFT: normalized-action MAE | 0.02 / 0.10 | 75.833% |
| TurboVLA: valid-action MAE | 0.50 / 0.10 | 75.950% |
| SpatialVLA: action-token error rate | 0.50 / 0.10 | 63.142% |
| StarVLA: MEAN(ABS((1−t) × (V_PRED−V_TARGET))) | 0.50 / 0.10 | 57.509% |
The weighting scales and valid-triplet information above have been added to Appendix A.4 of the paper.
Response 3: Training-Efficiency Comparison
Thank you for your valuable feedback. We conducted additional paired timing experiments for the four baselines on 10 A6000 GPUs, keeping the model, precision, batch size, optimizer, and data pipeline identical and changing only sampling organization and CSDR regularization. After 30 warm-up updates, we measured updates 31–40, including forward computation, backpropagation, and parameter updates, but excluding model loading and offline preparation. Each cell below reports “standard continuation / full CSDR,” in milliseconds per step. The timing components are mutually exclusive and sum to the complete step; token lengths are listed afterward.
CSDR relation computation adds 12.18–247.61 milliseconds per step, yet complete-step time decreases by 2.38%–10.49% in this experiment. Our analysis shows that same-instruction grouping reduces the mean padded length for OFT, SpatialVLA, and StarVLA, lowering computation and training time. Although TurboVLA's final length remains 21, its training pipeline first groups the instructions within a batch by three lengths, calls BERT separately for each group, and then combines the outputs. Same-instruction grouping can reduce the number of BERT calls for separate length groups, consistent with the approximately 25 ms reduction in forward time; its framework waiting time also decreases by approximately 33 milliseconds.
| Stage (ms/step) | OpenVLA-OFT | TurboVLA | SpatialVLA | StarVLA |
|---|---|---|---|---|
| Visible data-loading wait | 6921.263 / 6841.574 | 17.669 / 19.372 | 33.200 / 34.509 | 22.509 / 27.073 |
| Input processing and GPU transfer | 6.665 / 6.169 | 8.656 / 7.847 | 6.311 / 8.616 | 86.511 / 77.156 |
| Native model forward | 2528.523 / 2556.410 | 392.387 / 367.078 | 4048.183 / 3026.081 | 970.481 / 811.607 |
| Native task loss | 85.517 / 139.237 | 10.851 / 10.631 | 1010.949 / 1310.645 | 25.291 / 5.301 |
| CSDR relation computation (added) | 0.000 / 19.694 | 0.000 / 12.178 | 0.000 / 247.613 | 0.000 / 16.914 |
| Representation readout, regularization scheduling, and budget control | 2.406 / 9.337 | 3.267 / 9.370 | 27.916 / 36.879 | 2.829 / 7.134 |
| Backpropagation and gradient synchronization | 2777.529 / 2242.612 | 382.897 / 399.387 | 2450.512 / 2188.897 | 4124.391 / 3901.206 |
| Optimizer update | 35.466 / 34.811 | 16.021 / 16.799 | 36.262 / 35.018 | 1222.618 / 1217.557 |
| Training framework and other waiting | 57.938 / 45.618 | 85.503 / 52.752 | 581.786 / 446.970 | 223.574 / 200.526 |
| Complete step | 12415.307 / 11895.462 | 917.251 / 895.414 | 8195.119 / 7335.228 | 6678.204 / 6264.474 |
| Token length | OpenVLA-OFT | TurboVLA | SpatialVLA | StarVLA |
|---|---|---|---|---|
| Mean valid length | 85.833 / 85.833 | 14.387 / 14.387 | 280.967 / 280.967 | 119.768 / 119.768 |
| Mean padded length | 89.863 / 85.833 | 21.000 / 21.000 | 287.045 / 281.908 | 122.513 / 119.861 |
The training-efficiency analysis has been added to Section 4.7 and Appendix B.7 of the revised manuscript, with the revisions highlighted in blue. The revised manuscript also reports total training times for each baseline and compares update times when token lengths are identical. Please refer to the revised manuscript for further details.
Thank you again for your valuable feedback. We hope the above response addresses your concerns. If our reply has resolved your doubts, maybe you can consider increasing your Rating score. If you have any further questions, please let us know promptly. Thank you!
Response 1: Stability of Small Gains and Repeated Standard Continuation Training
Thank you for your valuable feedback. We summarized the results of three independent CSDR training runs under the same LIBERO evaluation configuration, as shown below. The sample standard deviations for the two baselines are 0.058 and 0.100 percentage points, respectively, indicating small variation across training runs.
We further reproduced Task-only with the same training seeds, budget, and evaluation initial scenes, and paired these runs with CSDR in the order shown in the table. The mean success rate across the three runs increases from 95.967% to 96.817% for OpenVLA-OFT and from 95.417% to 95.850% for TurboVLA; every paired comparison shows an improvement. The mean gains are 0.850 and 0.433 percentage points, respectively. The 95% paired t intervals computed from the three training differences are both above 0, supporting the stability of the gains across independent training repeats and further distinguishing CSDR's effect from simply increasing the number of training steps. These results demonstrate that CSDR provides stable improvements on LIBERO.
The LIBERO statistical analysis above has been incorporated into Table 1 of the revised manuscript, and the Task-only continuation-training repeats have been added to Appendix B.2. The revisions are highlighted in blue. Please refer to the revised manuscript for further details.