Dune: A Zeroth-Order Optimizer with Time-Conditioned Historical Error Regression for VLA Test-Time Adaptation
Organizations: Statistics and Mathematics, University of Glasgow · Shenzhen University of Advanced Technology · The University of Hong Kong · Beijing Normal University ·
Abstract
Test-Time Adaptation (TTA) enables Vision-Language-Action (VLA) models to handle distribution shifts during deployment without source-domain retraining. However, backpropagation-based adaptation incurs heavy memory overhead on resource-constrained robots, whereas zeroth-order (ZO) methods suffer from high estimation variance under limited perturbation budgets. We observe that while ZO estimation enables low-memory adaptation with rapid initial gains, it often stalls or degrades over long-horizon updates due to noisy gradient estimates. To address these limitations, we propose Dune, a zeroth-order optimizer that stabilizes long-horizon adaptation by coupling online historical error prediction with current perturbation observations. Specifically, Dune learns a lightweight, time-conditioned error predictor from past optimization dynamics and fuses its predictions with instantaneous ZO estimates, achieving effective variance reduction without backpropagation or extra forward passes. Additionally, we theoretically characterize how estimation variance restricts convergence in high-dimensional, small-population regimes, motivating the use of historical statistics. Extensive experiments on LIBERO and real-world robots demonstrate that Dune achieves superior success rates under a low memory footprint, with ablations confirming improved closed-loop stability and post-disturbance recovery.
Figures & tables
| Method | Type | Spatial | Object | Goal | Long | Mean SR (%) |
| SGD | FO | 24.0 | 34.0 | 31.6 | 7.2 | 24.2 |
| SignSGD | FO | 62.0 | 58.2 | 63.0 | 24.8 | 52.0 |
| AdamW | FO | 74.6 | 85.4 | 88.8 | 64.4 | 78.3 |
| MeZO | ZO | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 |
| n-SPSA | ZO | 0.0 | 0.0 | 0.0 | 0.0 | 0.0 |
| Dust | ZO | 46.4 | 59.4 | 57.2 | 26.8 | 47.5 |
| Method | SR (%) | Mean completion time (s) |
|---|---|---|
| AdamW | 32.0 | 102 |
| Dune | 26.0 | 144 |
| Method | GPU Memory | Minimum Compatible GPU |
|---|---|---|
| AdamW | 18,578MiB | RTX 3090 (24 GB) |
| Dust | 10,748MiB | GTX 1080 Ti / GTX TITAN X |
| Dune | 10,898MiB | GTX 1080 Ti / GTX TITAN X |
Test-Time Adaptation of VLAs
Test-time adaptation of vision-language-action models helps robot policies adjust to new views, tasks, objects, and execution errors during deployment.
See all 63 papers