Correction
Frontiers in Neurorobotics ·
Volume 20 - 2026 | https://doi.org/10.3389/fnbot.2026.1959872
Correction: ST-HADP: Spatio-Temporal hierarchical attention diffusion policy for long-horizon generalizable bimanual visuomotor imitation
- Xukun Liu1
- Fengjuan Xie1*corresponding author
- Shibo Liu1
- Xu Sun1
- Zhenyu Liu1
- Guangning Li1
- Shenggang Wei1
- Northwest Institute of Mechanical and Electrical Engineering, Xianyang, Shaanxi, China
Abstract
Keywords
There was a mistake in the figures and tables as published. During the proofreading stage, we explicitly requested that the font sizes within the figures be kept slightly smaller than that of the main text, in line with our original manuscript. However, this typesetting instruction was not carried out during the final production process, which compromised the clarity and readability of the data. To address both issues, we have now revised the figure layouts alongside upgrading the resolution, ensuring proper data clarity and typographical consistency.
The corrected Figures 1–5 appear below.
Additionally, in the Tables 1, 2, 4, 5, certain content was not centered as per the author correction version approved during the proofing stage, which deviated from the intended formatting.
Table 1. Hyperparameter configuration for ST-HADP.
| Hyperparameter | Value | Hyperparameter | Value |
|---|---|---|---|
| Training | Model Architecture | ||
| Learning rate | 1e-4 | Observation steps | 3 |
| Batch size | 256 | Action steps | 6 |
| Optimizer | AdamW | Diffusion step embed dim | 64 |
| Weight decay | 1e-6 | Encoder output dim | 128 |
| LR scheduler | Cosine | Down dims | [512, 1,024, 2,048] |
| Seed | 42 | Kernel size | 5 |
| Diffusion model | Point cloud encoder | ||
| Train timesteps | 100 | Input channels | 3 |
| Inference steps | 10 | Output channels | 128 |
| Beta range | [1e-4, 2e-2] | PointNet type | Pointnet |
Table 2. Model complexity comparison.
| Task | Model | Initial memory(GB) | Peak memory(GB) | Memory increment(GB) | Avg. inference time (s) |
|---|---|---|---|---|---|
| Put object cabinet | DP3 | 2.292 | 2.371 | 0.079 | 1.4399 |
| OURS | 2.296 | 2.376 | 0.080 | 1.5456 |
Table 4. Success rates on Aloha Agilex across different methods and tasks.
| Method | Turn switch | Place can basket | Press stapler | Place empty cup | Pick dual bottles |
|---|---|---|---|---|---|
| RDT | 35% ± 4% | 19% ± 3% | 41% ± 3% | 56% ± 2% | 42% ± 3% |
| Pi0 | 27% ± 3% | 41% ± 4% | 62% ± 3% | 37% ± 3% | 57% ± 4% |
| ACT | 5% ± 2% | 1% ± 0% | 31% ± 3% | 59% ± 4% | 31% ± 4% |
| DP | 36% ± 3% | 18% ± 3% | 6% ± 3% | 37% ± 2% | 24% ± 2% |
| DP3 | 46% ± 4% | 65% ± 3% | 69% ± 4% | 65% ± 3% | 60% ± 4% |
| ST-HADP | 61% ± 3% | 75% ± 3% | 76% ± 2% | 77% ± 2% | 70% ± 3% |
Table 5. Ablation studies on the “place can basket” task across six dual-arm robot configurations.
| Variant | Franka-Franka | ARX-Franka | UR5-ARX | Aloha-Aloha | ARX-ARX | Piper-Franka |
|---|---|---|---|---|---|---|
| w/o SAM | 62% ± 3% | 65% ± 2% | 66% ± 4% | 68% ± 3% | 64% ± 3% | 63% ± 2% |
| w/o TAM | 61% ± 2% | 64% ± 3% | 65% ± 3% | 67% ± 4% | 63% ± 4% | 62% ± 3% |
| w/o Ltemp | 65% ± 3% | 68% ± 2% | 69% ± 3% | 71% ± 3% | 67% ± 3% | 66% ± 2% |
| w/o Lattn | 67% ± 2% | 70% ± 2% | 71% ± 3% | 72% ± 3% | 69% ± 3% | 69% ± 2% |
| Concat (replace FiLM) | 64% ± 3% | 66% ± 2% | 67% ± 3% | 69% ± 3% | 66% ± 3% | 65% ± 2% |
| Full ST-HADP | 70% ± 2% | 73% ± 2% | 74% ± 3% | 75% ± 3% | 71% ± 3% | 72% ± 2% |
The original version of this article has been updated.
Generative AI statement
Any alternative text (alt text) provided alongside figures in this article has been generated by Frontiers with the support of artificial intelligence and reasonable efforts have been made to ensure accuracy, including review by the authors wherever possible. If you identify any issues, please contact us.
Publisher's note
All claims expressed in this articler are solely those of the authors and do not necessarily represent those of their affiliated organizations, or those of the publisher, the editors and the reviewers. Any product that may be evaluated in this article, or claim that may be made by its manufacturer, is not guaranteed or endorsed by the publisher.
Footnotes
Approved by: Frontiers Editorial Office, Frontiers Media SA, Switzerland
Citation
Liu, X., et al. (2026). Correction: ST-HADP: Spatio-Temporal hierarchical attention diffusion policy for long-horizon generalizable bimanual visuomotor imitation. Frontiers in Neurorobotics, Volume 20. https://doi.org/
- Received
- Accepted
- Published
- Volume
- 20 - 2026
Copyright
Copyright © 2026 Liu, Xie, Liu, Sun, Liu, Li and Wei.
License: https://creativecommons.org/licenses/by/4.0/
*Correspondence: Fengjuan Xie
Article metrics
- Views
- 0
- Downloads
- 0
- Citations
- 0
- Shares
- 0