BATCH: Behavior Alignment for Cross-Embodiment Learning from Humans
Organizations: Stanford University
Abstract
Human demonstrations provide a scalable source of diverse manipulation data, but leveraging them for robot learning remains challenging due to differences in embodiment, viewpoint, action spaces, and temporal structure. We introduce BATCH, a cross-embodiment learning framework that explicitly aligns human and robot behaviors. BATCH learns self-supervised temporal representations with a cycle-consistency objective to establish cross-embodiment correspondences without manual annotation. We further introduce two complementary training objectives that leverage these correspondences within an autoregressive videoaction diffusion model: behavior-aligned co-training and human-to-robot style transfer. On real-world manipulation tasks, BATCH improves generalization to unseen task configurations compared with training on unaligned data and prior cross-embodiment baselines, with additional gains from the style-transfer objective. Our results demonstrate that explicit behavior alignment enables more effective transfer from human demonstrations to robot policies.
Figures & tables
| Object in Container | Cup on Saucer | |||||||
| Overlap (%) | Prog. err. | Overlap (%) | Prog. err. | |||||
| Method | ID | OOD | ID | OOD | ID | OOD | ID | OOD |
| GTCC variants | ||||||||
| GTCC + DTW | 69.0 | 57.8 | .061 | .098 | 64.8 | 71.7 | .088 | .075 |
| DTW | 61.8 | 24.7 | .080 | .231 | 61.9 | 70.8 | .092 | .085 |
| rel. actions | 68.9 | 54.9 | .060 | .104 | 61.8 | 62.7 | .100 | .123 |
Appendix figures & tables6 assets
Supplementary material from the paper’s appendix.
Appendix
| Object in Container | Cup on Saucer | |
| Embodiment | single arm | bimanual |
| State / action dim | 8 (pos 3, quat 4, gripper 1) | 16 ( ) |
| Human action dim | 7 (one hand) | 14 (two hands) |
| Wrist cameras | 1 | 2 |
| Conditioning canvas | ||
| Clean tokens / frame | 525 | 750 |
Cross-Embodiment Robot Learning
Cross-embodiment robot learning transfers skills and policies across robots with different bodies, sensors, dynamics, and action spaces.
See all 374 papers