BATCH: Cycle-Consistent Behavior Alignment for Cross-Embodiment Learning
Learn cross-embodiment correspondences by training self-supervised temporal representations with a cycle-consistency objective, then use the resulting alignments to co-train an autoregressive video-action diffusion model with behavior-aligned objectives and an additional human-to-robot style-transfer objective.
Discovered 2026-11-10 · Sources checked 2026-11-10
Why it matters
Human demonstrations are diverse but often cannot be used directly for robot learning when embodiments, viewpoints, action spaces, and temporal structure differ; explicit behavior alignment is used to bridge this mismatch without manual correspondence annotation.
What changed
The approach adds (1) self-supervised temporal representation learning with cycle-consistency to align human and robot behaviors, and (2) two training objectives—behavior-aligned co-training and human-to-robot style transfer—applied within an autoregressive video-action diffusion model using the learned correspondences.
Where it helps
- Transferring manipulation skills from human demonstrations to robots with different embodiments
- Learning from videos when manual alignment between human and robot actions is unavailable
- Training action-generation models using behavior-aligned cross-domain correspondences
- Improving robustness to unseen task configurations in manipulation tasks
- Adapting demonstration style to match robot execution characteristics
Evidence & reproduction
Reproduction not evaluated. A linked paper or repository does not establish independent reproduction.
Limitations & open questions
The method is described for cross-embodiment manipulation settings and depends on learning temporal correspondences via the specified cycle-consistency setup.
Related World Models papers
Sources
- BATCH: Behavior Alignment for Cross-Embodiment Learning from Humans ↗
arXiv · primary · Checked 2026-11-10
