Event
Compositional Generalization in Diagnostic Sequence Robot Tasks
1 reports · 1 sources
Yixiao Wang et al. published a research paper analyzing compositional generalization in sequential robot manipulation tasks from the perspective of instruction-space coverage, decomposing the generalization gap into three sources: marginal instruction shift, instruction-composition shift, and context-action shift. The paper claims that exhaustively enumerating instruction tuples is not necessary; covering a structured subset of action-relevant dependencies suffices to achieve strong out-of-distribution performance. For semantically related tasks, effective coverage must capture relational structure rather than relying solely on factor diversity. Sparse-training failures often stem from instruction guidance rather than missing underlying skills; fine-tuning with just one demonstration per task can raise out-of-distribution success rate from 0.4% to 54.7%.
Report timeline
- Diagnosing Compositional Generalization in Sequential Robot Tasks
This paper studies compositional generalization in sequential robot manipulation tasks from the perspective of instruction-space coverage, decomposing the generalization gap into three sources: marginal instruction shift, instruction-compositional shift, and context-action shift. Results show that exhaustive instruction-tuple enumeration is unnecessary; a structured subset covering action-relevant dependencies can achieve strong out-of-distribution performance. Sparse-training failures often stem from instruction steering rather than missing low-level skills, and fine-tuning with just one demonstration per task raises out-of-distribution success from 0.4% to 54.7%.
arXiv cs.RO ·