MoE VLAs spontaneously compose robot skills from data alone
What if robots could learn to break down tasks into primitives without explicit instruction?
A new paper accepted at the ICML 2026 Workshop on Compositional Learning explores how Mixture-of-Experts (MoE) architectures can unlock emergent compositional skills in robot learning. The authors—Shlok Shah, Rhiaan Jhaveri, Tharun Kumar Tiruppali Kalidoss, Chirayu Nimonkar, and Ishaan Javali—train Vision-Language-Action (VLA) models with a simplified MoE action head on expert demonstrations. Crucially, the model has no pre-specified notion of task decomposition or hierarchy; the routing mechanism must learn end-to-end how to sequence experts to solve tasks.
The results are striking: the MoE variant matches the task performance of a standard monolithic baseline, but its learned experts show clear specialization. The router implicitly learns to perform high-level sequencing, while each expert corresponds to a qualitatively distinct low-level behavior (e.g., grasping, pushing, turning). This suggests that the MoE framework can produce modular, interpretable robot policies that emerge purely from data—a step toward policy architectures that are both reusable and easier to debug.
- MoE VLA matches monolithic baseline performance while enabling expert specialization.
- Learned experts correspond to distinct low-level behaviors like grasping or pushing, reused across tasks.
- Router implicitly learns high-level task decomposition without explicit supervision or predefined hierarchies.
Why It Matters
This work paves the way for modular, interpretable robot policies that can adapt quickly to new tasks using reusable primitives.