Event

Proposes Move-Then-Operate dual-expert VLA manipulation framework

1 reports · 1 sources

Haoming Xu et al. propose the Move-Then-Operate vision-language-action (VLA) framework, which explicitly splits robotic manipulation into two stages: coarse movement (move) and contact-critical interaction (operate), routed by a learnable phase selector to dual-expert policies. The paper has been accepted to ICML 2026, and the code has been released. On manipulation tasks in the RoboTwin2 benchmark, the method achieves an average success rate of 68.9%, which the authors state is 24% higher than the monolithic π₀ baseline.