A new paper proposes REA, a runtime that lets a slow reasoning model and fast robot policies run together without the fast ones waiting on the reasoning model.
Routing: A route memory stores each policy's measured success ahead of time, a judge gate consults the brain at policy boundaries, and a shared-budget ladder handles bounded recovery.
LIBERO-Pro: Across 3,000 test episodes on 60 tasks, REA reaches 95.9% success against 94.7% for the best single-agent baseline, with 25% fewer control steps and 17% less control time.
RoboTwin 2.0: In 100 episodes, REA scores 84% with VLA policies and 86% with VLA+code, against 80% for the strongest single VLA baseline.
Transfer: The paper reports it also works on BEHAVIOR-1K and a real Unitree G1 humanoid.
Sign in to comment.
TL;DR of the paper, generated automatically.
A new paper proposes REA, a runtime that lets a slow reasoning model and fast robot policies run together without the fast ones waiting on the reasoning model.