REA Runtime Hits 95.9% on LIBERO-Pro, Cuts Control Steps 25%

Across 3,000 test episodes on 60 LIBERO-Pro tasks, a new runtime called REA reaches 95.9% success, beating the best single-agent baseline at 94.7% while using 25% fewer control steps and 17% less control time. The authors, from the National University of Singapore's HPC-AI Lab, UC Berkeley, the University of Tokyo and three other labs, published the paper on Hugging Face.

REA splits an embodied system into a slow deliberative "brain" (foundation models), an embodied router, and reactive "cerebellum" policies like VLAs and whole-body controllers. The single rule, as the paper puts it, is that "no running policy waits on the brain." Instead of calling the foundation model after every failure, REA keeps a route memory that stores each policy's measured success offline, a boundary-triggered judge gate that only consults the brain between policy executions, and a shared-budget escalation ladder for bounded recovery.

On RoboTwin 2.0 (100 episodes), REA with VLA policies hits 84% and with VLA plus code 86%, against 80% for the strongest single VLA baseline MEDex. On a real Unitree G1 humanoid walk-to-push, REA scored 8 of 10 trials versus 5 of 10 for Agents SDK and 2 of 10 for ASPIRE, with p=0.023. On an AlohaMini2 block placement it went 10 of 10.

The route memory itself is the efficiency story. The authors measure a blocking advisor at 2.62 seconds per episode, roughly 20% of control time; precomputing route success means those seconds don't get spent at inference. In their development-cost table, route memory took 1,790 episodes and zero tokens, against 89.0M input tokens for Codex + GPT-6.

The paper is one of several embodied-research entries our research tracker has logged over the past week.