REA Runtime Lifts LIBERO-Pro Task Success to 95.9%, Versus 94.7% for the Best Single Agent

Large models reason well, but they are too slow and expensive to control a manipulator directly. Fast policies such as VLA models and whole-body controllers react within fractions of a second, but they generalize poorly. A group of authors from the National University of Singapore, UC Berkeley, the University of Tokyo, the Hong Kong University of Science and Technology and other institutions proposes REA (Routed Embodied Agent), a single runtime in which the "brain" and the "cerebellum" work asynchronously. Over 3000 test episodes across 60 LIBERO-Pro tasks, the system reached 95.9% success, ahead of the best single agent, Agents SDK + GPT-6 (94.7%).

Why the "agent + fast policy" pairing breaks down

The authors begin with the cost of a reasoning agent. A coding agent spent an average of 43 minutes on a single LIBERO task, and a Codex + GPT-6 pipeline consumed 89 million input tokens generating programs for the 60 LIBERO-Pro tasks. Using the model as a direct controller does not work either: when it emitted single tokens or Cartesian IK actions, it solved 1 of 1987 episodes and 0 of 600. Meanwhile, VLA models that exceed 90% on standard LIBERO drop noticeably on the perturbed LIBERO-Pro set.

Simply gluing the two levels together produces two typical problems. If the agent calls a fast policy and waits for the response, the robot sits idle: an advisor agent that the router waited on cost 2.6 seconds per episode, about 20% of control time. If the fast policy calls the agent only on error, that does not help either: a persistent error triggers the "brain" again and again, while most failures raise no exception at all. A router with an exception-based trigger escalated only 6 times across 590 episodes, even though 166 of them ended in failure. Rerunning is risky too: if the task is already solved, a rerun can undo the result.

Three levels and a "golden rule"

REA is built on one rule: no running policy waits for the brain. The system has three levels. System 2 is the agent, which generates policies offline or online and gives advice or issues verdicts. System 1 is the "embodied router": it does not produce actions itself, but decides which policy to launch and when to consult the agent. System 0 consists of fast policies: frozen programs, VLAs or whole-body controllers that return control at segment boundaries (for example, a program pauses every 100 steps).

The router relies on three mechanisms:

• Route memory. Before evaluation, each policy is run on separate development episodes, and its success rate and average step count are recorded. At test time, choosing a policy becomes a plain table lookup, with no agent call.
• Judge gate. When a policy returns control, a vision-model judge compares the initial and current frames against the instruction in parallel and decides whether the task is solved. If it is, the system stops and does not spoil a finished result. This catches "silent" failures that raise no exception. The judge's threshold and prompt were frozen on 216 labeled examples before testing.
• Escalation ladder with a shared budget. Subsequent policies launch from the current state, and all launches and brain calls draw on a single budget of steps and time, so an episode never exceeds the budget of a single policy.

What the experiments showed

According to the authors, REA was the fastest of the methods measured and came within 1% of a "retrospective oracle," which picks the best policy for each task after the fact. In the paper's text, REA's control time is 17% lower than that of Agents SDK + GPT-6 and 25% lower than that of the blocking advisor. The abstract states the reductions in control time and steps as 17% and 25%, respectively.

In a separate study on RoboTwin 2.0 (50 clean and 50 randomized episodes each), REA with VLA policies scored 84%, and with VLA and code it scored 86%. The strongest single VLA model, MEDex, reached 80%. On BEHAVIOR-1K, the system solved 27 of 50 episodes against 25 for an agent built on Agents SDK, but the authors themselves note that the difference is not statistically significant. In a test on 200 RoboTwin scenes, the result was 97/200 against 75/200 for Codex CLI. On a real Unitree G1 humanoid, the robot walked up to a chair and pushed it in 8 of 10 attempts.

Limitations

Most of the results come from simulators, and the real-world experiment is limited to ten attempts on a single task. The advantage on LIBERO-Pro is about 1.2 percentage points, so sample size and the paired statistical tests that the authors apply matter. The policy pool is still prepared offline by the agent, so the cost of generating programs does not go away. Still, the takeaway for practitioners is clear: take the expensive agent out of the control loop. Let it prepare policies in advance and act as a judge at segment boundaries, while a cheap router with a memory of past results handles policy selection.