Executing a vision-language-action (VLA) policy is not just about generating actions; it also involves deciding when to keep going and when to step in. A new paper introduces ACRO, an actor-critic rollout orchestration layer designed to make those decisions explicit. The layer sits on top of a VLA policy and learns when an ongoing attempt is likely to succeed or when intervention is needed.

The core idea is to frame rollout orchestration as a sequential decision problem. An actor-critic architecture is used to evaluate the current rollout and decide whether to continue or to intervene, potentially improving the chances of success. The abstract does not include experimental results, so the reported effectiveness of ACRO is not yet detailed in this source.

The work addresses a practical gap in VLA deployment: knowing when to trust the policy and when to correct it. By adding a learned orchestration layer, ACRO offers a potential path toward more robust execution of embodied tasks, though further details on performance and benchmarks would be needed to assess its impact.