Generalist vision-language-action (VLA) models can perform many open-world manipulation tasks, but they lack two abilities needed for real-world assistance: personalization and reliable execution. An instruction like "make me a sandwich" requires grounding subtask choices in persistent user preferences, not a generic solution, and deployment demands explainable safety assurance that prevents hazards and recovers from failures. The CarryOn system, described in a recent arXiv preprint, is a dual-loop harness designed to add these capabilities to a frozen VLA policy.

The planning loop maintains user-level and scene-level context that is continually updated against new observations and revised through interaction, enabling personalized task selection across conversations and changing scenes. The execution loop actively supervises policy rollouts, shielding actions from predicted hazardous contact. When execution fails, the system calls on reusable control-based recovery skills that restore task progress and are refined through reflection on earlier failures.

In experiments on a real robot and in simulation, CarryOn dispatched preference-consistent subtasks across different users and scenes. It also reduced the violation rate of an already safety-finetuned VLA from 52.6% to 6.1%, and increased recovery success from 13% to 80% when refined skills were reused. The authors position CarryOn as an augmentation layer rather than a replacement for the underlying VLA, keeping the policy frozen while adding the missing structure for personalized, reliable assistance.