CarryOn: A Dual-Loop Harness for Personalized and Reliable Robot Assistance

CarryOn. A Dual-Loop Harness for Personalized and Reliable Robot Assistance. By Zixuan Wang and nine others, from the Hong Kong University of Science and Technology, the Chinese University of Hong Kong, and one other institution. Published on the fifteenth of October. Its code repository is a placeholder: a readme, and nothing to run yet.

Generalist vision-language-action models have enabled diverse open-world manipulation, but real-world robotic assistance asks for two capabilities they currently lack. The first is personalization. Carrying out an instruction like make me a sandwich means grounding the breakdown into subtasks in the user's persistent preferences, rather than relying on one generic solution. The second is reliable execution. Deployment demands explainable safety assurance that proactively prevents hazards, and that recovers gracefully from the occasional physical failure. A policy that manipulates well in general still knows nothing about this particular user, and still has no account of what to do when a grasp goes wrong.

CarryOn is a dual-loop harness designed to augment a frozen VLA policy. The policy is left as it is. Two loops around it supply what it lacks: a planning loop for personalization, and an execution loop for reliability. One decides what should be done for this person in this scene. The other watches how it is being done, and steps in when it goes badly.

The planning loop maintains persistent context at two levels, one for the user and one for the scene. That context is continually grounded against new observations and revised through interaction, which enables personalized task selection and execution grounding across conversations and changing scenes. The execution loop actively supervises policy rollouts, shielding actions from predicted hazardous contact. When execution fails anyway, the system invokes reusable, control-based recovery skills that restore task progress. Those skills are refined through reflection on earlier failures, so a recovery that has been needed before improves the next time.

CarryOn was evaluated on a real robot and in simulation. It dispatches preference-consistent subtasks across users and scenes. It reduces the violation rate of a VLA that had already been safety-finetuned from fifty-two point six percent to six point one percent. And when refined skills are reused, recovery success rises from thirteen percent to eighty percent.

The design keeps the policy frozen and places memory and supervision outside it. Preferences live in the planning loop, hazards and recoveries in the execution loop. The steepest drop in violations comes on a model that had already been tuned for safety, which locates the gain in the harness rather than in the policy underneath.