CarryOn
An enhancement framework for robotic VLA models that superimposes two coordinated loops, a planning loop and an execution loop, on top of a frozen underlying VLA policy. These loops are responsible for maintaining persistent personalized context and for safety protection and failure recovery, respectively, improving the personalization and reliable execution of robotic assistants without retraining.
Timeline (last 90 days)
The CarryOn framework adopts a dual-loop architecture, layering a planning loop and an execution loop, two loops that work in coordination, on top of a frozen VLA policy.
CarryOn's planning loop maintains two layers of persistent context, at the user level and the scene level, recording across conversations the individual's personal preferences and the current physical environment state, respectively.
CarryOn's execution loop actively supervises the VLA policy's action outputs, masking actions predicted to involve dangerous contact, achieving prediction-based proactive protection.
In experiments, CarryOn reduced the safety violation rate from 52.6% to 6.1%.
When CarryOn reuses refined recovery skills, the task recovery success rate rises from 13% to 80%.
CarryOn's execution loop assesses potential risks before actions are executed, intercepting operations that could lead to accidents in advance.
CarryOn's recovery skills are based on control methods, are reusable and highly parameterized, and are refined through reflection on past failure cases.
CarryOn's planning loop continuously updates the user-level and scene-level context with new observations and interaction feedback, avoiding stale memory.
The boundary of the CarryOn framework lies in that the context maintenance of the planning loop depends on high-quality interaction history, and the execution loop's recovery skill library needs to be accumulated in advance.
All Facts (9)
The CarryOn framework adopts a dual-loop architecture, layering a planning loop and an execution loop, two loops that work in coordination, on top of a frozen VLA policy.
50%UnverifiedCarryOn's planning loop maintains two layers of persistent context, at the user level and the scene level, recording across conversations the individual's personal preferences and the current physical environment state, respectively.
50%UnverifiedCarryOn's execution loop actively supervises the VLA policy's action outputs, masking actions predicted to involve dangerous contact, achieving prediction-based proactive protection.
50%UnverifiedIn experiments, CarryOn reduced the safety violation rate from 52.6% to 6.1%.
50%UnverifiedWhen CarryOn reuses refined recovery skills, the task recovery success rate rises from 13% to 80%.
50%UnverifiedCarryOn's execution loop assesses potential risks before actions are executed, intercepting operations that could lead to accidents in advance.
50%UnverifiedCarryOn's recovery skills are based on control methods, are reusable and highly parameterized, and are refined through reflection on past failure cases.
50%UnverifiedCarryOn's planning loop continuously updates the user-level and scene-level context with new observations and interaction feedback, avoiding stale memory.
50%UnverifiedThe boundary of the CarryOn framework lies in that the context maintenance of the planning loop depends on high-quality interaction history, and the execution loop's recovery skill library needs to be accumulated in advance.
50%