Event
RoboDuo: A Closed-Loop Framework Connecting Dual VLAs via Code
1 reports · 1 sources
Jie Wang, Bo Zhang, Renzhi Lu, and Jianyu Chen propose RoboDuo, a closed-loop dual-system framework that uses code as the interface between a coding agent and two VLAs: the coding agent writes programs that let the full-task VLA attempt the entire task and check whether it is completed; when it fails, the atomic VLA completes the remaining steps. The authors state that the framework wraps an existing full-task VLA without updating its weights. The authors state that RoboDuo stably improves X-VLA and π0.5 on LIBERO and LIBERO-PRO, with the largest gains on long-horizon tasks, and that it recovered failed executions of a frozen full-task VLA on a real robot. The team also built the LIBERO-Atomic, LIBERO-Code, and LIBERO-Judge datasets from the same set of LIBERO demonstrations.
Report timeline
- RoboDuo: Using Code to Enable Attempt, Check, and Atomic Recovery for Vision-Language-Action Models
RoboDuo proposes a closed-loop dual-system framework that uses code as the interface between a coding agent and two VLAs: the coding agent writes programs that let a full-task VLA attempt the entire task and check whether it is complete; on failure, an atomic VLA completes the remaining steps.
arXiv cs.RO ·