MoManipVLA: Transferring Vision-language-action Models for General Mobile Manipulation

TL;DR

MoManipVLA addresses the generalization gap in mobile manipulation by transferring pre-trained fixed-base Vision-Language-Action (VLA) policies to mobile platforms. It uses VLA-predicted end-effector waypoints to drive a bi-level trajectory optimization that jointly plans the mobile base and robot arm under reachability, smoothness, and collision constraints. The approach leverages upper-level base waypoint optimization to widen the manipulator policy space and a lower-level end-effector optimization guided by VLA, solved with Dual Annealing and SLSQP refinement. Experimental results on OVMM and real-world tasks show improved success rates (≈4.2% SR gain) and data efficiency (≈50 expert episodes for deployment), demonstrating strong generalization from fixed-base to mobile manipulation.

Abstract

Mobile manipulation is the fundamental challenge for robotics to assist humans with diverse tasks and environments in everyday life. However, conventional mobile manipulation approaches often struggle to generalize across different tasks and environments because of the lack of large-scale training. In contrast, recent advances in vision-language-action (VLA) models have shown impressive generalization capabilities, but these foundation models are developed for fixed-base manipulation tasks. Therefore, we propose an efficient policy adaptation framework named MoManipVLA to transfer pre-trained VLA models of fix-base manipulation to mobile manipulation, so that high generalization ability across tasks and environments can be achieved in mobile manipulation policy. Specifically, we utilize pre-trained VLA models to generate waypoints of the end-effector with high generalization ability. We design motion planning objectives for the mobile base and the robot arm, which aim at maximizing the physical feasibility of the trajectory. Finally, we present an efficient bi-level objective optimization framework for trajectory generation, where the upper-level optimization predicts waypoints for base movement to enhance the manipulator policy space, and the lower-level optimization selects the optimal end-effector trajectory to complete the manipulation task. In this way, MoManipVLA can adjust the position of the robot base in a zero-shot manner, thus making the waypoints predicted from the fixed-base VLA models feasible. Extensive experimental results on OVMM and the real world demonstrate that MoManipVLA achieves a 4.2% higher success rate than the state-of-the-art mobile manipulation, and only requires 50 training cost for real world deployment due to the strong generalization ability in the pre-trained VLA models.

Figures (3)

• Figure 1: Transferring pre-trained VLA models to mobile manipulation significantly enhances the generalization ability of the policy across tasks and environments. Our MoManipVLA can complete diverse household tasks such as object picking, object delivery and drawer opening in large working area. Guided by the waypoints from VLA models, the motions of the mobile base and the robot arm are jointly generated with physical feasibility constraints.
• Figure 2: The pipeline of MoManipVLA. The pre-trained VLA models predict highly generalized end-effector waypoints to guide the mobile manipulation task, through which the trajectory of the mobile base and the robot arm can be generated with objectives of physical feasibility. The objectives consider the reachability, smoothness and collision, and the trajectory is acquired via bi-level optimization.
• Figure 3: Real-world mobile manipulation visualization results. We demonstrate a real-world mobile robotic Pick-and-Place action sequence, which generates mobile manipulation trajectories through iterative bi-layer search of the base and arm.