Skip to content

MoManipVLA: Transferring Vision-Language-Action Models for General Mobile Manipulation

Conference: CVPR 2025
arXiv: 2503.13446
Code: Yes (Project Page)
Area: Robotics / Mobile Manipulation
Keywords: Vision-Language-Action Models, Mobile Manipulation, Bi-level Trajectory Optimization, VLA Transfer, Reachability

TL;DR

Proposes MoManipVLA to transfer pre-trained fixed-base VLA models to mobile manipulation scenarios. By jointly planning base movement and manipulator trajectories using a bi-level trajectory optimization (optimizing reachability, smoothness, and collision avoidance), it achieves a 66.1% success rate (+4.2%) on the OVMM benchmark and can be deployed in the real world with only 50 demonstrations.

Background & Motivation

Background

Background: VLA models (such as RT-2, Octo) perform excellently in fixed-base manipulation, but mobile manipulation requires coordinated planning of both the base and the manipulator—the base must move to an appropriate position before the arm can reach the target.

Limitations of Prior Work: Directly applying end-effector trajectories from fixed-base VLAs to mobile robots fails because the target position may exceed the reach of the arm length from the current base position. Additional base motion planning is required.

Key Challenge: VLAs output end-effector waypoints without including base motion information. Training a mobile manipulation VLA from scratch is extremely data-costly.

Key Insight: Keep the VLA unchanged and add a bi-level optimizer on top of its output waypoints: the inner loop solves for joint angles using IK, while the outer loop optimizes the base position to maximize reachability, maximize smoothness, and minimize collisions.

Core Idea: VLA-generated waypoints + bi-level trajectory optimization (reachability/smoothness/collision) = low-cost mobile manipulation transfer.

Method

Key Designs

  1. Bi-level Trajectory Optimization:

    • Function: Jointly plans base movement and manipulator trajectories.
    • Mechanism: The outer loop optimizes the base position sequence \(\mathbf{x}_b\) and arm joint angles \(\boldsymbol{\theta}\), with the objective function \(\mathcal{O} = 10\mathcal{F}_r + 1\mathcal{F}_s + 0.6\mathcal{F}_c\), where \(\mathcal{F}_r\) is the IK reachability (imposes a large constant penalty if unreachable), \(\mathcal{F}_s\) is smoothness (first-order difference of base and joint angles), and \(\mathcal{F}_c\) is collision (signed distance function to obstacles).
    • Design Motivation: Reachability has the highest weight (10) because the primary challenge of mobile manipulation is "whether it can be reached".

Loss & Training

The VLA is fine-tuned using a small amount of embodiment-specific data. Base optimization requires no learning—it is solved directly during inference. Only 50 expert demonstrations are required for real-world deployment.

Key Experimental Results

Main Results

Method OVMM Overall Success Rate Pick Success Rate
SOTA Baseline 61.9% 50.2%
MoManipVLA 66.1% 62.6%

Key Findings

  • Reachability is the most critical constraint: removing it drops the success rate from 66.1% to 48.2%.
  • Only 50 demonstrations are needed to achieve a 40% success rate in the real world.
  • Inference latency is 693ms (vs. 742ms for direct optimization).
  • Smoothness constraint contributes significantly: removing it increases trajectory jitter and reduces the success rate to 56.3%.
  • Collision avoidance constraint plays an obvious role in cluttered scenarios; without this constraint, the success rate drops to 60.1%.

Highlights & Insights

  • Zero-data base planning—no mobile manipulation data is required, as the base trajectory is solved by the optimizer during inference.
  • VLA reuse strategy—the VLA is not retrained, and only a planning optimization layer is added.
  • The decoupled design fully reuses the generalization capability of the VLA, avoiding the expensive data collection costs of mobile manipulation.
  • The bi-level optimization framework is elegantly designed, with the outer loop optimizing the base position and the inner loop solving joint angles, offering clear logic and scalability.

Limitations & Future Work

  • Relies on the quality of visual segmentation masks (success rate drops to 23.7% without masks).
  • Trajectory length is limited to < 150 steps.
  • Requires embodiment-specific fine-tuning.
  • Computational overhead of the optimizer solving base trajectories during inference may become a bottleneck in complex environments.
  • Only validated in tabletop manipulation scenarios; larger-scale mobile manipulation (such as household navigation + retrieval) is not covered.
  • The perception range of the VLA model is limited by a single camera view; multi-camera fusion might further improve performance.
  • For tasks requiring fine manipulation (e.g., insertion, tightening), waypoint-level planning may lack sufficient precision.
  • Currently, the objective function weights (10:1:0.6) of the optimizer need to be adjusted for different robot platforms.
  • Multi-arm or bimanual mobile manipulation scenarios have not been explored.

Rating

  • Novelty: ⭐⭐⭐⭐ Practical decoupled design of VLA + trajectory optimization.
  • Experimental Thoroughness: ⭐⭐⭐⭐ Simulation + Real-world.
  • Writing Quality: ⭐⭐⭐⭐ Clear.
  • Value: ⭐⭐⭐⭐ Low-cost transfer of VLA to mobile manipulation.