Research map / Vision-language and multimodal / Robot action and navigation
MoManipVLA: Transferring Vision-language-action Models for General Mobile Manipulation
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Transfers a fixed-base VLA (OpenVLA-7B fine-tuned with LoRA) to mobile manipulation by using its end-effector waypoints and a bi-level trajectory optimization over base and arm with reachability, smoothness, and collision costs.
Problem
Fixed-base VLA models generalize well across manipulation tasks but cannot produce coordinated base and arm motions needed for mobile manipulation, while conventional mobile manipulation methods generalize poorly due to limited demonstration data.
Approach
Use pre-trained VLA models to predict end-effector waypoints, then generate whole-body base and arm trajectories by minimizing physical-feasibility costs (IK reachability, smoothness, ESDF-based collision) with a bi-level optimization: the upper level optimizes base poses to enlarge the arm's policy space and the lower level selects the end-effector trajectory.
Key results
- On OVMM, MoManipVLA reaches 49.4% overall success rate, versus 38.2% for the best prior baseline (KUZHUM), a 4.2-point improvement from the abstract's comparison.
- Real-world tests with 50 fine-tuning samples achieve 40% success on Put in Bowl, 30% on Stack Block, and 10% on Open Drawer (10 trials each).
- Ablation shows removing the GT masks drops overall success to 11.3%, and removing the reachability cost is the largest drop among cost terms (48.2% vs 49.4%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| transfer OpenVLA-7B waypoints to mobile manipulation using MoManipVLA | MoManipVLA · OVMM | Overall SR: 38.2% -> 49.4% | Helped |
| use bi-level search instead of direct search (ablation) | MoManipVLA · OVMM | Overall SR: 48.5% -> 49.4% | Helped |
| use bi-level search instead of direct search (ablation) | MoManipVLA · OVMM | Latency: 742.9 -> 693.1 | Helped |
| use ground-truth masks instead of instance segmentation (ablation) | MoManipVLA · OVMM | Overall SR: 11.3% -> 49.4% | Helped |
| add the reachability cost (ablation) | MoManipVLA · OVMM | Overall SR: 48.2% -> 49.4% | Helped |
| add the smoothness cost (ablation) | MoManipVLA · OVMM | Overall SR: 48.7% -> 49.4% | Helped |
| add the collision cost (ablation) | MoManipVLA · OVMM | Overall SR: 48.9% -> 49.4% | Helped |
| remove the reachability cost (ablation) | MoManipVLA · OVMM | Latency: 693.1 -> 682.7 | Helped |
| remove the smoothness cost (ablation) | MoManipVLA · OVMM | Latency: 693.1 -> 688.2 | Helped |
| remove the collision cost (ablation) | MoManipVLA · OVMM | Latency: 693.1 -> 692.1 | Helped |
Surprising findings
- Using vision foundation model masks (Detic) instead of ground-truth masks sharply degrades overall success to 11.3%, showing perception quality is the main bottleneck.
Limitations
- Performance is bounded by the underlying VLA models and the method works only in highly constrained movement spaces.
- Cannot generate effective motions for long-horizon tasks without a task planning module.
- Other moving platforms require additional hand-designed motion planning constraints.
Experimental setup
Datasets: OVMM expert trajectories (collected with heuristic baseline)
Benchmarks: OVMM
Metrics: overall success rate partial success rate steps latency
Baselines: OVMM (RL) OVMM (Heuristic) RoboAI KUZHUM UniTeam
Builds on: OpenVLA OVMM ReKep nvblox
Closest papers by meaning
- Exploring the Limits of Vision-Language-Action Manipulation in Cross-task Generalization (NeurIPS 2025 · similarity 0.89)
- VideoVLA: Video Generators Can Be Generalizable Robot Manipulators (NeurIPS 2025 · similarity 0.89)
- ForceVLA: Enhancing VLA Models with a Force-aware MoE for Contact-rich Manipulation (NeurIPS 2025 · similarity 0.89)
- HAMSTER: Hierarchical Action Models for Open-World Robot Manipulation (ICLR 2025 · similarity 0.89)
- AC-DiT: Adaptive Coordination Diffusion Transformer for Mobile Manipulation (NeurIPS 2025 · similarity 0.89)
- ActionUNet: Improving Robustness of VLA Models with Efficient Multi-scale Fine-tuning (NeurIPS 2026 · similarity 0.88)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.