MOMASPLAT augments mobile manipulation demonstrations
Original title: MOMASPLAT: Dynamic Novel-View Demonstration Augmentation for Mobile Manipulation
MOMASPLAT more than doubles success rate on tasks with unseen base poses.
Mobile manipulation policies fail on viewpoint changes from navigation drift. MOMASPLAT synthesizes full trajectories including dynamic interactions using 3D Gaussian Splatting. It combines multiview observations, foundation models, and action retargeting to improve robustness across five real-world tasks.
- More than doubles average success rate
- Maintains non-trivial success at large displacements
- Reduces need for large-scale real-world data
- Evaluated on five tasks with rigid and articulated objects
From the original article
1 min read at arxiv.orgVisuomotor policies trained through imitation learning are highly sensitive to changes in camera viewpoint. This is particularly problematic for mobile manipulation (MoMa), where robots rarely execute manipulation from the exact same base pose used during demonstration collection due to navigation uncertainty, localization drift, and environmental constraints.
By Sateesh Kumar · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.