MOMASPLAT: Dynamic Novel-View Demonstration Augmentation for Mobile Manipulation

Sateesh Kumar, Arnav Balaji, Arpit Bahety, Skand Peri, Georgios Pavlakos, Roberto Martín-Martín

November 2026

Synthesized Pick-Place Bottle demonstrations rendered from many sampled base poses.

TL;DR MoMaSplat rebuilds the scene and the manipulated object as separate Gaussian splats from a mobile robot's own data, then re-renders whole demonstrations, including the object moving, from new base poses. On five real TIAGo tasks it succeeds on 37% of rollouts from unseen base poses, against 17% for the best baseline.

What they do

Visuomotor policies break when the camera moves, and a mobile manipulator rarely stops at exactly the base pose where it was taught. Earlier novel-view augmentation treats the scene as static, so it can re-render the approach but not the part where the object moves. MoMaSplat reconstructs the static scene and the manipulated object separately, re-renders full demonstrations from sampled base poses with retargeted actions, and trains a Diffusion Policy on real plus synthetic data.

How it works

Scene model. Navigation runs toward the demo base poses give multi-view coverage. COLMAP and WildGaussians build a 3DGS scene with robot and object masked out, scaled to metric with the onboard depth camera.
Object model. The object is segmented with Grounded-SAM and SAM 2, and FastVGGT estimates object-relative camera poses. Navigation frames, where the object is still, align these to the scene frame with a Sim(3) fit. The object splat (for articulated objects, only the moving link) is optimized jointly with its poses.
Synthesis. Base poses are sampled in a 1.2 x 1.2 m region and kept only if IK can follow the end-effector path relative to the object. Scene, object and an OmniGibson-rendered robot are depth-composited, and actions and proprioception are rotated into the new base frame.
Training. 10 synthetic demos per source demo, sampled 0.75:0.25 against the real ones.

Results

Five real tasks on a TIAGo, 12 unseen base poses x 3 rollouts each: 37% average success, vs 17% for One-Thousand Trajectories, 8% for VISTA and 6% for source-only and action augmentation. At the original poses all methods are comparable.
Where every baseline is at zero: 33% on Pick-Place Bottle at 30 cm, 50% on Close Laptop at 40 cm.
With 10, 50 or 90 source demos: 33%, 42%, 50% at 10-20 cm offsets; 90 demos without augmentation give 8%.
In simulation, object pose error is 1.46 cm and 4.01°; the scene renders at 26.40 dB PSNR, the object at 18.18 dB.

Limitations

Assumes mostly static scenes and rigid, visible objects. Surfaces hidden by the object at the start, such as under a lifted bottle or inside a drawer, are never reconstructed. Synthetic demos replay the source object motion without contact physics, depend on reliable segmentation, and are only as good as the navigation coverage.

Videos

Synthesized Open Drawer demonstrations from many sampled base poses.

Gaussian-splat model of the manipulated bottle, reconstructed separately from the scene.

Figures

From Figure 1: Starting from data collected at the original base pose, MoMaSplat synthesizes demonstrations from samp... — From Figure 1: Starting from data collected at the original base pose, MoMaSplat synthesizes demonstrations from sampled novel base poses, including the object as the robot moves it. Top right is the source demonstration, bottom right the synthesized one. Training on them lets the robot act from unseen deployment-time base poses.

From Figure 2: Overview. The static scene is reconstructed as a Gaussian splat with structure-from-motion on navigati... — From Figure 2: Overview. The static scene is reconstructed as a Gaussian splat with structure-from-motion on navigation data. Each manipulated object is segmented, its per-frame pose estimated by a feed-forward 3D foundation model, aligned to the scene with a Sim(3) transform and fit as its own splat. For a sampled base pose, scene, object and a simulator-rendered robot are composited and paired with transformed actions; the policy trains on source plus generated demonstrations.

From Figure 4: Success rate by distance from the closest original base pose, over 12 unseen base poses with 3 trials ... — From Figure 4: Success rate by distance from the closest original base pose, over 12 unseen base poses with 3 trials each per task. The gray region marks the original poses; MoMaSplat keeps higher success as the base moves away.

Abstract

Visuomotor 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. Recent approaches use novel-view synthesis (NVS) to augment demonstrations with unseen viewpoints, but they assume static scenes and therefore fail to re-render the manipulated object during the post-contact interaction phase, leaving the most critical parts of the demonstration unaugmented. We present MOMASPLAT, a novel-view demonstration augmentation framework for mobile manipulation that synthesizes full manipulation trajectories from unseen robot base poses, including dynamic object interactions. Our key insight is that standard MoMa data collection naturally provides the multiview observations needed to reconstruct both the static scene and the manipulated object in 3D. MOMASPLAT combines 3D Gaussian Splatting, feed-forward 3D foundation models, and kinematics-based action retargeting to separately reconstruct the scene and manipulated object, render them from novel viewpoints, and generate geometrically consistent synthetic demonstrations paired with transformed actions. We evaluate MOMASPLAT on five real-world mobile manipulation tasks involving both rigid and articulated objects across three scenes, and validate the fidelity of its synthesized demonstrations against ground truth in simulation. MOMASPLAT substantially improves robustness to unseen base poses, more than doubling the average success rate of the strongest baseline and maintaining non-trivial success at large base displacements where all baselines fail, while reducing the need for largescale real-world data collection.

Summaries are written automatically from each paper's text. Figures and videos come from the paper and its project page; all credit goes to the authors.

Matt Strong