GeomVLA: Unifying Scene, Motion, and Action in 3D

Ziyin Xiong 1
Nikos Gkanatsios 2
Moritz Reuss 3
Katerina Fragkiadaki 1
CoRL 2026
1Carnegie Mellon University, 2Wayve, 3NVIDIA
Paper Code arXiv Slides
GeomVLA overview animation

Abstract

We present GeomVLA, a Vision-Language-Action model that unifies perception, latent scene motion prediction, and action generation within a shared 3D coordinate frame. Our approach lifts VLM features into spatially grounded 3D scene tokens by adjusting their positional encodings while preserving pretrained visual-language priors. We further introduce a 3D Scene Trajectory Denoiser, a task-conditioned latent motion representation that captures how scene points are expected to move in 3D space. The predicted scene trajectory acts as a latent geometric reasoning process that conditions a 3D flow-based action denoiser through 3D-aware attention.

GeomVLA achieves state-of-the-art performance on CALVIN, strong performance on LIBERO and RoboTwin2.0, and outperforms strong baselines in real-world manipulation settings without large-scale robot-action pretraining. Extensive ablations demonstrate that future-motion reasoning alone is insufficient: the primary gains arise from maintaining geometric consistency between scene representation, motion prediction, and robot actions throughout the entire perception-to-action pipeline.

Method

GeomVLA model architecture
Multi-view RGB observations and language are lifted into 3D scene tokens, latent scene motion is predicted with a 3D Scene Trajectory Denoiser, and actions are generated by a 3D-aware denoising policy, with an optional joint-angle decoder for joint-space control.

Scene Trajectory Prediction

Simulation Rollouts

Real-world Deployment

Quantitative Results

table/graph for sim and real-world

Ablations

table/graph for key ablations

BibTeX

If you find this work useful, please consider citing our paper:

@misc{xiong2026geomvla,
  author = "{Ziyin Xiong, TODO}",
  title = "GeomVLA: Unifying Scene, Motion, and Action in 3D",
  year = "2026",
  howpublished = "\url{TODO}",
}