CoRL 2026

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

Connect perception, future scene motion, and robot control in one metric space. 3D-Grounded. No robot-action pretraining required.

Ziyin Xiong ¹

Nikolaos Gkanatsios ¹

Moritz Reuss ²

Katerina Fragkiadaki ¹

¹Carnegie Mellon University, ²NVIDIA

Paper Code arXiv Slides Thread

Problem

Vanilla VLAs map 2D features directly to 3D actions, while motion-guided policies predict future in spaces disconnected from control. Both suffer from a geometric mismatch with robot actions.

Method

GeomVLA lifts vision-language features into 3D, predicts task-conditioned 3D scene trajectories as latent motion reasoning, and uses them to guide action generation in the same metric space.

Result

Without robot-action pretraining, GeomVLA reaches 4.624 on CALVIN, 98.4% on LIBERO, 84.8% on representative RoboTwin2.0 tasks, and 62.5% across eight real-world tasks.

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.

Architecture

Future Scene Motion

Given RGB-D observations and a language instruction, the 3D Scene Trajectory Denoiser predicts how scene points should move over time. Rather than serving as an explicit open-loop plan, these trajectories provide a latent geometric reasoning signal for closed-loop action generation.

Simulation Rollouts

Real-world Deployment

We deploy a multi-task policy on the right ViperX300S arm of an ALOHA2 platform. Demonstrations are collected with a WidowX250S leader arm, while observations are captured by two elevated Azure Kinect DK cameras and a wrist Intel RealSense D405.

Insert the marker into the cup.

Put the green bottle into the green cup, and put the blue bottle into the blue cup.

Rank the three bottles in order of pink, green and blue.

Stack two blocks.

Pick up the marker from one cup and place into another cup.

Place the three bottles in a triangle.

Uncap the marker.

Stand bottle upright.

Insert the marker into the cup.

Put the green bottle into the green cup, and put the blue bottle into the blue cup.

Rank the three bottles in order of pink, green and blue.

Stack two blocks.

Pick up the marker from one cup and place into another cup.

Place the three bottles in a triangle.

Uncap the marker.

Stand bottle upright.

Quantitative Results

3D VLAs Motion-guided Other baselines GeomVLA and variants

CALVIN ABC_D

5.0 4.5 4.0 3.5

UniVLA

π₀.₅

OpenVLA-OFT

X-VLA

FLOWER

2D-JA

GeomVLA w/o motion

GeomVLA-JA

GeomVLA

Average sequence length, reported out of 5.

RoboTwin2.0 5-task

90 75 60 45 30

PoseVLA

OpenVLA-OFT

X-VLA

SimpleVLA-RL

π₀.₅

2D-JA

GeomVLA w/o motion

GeomVLA-JA

GeomVLA

Average success rate over five representative RoboTwin2.0 tasks.

LIBERO

100 90 80 70 65

SpatialVLA

4D-VLA

PoseVLA

TraceVLA

CoT-VLA

WorldVLA

RoboFlow4D

FlowVLA

mimic-video

UniVLA

LaMP

π₀.₅

FLOWER

OpenVLA-OFT

X-VLA

SimpleVLA-RL

GeomVLA w/o motion

GeomVLA

LIBERO success rates. Select a benchmark split to update the bars.

RoboTwin2.0 50-task Results

Task	π₀.₅		GeomVLA		Task	π₀.₅		GeomVLA
Task	Easy	Hard	Easy	Hard	Task	Easy	Hard	Easy	Hard
Adjust Bottle	86	92	100	100	Place Can Basket	70	70	84	76
Beat Block Hammer	84	80	88	90	Place Cans Plasticbox	74	86	94	92
Move Playingcard Away	74	76	98	96	Place Dual Shoes	62	64	84	86
Move Stapler Pad	56	64	72	76	Place Mouse Pad	44	36	88	90
Place A2B Right	82	72	96	90	Place Object Scale	80	70	96	98
...
Place Bread Basket	74	68	86	86	Place Phone Stand	64	66	76	84
Place Bread Skillet	72	66	92	80	Place Shoe	88	84	96	100
Place Burger Fries	78	82	96	94	Stamp Seal	88	74	96	96
Move Pillbottle Pad	80	88	96	94	Turn Switch	56	58	72	72
Move Can Pot	92	88	100	98	Place Empty Cup	90	90	100	96

Average over all 50 RoboTwin2.0 tasks: π₀.₅ Easy 75.9 / Hard 75.7, GeomVLA Easy 78.6 / Hard 76.2.

Real-world ALOHA2

π₀.₅ Joint-angle variants GeomVLA and ablation

70 50 30 10 0

π₀.₅

2D-JA

GeomVLA-JA

GeomVLA w/o motion

GeomVLA

Average success rate over eight real-world tabletop manipulation tasks.

What Makes GeomVLA Work?

Ablations isolate the contributions of future-motion reasoning, explicit 3D representation, motion-latent design, and geometric consistency across the perception-to-action pipeline.

Hover a point to expand notes and highlight related ablations.

4.04.24.44.6

GeomVLA w/o motion

4.508

2DGeomVLA w/o motion

4.411

2DGeomVLA

4.437

2DGeomVLA LaMP-style

4.462

GeomVLA w/ 2D motion

4.085

GeomVLA w/ denoised trace

4.047

GeomVLA w/ denoised latent

4.334

GeomVLA w/ partial denoising

4.484

GeomVLA encoder-only

4.443

GeomVLA w/ arbitrary frame

4.596

GeomVLA

4.624

BibTeX

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

@inproceedings{xiong2026geomvla,
  title     = {{GeomVLA}: Unifying Scene, Motion, and Action in 3D},
  author    = {Xiong, Ziyin and Gkanatsios, Nikos and Reuss, Moritz and Fragkiadaki, Katerina},
  booktitle = {Proceedings of the Conference on Robot Learning (CoRL)},
  year      = {2026}
}