GaussVLA: Geometry-Aware Spatial Reasoning for Vision-Language-Action Model

Sarowar, Md Selim; Islam, Md Tanvir; Kim, Sungho; Ahn, Sangtae

GaussVLA: Geometry-Aware Spatial Reasoning for Vision-Language-Action Model

Md Selim Sarowar¹, Md Tanvir Islam², Sungho Kim^1,†, Sangtae Ahn²

¹Yeungnam University ²Kyungpook National University
^†Corresponding author

BMVC 2026 · United Kingdom

arXiv Paper OpenReview Cite

The arXiv release includes the supplementary material.

GaussVLA executing long-horizon LIBERO tasks end to end.

0%

LIBERO-Spatial success

0%

LIBERO average success

0M

Trainable parameters

0 ms

Inference latency

Abstract

Vision-Language-Action (VLA) models encode visual observations as flat 2D patch tokens that carry no intrinsic geometric structure, and augmenting them with dense monocular depth injects per-pixel scalar values that encode neither surface orientation nor geometric confidence. This leaves the policy with limited structured spatial reasoning for action prediction. We propose GaussVLA, a mamba-based VLA that incorporates two custom modules: Gaussian Spatial Tokenizer (GST) to lift frozen semantic and depth features into compact 3D Gaussian tokens, pools geometrically salient regions with learned queries, and Depth-Aware Chain-of-Thought (DA-CoT) that performs structured, non-autoregressive geometric reasoning under language and flow-time conditioning. Across both simulation and real-world evaluations, GaussVLA demonstrates strong spatial-manipulation performance while remaining parameter efficient. On LIBERO, it achieves 93.5% average success and 100.0% success on the Spatial suite with only 200M parameters, improving over SpatialVLA by 19.7% relative average success while remaining significantly more parameter-efficient. Our experimental results suggest that geometry-aware representations can provide a practical path toward more efficient and deployable robot manipulation.

Method

Pick a module to see what it contributes. Click any figure to enlarge it.

GaussVLA overview: flat 2D tokens lack geometry, while 3D Gaussian Spatial Tokens recover grounding, contact geometry, metric distance and coarse waypoints — **Overview of GaussVLA.** (a) Flat 2D tokens and weak depth cues leave the policy with ambiguous 3D understanding. (b) GST converts each 2D patch into a 3D Gaussian token and pools it with learned spatial-attention queries, yielding 3D grounding, contact geometry, metric distance and coarse waypoints. (c) Accumulated attention maps: without GST spatial reasoning attention is too sparse and the rollout fails; with GST it concentrates on the task-relevant region and succeeds.

What it does

Lifts frozen semantic (SigLIP) and dense depth (Depth Anything V2) features into compact 3D Gaussian tokens, then pools geometrically salient regions with learned queries instead of a flat patch grid.

What it buys

Adding GST alone lifts the LIBERO average from 78.1% to 90.5%, and LIBERO-PRO from 11.2% to 29.0%, for +11.2M parameters and +1.4 ms of latency.

Gaussian Spatial Tokenizer pipeline and attention maps — **GST pipeline**, from 2D patch to 3D Gaussian token to spatial attention pooling.

What it does

Performs structured, non-autoregressive geometric reasoning over the Gaussian tokens under language and flow-time conditioning, so reasoning cost does not grow with a token-by-token decode.

What it buys

DA-CoT is what carries the long-horizon suites: 83.0% vs 77.0% on LIBERO-Long and 92.3% vs 78.2% on LIBERO-90 at 20 trials.

Bar chart comparing success rate with and without DA-CoT on LIBERO-Long and LIBERO-90 at 20 and 40 trials — **DA-CoT ablation.** Success rate with and without DA-CoT on LIBERO-Long and LIBERO-90, at 20 and 40 evaluation trials. The gain is largest on LIBERO-90, where the reasoning horizon is longest.

LIBERO Benchmark

Click any column header to sort. Toggle the filter to compare only against similarly sized models.

Show only models under 1B parameters sortable

Method	Venue	Spatial	Object	Goal	Long	Average	Params
Diffusion Policy	RSS'23	78.3	92.5	68.3	50.5	72.4	80M
MaIL	CoRL'24	53.8	81.5	56.3	41.7	58.3	24M
QueST	NeurIPS'24	89.0	90.0	88.4	87.0	88.6	152M
OpenVLA	CoRL'24	84.7	88.4	79.2	53.7	76.5	7B
SpatialVLA	RSS'25	88.2	89.9	78.6	55.5	78.1	4B
ThinkAct	NeurIPS'25	88.3	91.4	87.1	70.9	84.4	7B
CoT-VLA	CVPR'25	81.5	91.6	87.6	69.0	82.43	7B
Mask2Act	BMVC'25	76.1	68.7	75.1	30.6	62.6	7B
TraceVLA	ICLR'25	84.6	85.2	75.1	54.1	74.8	4B
π₀	RSS'26	90.0	86.0	95.0	73.0	86.0	3.3B
SUREFlow	IROS'26	94.8	91.0	93.8	90.2	92.5	179.1M
GaussVLA Ours	BMVC'26	100.0	95.8	95.3	83.0	93.5	200M^‡

Success rate (%) on the four LIBERO suites. ^‡200M trainable parameters; the full system is roughly 1B including the frozen SigLIP and Depth Anything V2 backbones.

LIBERO rollouts from initial state through execution to end state for the Spatial, Goal, Object and Long suites — **Representative LIBERO rollouts.** GaussVLA executing one task from each suite, from initial state to completion.

Build the Model

Switch the two modules on and off to see the exact cost and benefit of each. Every number below is measured.

GST Gaussian Spatial Tokenizer DA-CoT Depth-Aware Chain-of-Thought

Vanilla GaussVLA

Visual input: flat 2D tokens

LIBERO78.1%

LIBERO-PRO11.2%

Total params179M

Trainable158M

GFLOPs3.50

Latency10.85 ms

Turn on a module to see its effect.

Two hyperparameter sweep curves, both peaking at 93.5 percent LIBERO average — **Hyperparameter sweep.** LIBERO average success across two loss coefficients. Both peak at the reported 93.5%, at 0.05 (left) and 0.10 (right), and degrade sharply beyond 0.5.

Collage of failure case rollout frames on a bowl and plate manipulation task — **Failure cases.** Remaining errors concentrate on fine contact geometry, where the gripper reaches the right region but the grasp does not settle.

Real Robot (SO-101)

Hover a bar to read its exact success rate.

ACT SpatialVLA GaussVLA (Ours)

52.0

61.3

75.7

Pick-Place

34.0

39.7

47.3

Stacking

21.0

37.8

53.5

Sorting

35.66

46.27

58.83

Average

Multi-task success rate (%) on the SO-101 arm.

Single-task Pick-Place, in-distribution and out-of-distribution
Method	ID	OOD	Avg.
ACT	57.0	27.8	42.4
SpatialVLA	69.0	35.8	52.4
GaussVLA Ours	81.0	46.7	63.9

Real-world SO-101 evaluation comparing GaussVLA with ACT and SpatialVLA, plus rollout frames of a physical stacking task — **Real-world SO-101 evaluation.** (a) Multi-task settings, (b) single-task Pick-Place under ID and OOD conditions, and (c) rollout frames for `Pick_the_red_cube_and_stack_on_the_small_wheel`.

Single-Episode Rollout

One LIBERO episode at full length, uncut.

Video Presentation

BibTeX

@inproceedings{Sarowar_2026_BMVC,
author    = {Md Selim Sarowar and Md Tanvir Islam and Sungho Kim and Sangtae Ahn},
title     = {GaussVLA: Geometry-Aware Spatial Reasoning for Vision-Language-Action Model},
booktitle = {37th British Machine Vision Conference 2026, {BMVC} 2026, Lancaster, UK, November 23-26, 2026},
publisher = {BMVA},
year      = {2026},
url       = {https://bmva-archive.org.uk/bmvc/2026/assets/papers/Paper_121/paper.pdf}
}