GeoBridge-VLA: Geometry-Aware Residual Adaptation for Vision-Language-Action Models

Hyun Song; Kangmin Kim; Loren Jinsoo Um; Minhui Han; Jaehyeok Park; Taewan Cho; Andrew Jaeyong Choi

cs.CVOct 4, 2026

GeoBridge-VLA: Geometry-Aware Residual Adaptation for Vision-Language-Action Models

Authors: Hyun Song, Kangmin Kim, Loren Jinsoo Um, Minhui Han, Jaehyeok Park, Taewan Cho, Andrew Jaeyong Choi

Organizations: School of Computing, iRASC Lab., Gachon University

Abstract

Vision-language-action (VLA) models encode semantic information from vision-language pretraining, but manipulation also requires precise spatial reasoning. We present GeoBridge-VLA, a two-stage method for learning geometric features from a pretrained VLA's frozen visual encoder and using them for action prediction. Stage I trains a feature bridge and geometry decoder with depth supervision. Stage II freezes these modules and trains a gated residual interface together with the action-side projections and action expert. The residual augments the existing visual tokens without adding a second image encoder or increasing the token count. Deployment requires RGB, robot state, and language, but no depth observations. Under matched evaluation conditions, GeoBridge-VLA achieves 70.9% success on LIBERO, compared with 60.0% for SmolVLA. Disabling the residual in the same trained checkpoint reduces success from 70.90% to 69.85%, with mixed effects across suites. On a physical ROBOTIS OMY robot, GeoBridge-VLA succeeds in 148 of 200 trials (74.0%) across four tasks, compared with 108 of 200 (54.0%) for SmolVLA.

Figures & tables

Fig. 2: Geometry integration in VLA policies. (a) A native RGB-only VLA uses its visual tokens for action conditioning. (b) A separate-depth design adds a second RGB encoder and a fusion module. (c) GeoBridge-VLA shares the native visual encoder, decodes geometry through a feature bridge and K4-based geometry decoder, and adds a learned residual to the existing visual tokens. Language and robot-state inputs are omitted here for clarity.

Fig. 3: The four LIBERO evaluation suites [ 31 ] . Representative RGB observations and language instructions illustrate spatial relations (Spatial), object variation (Object), goal variation (Goal), and multi-step manipulation (Long). These panels introduce the task settings used in Table II .

Fig. 4: ROBOTIS OMY task settings. Each task panel pairs the wrist camera (left) with the external camera (right). Top row: T1, red-cube pick-and-place; T2, drawer closing. Bottom row: T3, bowl retrieval and transfer; T4, size-ordered block stacking. The photos document camera views and manipulation settings. Table V reports task success rates for SmolVLA and GeoBridge-VLA.

Task	Language instruction
T1: Cube transfer	Pick up the red cube on the right side and place it on the blue plate on the left side.
T2: Drawer closing	Close the drawer.
T3: Bowl transfer	Pick up the red bowl in the drawer and place it on the blue plate on the right side.
T4: Block stacking	Stack the three red cubes on the blue plate, from largest at the bottom to smallest at the top.

TABLE I: OMY tasks and language instructions.

LIBERO Suite	SmolVLA (SR)	GeoBridge-VLA (SR)	Change ( $\Delta$ , pp)
Spatial	68.8%	73.4%	+4.6
Object	65.6%	86.4%	+20.8
Goal	72.8%	76.6%	+3.8
Long	32.8%	47.2%	+14.4
Overall	60.0%	70.9%	+10.9

TABLE II: LIBERO success rates (%). Both methods use the same evaluation conditions. Overall is the suite average;

\Delta

is the change in percentage points.

Fig. 5: LIBERO-Long task sequence. Frames 0, 125, and 251 illustrate progress through a multi-step mug-and-pudding manipulation task.

Model	AbsRel $\downarrow$	RMSE (m) $\downarrow$	$\delta_{1}$ $\uparrow$
SmolVLA shared encoder	0.030	0.1316	0.975
DA3-Mono-Large [ 13 ]	0.036	0.1549	0.988

TABLE III: Shared-encoder depth prediction. Results on the LIBERO front-view validation set.

Fig. 6: Shared-encoder depth predictions. Qualitatively selected LIBERO front views show RGB (left), reference depth (middle), and predictions (right), illustrating robot, object, and tabletop geometry.

Fig. 7: OMY shared-encoder depth predictions. RGB observations (left) and predicted relative depth (right) for (a) the front view and (b) the wrist view. Examples are qualitatively selected from the training set.

LIBERO Suite	Normal	Zero	$\Delta$ (pp)
Spatial	73.4%	70.2%	+3.20
Object	86.4%	87.6%	-1.20
Goal	76.6%	77.2%	-0.60
Long	47.2%	44.4%	+2.80
Overall	70.9%	69.85%	+1.05

TABLE IV: Same-checkpoint residual intervention (%). Normal and zero use the same trained checkpoint without retraining. Zero sets

\Delta z=0

at inference;

\Delta

is normal minus zero in percentage points; positive values indicate higher success with the residual enabled.

Task	SmolVLA	GeoBridge-VLA	$\Delta$ (pp)
T1: cube transfer	60.0%	80.0%	+20.0
T2: drawer closing	70.0%	86.0%	+16.0
T3: bowl transfer	50.0%	70.0%	+20.0
T4: block stacking	36.0%	60.0%	+24.0
Overall	54.0%	74.0%	+20.0

TABLE V: OMY success rates (%). Each method is evaluated over 50 trials per task. Overall is the unweighted mean across the four tasks;

\Delta

denotes GeoBridge-VLA minus SmolVLA in percentage points.

Decoder	AbsRel $\downarrow$	RMSE (m) $\downarrow$	$\delta_{1}$ $\uparrow$
SPACE-CLIP V1	0.05239	0.06134	0.95497
SPACE-CLIPv2	0.04321	0.04993	0.96678

TABLE VI: Standalone depth comparison. Matched CLIP backbone and 9,984-image LIBERO test split.

Explore similar work

HABILIS Brain 0: Geometry-Change Supervision for Vision-Language-Action and Residual Flow Recovery
Sep 22, 2026Jinu Pahk, Jesoon Kang, Taegeon Park +4Photometric Supervision Egocentric Video
Tommoro Robotics
G $^3$ VLA: Geometric inductive bias for Vision-Language-Action Models
Jun 23, 2026Yue Peng, Yongzhe Zhao, Artur Habuda +5Diffusion-Based Vision-Language-Actions Visual Geometry Grounded Transformer
New York University Shanghai · Technical University of Denmark · MBZUAI - Mohamed bin Zayed University of Artificial Intelligence
Latent Bridge: Feature Delta Prediction for Efficient Dual-System Vision-Language-Action Model Inference
May 4, 2026Yudong Liu, Yuan Li, Zijia Tang +12Robotic Manipulation Latent Variable
1Duke University · 2Qualcomm AI Research · University of Florida

Sep 22, 2026, cs.ROJ/K move · Enter open · S save

HABILIS Brain 0: Geometry-Change Supervision for Vision-Language-Action and Residual Flow Recovery

Jinu Pahk, Jesoon Kang, Taegeon Park, Jisu An+3

Tommoro Robotics

Photometric Supervision Egocentric Video

Open paper AskPDF arXiv

Vision-language-action policies benefit from geometric supervision, but current-frame geometry alone does not explicitly describe the changes associated with manipulation. This design is motivated by the goal of learning an embodiment-agnostic visual interface that can be pretrained across robot and egocentric video before robot-specific action alignment. We introduce Geometry-Change VLA (GC-VLA), which learns to predict multiview future-current geometry-change tokens from current observations. Offline frame pairs define a nominal 0.5-second prediction horizon; future observations are used only to construct training targets. Stage 1 trains a geometry-change vision-language model (GC-VLM). Stage 2 introduces a continuous ActionExpert and aligns it with robot actions while stopping action-flow gradients at the VLM interface. Stage 3 enables these gradients to update the trainable VLM components jointly with the ActionExpert. Stage 4 freezes GC-VLA and applies Geometry-Conditioned Residual Flow (GCRF), using a binary intervention router and a single bounded residual velocity policy learned from closed-loop feedback. GC-VLA achieves 95.20% success on LIBERO, and GC-VLA with GCRF achieves 99.55%. Inference uses current observations and the learned GC representation without executing the offline target encoders.

Abstract

Explore similar work

HABILIS Brain 0: Geometry-Change Supervision for Vision-Language-Action and Residual Flow Recovery

G3^33VLA: Geometric inductive bias for Vision-Language-Action Models

Latent Bridge: Feature Delta Prediction for Efficient Dual-System Vision-Language-Action Model Inference

G $^3$ VLA: Geometric inductive bias for Vision-Language-Action Models