FoveaVLA: Foveation Actions Replace the Wrist Camera in VLAs

Bingao Chen; Lihan Zha; C. Karen Liu; Guy Tevet

cs.ROOct 14, 2026

FoveaVLA: Foveation Actions Replace the Wrist Camera in VLAs

Authors: Bingao Chen, Lihan Zha, C. Karen Liu, Guy Tevet

Organizations: Stanford University · Princeton University

Abstract

We present FoveaVLA, a Vision-Language-Action (VLA) model that replaces the wrist camera with learned foveation actions. Wrist cameras provide close-up detail but restrict the view to the end-effector's surroundings. External cameras see the whole scene, but the policy receives them at low resolution, so the fine detail needed for manipulation is lost. FoveaVLA recovers this detail by predicting, alongside each action chunk, the interaction region to observe at the next policy query. The policy then receives crops of these regions, taken from the original-resolution images, together with the full views. We build FoveaVLA on $π_{0.5}$ in two configurations, Single and Dual, using one or two external stereo cameras with full and foveal RGB and depth inputs, and post-train both on DROID with RoboInter interaction annotations. Across four real-world task suites, FoveaVLA-Single achieves 70.8% success with one external RGB-D view, and FoveaVLA-Dual achieves 77.1%, versus 52.1% for the wrist-equipped $π_{0.5}$ . Both match the baseline on plain pick-and-place and improve success with distractors, cloth manipulation, and far-away objects.

Figures & tables

Fig. 2: FoveaVLA overview (Single shown). Full and foveal RGB images and aligned depth maps are encoded together with the task instruction and robot state. From this shared representation, the action expert generates an action chunk, while the language-model head autoregressively generates interaction boxes that select the foveal crops for the next observation. Dual adds the same four image inputs from a second external view and predicts one box per view.

Fig. 3: Interaction boxes and foveal crops. RoboInter annotations on DROID define interaction boxes (green) enclosing skill-relevant components; square foveal crops (purple) provide detailed visual inputs.

Fig. 4: Real-world task success of wrist-equipped

\pi_{0.5}

and the two FoveaVLA variants, which use only external camera views. Each suite contains

12

shared trials. Bar labels are rounded to the nearest integer; Table I reports precise success rates, SEM, and all baselines.

Method	Pick-and-Place (plain)	Pick-and-Place (w/ Distractors)	Fold/Unfold	Move (far objects)	Overall
$\pi_{0.5}$ [ 23 ]	100.0 $\pm$ 0.0	25.0 $\pm$ 13.1	41.7 $\pm$ 14.9	41.7 $\pm$ 14.9	52.1 $\pm$ 7.3
$\pi_{0.5}$ -Blackout	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0
$\pi_{0.5}$ -Single-External	16.7 $\pm$ 11.2	0.0 $\pm$ 0.0	0.0 $\pm$ 0.0	8.3 $\pm$ 8.3	6.3 $\pm$ 3.5
$\pi_{0.5}$ -Dual-External	50.0 $\pm$ 15.1	16.7 $\pm$ 11.2	16.7 $\pm$ 11.2	16.7 $\pm$ 11.2	25.0 $\pm$ 6.3
ECoT-inspired [ 49 ]	66.7 $\pm$ 14.2	16.7 $\pm$ 11.2	16.7 $\pm$ 11.2	25.0 $\pm$ 13.1	31.3 $\pm$ 6.8
$\pi_{0.5}$ -Single-External-RGB-D	8.3 $\pm$ 8.3	8.3 $\pm$ 8.3	0.0 $\pm$ 0.0	8.3 $\pm$ 8.3	6.3 $\pm$ 3.5

TABLE I: Real-world success (%, mean

\pm

SEM) on

12

shared trials per suite; Overall pools all

48

. Bold / underline : best/second-best; tied scores share the mark.

Fig. 5: Image attention analysis on a Move episode. $\pi_{0.5}$ [ 23 ] attends almost exclusively to the wrist view for the whole episode, while FoveaVLA keeps a steady share on the full external views and draws close-up detail from the foveal crops. Left: attention share per view over time, with three keyframes marked. Right: attention maps at those keyframes; each model is colored relative to its own 99th percentile, so brightness is not comparable across the two panels.

Task	$\pi_{0.5}$ -DROID	FoveaVLA-Dual (ours)
Pick-v1.5	18.1	33.7
Pick-v2-Classic	6.4	9.7
Pick-v2-Filament	7.0	10.6
Pick-v2-RandCam	8.0	5.5
Pick&Place-v2	12.0	17.0
Pick&Place-NextTo-v2	10.3	14.9

TABLE II: Zero-shot success (%) on MolmoSpaces.

\pi_{0.5}

-DROID: reported leaderboard reference; our camera and depth protocol is detailed in Appendix C . Average: unweighted task mean. Bold : better score.

Method	FT	Frame	[email protected] $\uparrow$	mIoU $\uparrow$	Wait (ms) $\downarrow$
$\pi_{0.5}$ -DROID	No	$t$	0.002	0.009	0
PaliGemma + PG	No ∗	$t+w$	0.132	0.207	427
SAM3 + PG	No ∗	$t+w$	0.305	0.347	690
YOLO11s	Yes	$t+w$	0.488	0.492	29
FoveaVLA (ours)	Yes	$t$	0.483	0.465	0

TABLE III: Held-out DROID box localization. FT: box fine-tuning; PG: projected gripper (

*

: fitted on training annotations). Wait: added delay after a new observation arrives. Bold / underline : best/second-best, including ties.

Appendix figures & tables9 assets

Supplementary material from the paper’s appendix.

Appendix

Primitive skills	Required components
Pick, push, pull, press, twist	Gripper and object/tool
Transfer, place, slide, fold, pour, insert, throw, manipulate, shake, strike	Gripper, object/tool, and target region

Appendix

TABLE IV: Required components for interaction-box construction from RoboInter primitive skills.

Setting	Value
Initialization	$\pi_{0.5}$ base checkpoint
Training updates / global batch	$100{,}000$ / $256$
Training hardware	$8$ NVIDIA H100 GPUs
Optimizer	AdamW
Adam $(\beta_{1},\beta_{2})$ / $\epsilon$	$(0.9,0.95)$ / $10^{-8}$
Weight decay / gradient norm clip	$10^{-10}$ / $1.0$

Appendix

TABLE V: Training and inference settings for FoveaVLA-Single and FoveaVLA-Dual.

Tasks	Renderer	External-camera layout
Pick-v1.5; Open-v1; Close-v1	Classic	Original external view plus fixed right-shoulder view
Pick-v2-Classic	Classic	Left/right shoulder views
Pick-v2-Filament; all three Pick&Place-v2 tasks	Filament	Left/right shoulder views
Pick-v2-RandCam	Filament	Randomized views on opposite sides

Appendix

TABLE VI: Camera layouts and RGB renderers for our MolmoSpaces evaluation.

Fig. 6: Representative real-world task instances. Initial and completed states for Pick-and-Place, Pick-and-Place with Distractors, Move, and Fold/Unfold, using the trial identifiers in Tables VII – X . The first three rows show FoveaVLA-Single; the last shows FoveaVLA-Dual from one of its two external views.

Fig. 7: Foveal observations during a successful distractor trial. FoveaVLA-Single executes “Place the black marker in the yellow bowl” (trial $47$ ). Columns show three moments in temporal order. Each column shows the full external RGB view and the foveal RGB crop from the same frame (depth omitted). The crops show the target marker during approach and transport, and the marker in the target bowl after release.

Trial	Objects and initial positions	Language instruction
0	S: red marker (BR); T: blue bowl (FR)	Place the marker in the blue bowl
1	S: black marker (FL); T: blue bowl (BL)	Put the pen in the bowl
2	S: red marker (BL); T: orange bowl (FR)	Place the red marker in the orange bowl
3	S: small cube (FL); T: orange bowl (FR)	Place the small block in the bowl
4	S: blue cube (FR); T: blue bowl (BL)	Place the block in the bowl
5	S: small cube (BR); T: blue bowl (FR)	Place the small block in the blue bowl

Appendix

TABLE VII: Pick-and-Place trial specifications. S denotes the source object and T the target container.

Trial	Objects and initial positions	Language instruction
12	S: blue cube (BL); R: green cup (BR)	Move the cube next to the cup
13	S: small cube (FR); R: white cup (FL)	Move the cube next to the white mug
14	S: blue cube (BR); R: silver cup (BL)	Move the block next to the cup
15	S: blue cube (FL); R: white cup (BR)	Move the blue block near the white mug
16	S: silver cup (BL); R: yellow bowl (BR)	Move the cup near the yellow bowl
17	S: silver cup (FL); R: orange bowl (BL)	Move the silver mug near the bowl

Appendix

TABLE VIII: Move trial specifications. S denotes the source object and R the reference object.

Trial	Objects and initial positions	Language instruction
24	T: blue towel (FL)	Fold the blue towel
25	T: grey towel (FR)	Fold the grey towel
26	T: pink towel (BL)	Fold the pink towel
27	T: blue towel (FL); D: grey towel (BL)	Fold the blue towel
28	T: grey towel (FR); D: pink towel (BR)	Fold the grey towel
29	T: pink towel (BL); D: blue towel (FL)	Fold the pink towel

Appendix

TABLE IX: Fold/Unfold trial specifications. T denotes the target towel and D the distractor. Fold trials start flat; unfold trials start with one half-fold.

Trial	Objects and initial positions	Language instruction
36	S: red marker (BL); T: blue bowl (FM); D: blue marker (BR)	Place the red marker in the blue bowl
37	S: blue marker (BR); T: orange bowl (FM); D: black marker (BL)	Place the blue marker in the orange bowl
38	S: black marker (BL); T: yellow bowl (FM); D: red marker (BR)	Place the black marker in the yellow bowl
39	S: red marker (BR); T: blue bowl (FL); D: blue marker (BL); orange bowl (FR)	Place the red marker in the blue bowl
40	S: blue marker (BL); T: orange bowl (FR); D: black marker (BR); yellow bowl (FL)	Place the blue marker in the orange bowl
41	S: black marker (BR); T: yellow bowl (FL); D: red marker (BL); blue bowl (FR)	Place the black marker in the yellow bowl

Appendix

TABLE X: Pick-and-Place with Distractors trial specifications. S denotes the target marker, T the target bowl, and D all distractors.

Foveated Vision

See all 16 papers

WARP-VLA: Wrist-Camera Adaptation for View-Robust Policy Execution in Vision-Language-Action Models
Oct 8, 2026Junmyeong Lee, Dongmin Shin, Min-Gyu Park +3Robustness of VLA Models Vision-Language-Action Models
Yonsei University. · Korea Electronics Technology Institute(KETI).
VersaCamVLA: Camera-Configurable VLA Policies for Robotic Manipulation
Oct 8, 2026Boyao Han, Chen Shi, Jingjing Qian +2Vision-Language-Action Models Robotic Manipulation
The Chinese University of Hong Kong, Shenzhen · Shenzhen Loop Area Institute · Harbin Institute of Technology, Shenzhen
FocusVLA: Focused Visual Utilization for Vision-Language-Action Models
Mar 30, 2026Yichi Zhang, Weihao Yuan, Yizhuo Zhang +2Vision-Language-Action Models Robotic Manipulation
Harbin Institute of Technology, Shenzhen, China · DaiMon Robotics, China · Nanjing University, Nanjing, China +1

Oct 8, 2026J/K move · Enter open · S save

WARP-VLA: Wrist-Camera Adaptation for View-Robust Policy Execution in Vision-Language-Action Models

Junmyeong Lee, Dongmin Shin, Min-Gyu Park, Wooseok Jeon+2

Yonsei University. · Korea Electronics Technology Institute(KETI).

Robustness of VLA Models Vision-Language-Action Models Test-Time Adaptation of VLA Models Robot Manipulation Benchmarks

Open paper AskPDF arXiv

Despite recent advances in Vision-Language-Action models (VLAs) for robotic manipulation, their performance remains sensitive to changes in camera configuration. The problem becomes more evident in cross-setup deployment, as reproducing the exact camera pose used for training is nearly impossible. Unlike fixed external views, wrist views are more challenging because the camera moves with the robot, causing even small mounting variations to alter fine-grained geometric cues. To address this, we propose WARP-VLA, a camera-view robust VLA for diverse wrist camera configurations. WARP-VLA adopts a Mixture-of-Experts (MoE) architecture where individual experts learn view-specific feature transformations, and a router combines them based on implicit view information. This allows the policy to be deployed without requiring camera extrinsic parameters as additional input. Through experiments on the LIBERO benchmark, WARP-VLA improves the average success rate of pi-0.5 from 39.2% to 78.3% under wrist-view perturbations. The real-robot experiments further show that the feature-level adaptation learned in simulation successfully transfers to diverse deployment settings. To facilitate reproducibility and future research, we release our wrist viewpoint robustness benchmark and a plug-and-play implementation.

Abstract

Foveated Vision

Explore similar work

WARP-VLA: Wrist-Camera Adaptation for View-Robust Policy Execution in Vision-Language-Action Models

VersaCamVLA: Camera-Configurable VLA Policies for Robotic Manipulation

FocusVLA: Focused Visual Utilization for Vision-Language-Action Models