CoRe-VLA: Preserving Cross-View Coordination in VLAs under Camera Shifts

Tianhang Pan; Xuanhao Wang; Yiwen Pang; Bo Zhou; Jun Yang; Min-Ling Zhang; Shimin Di

cs.ROSep 29, 2026

CoRe-VLA: Preserving Cross-View Coordination in VLAs under Camera Shifts

Authors: Tianhang Pan, Xuanhao Wang, Yiwen Pang, Bo Zhou, Jun Yang, Min-Ling Zhang, Shimin Di

Organizations: Southeast University, Nanjing, China · National Center of Technology Innovation for EDA, Nanjing, China

Abstract

VLAs combine pretrained vision-language representations with action generation to enable language-guided control across diverse tasks, becoming a mainstream paradigm in embodied intelligence. However, multiple studies have reported VLA's substantial declines in task success under camera shifts, revealing a key vulnerability that limits reliable deployment. To address this vulnerability, existing methods collect paired observations of the same scene from different viewpoints to fine-tune the VLA or train visual adaptation modules. Unfortunately, they require additional data collection and VLA training costs. In this paper, we first identify \emph{cross-view coordination breakdown} under external camera shifts: the robot may rely too heavily on wrist-view cues and consequently execute subtasks in the wrong order when losing global view. Motivated by this, we propose CoRe-VLA, a plug-and-play framework requiring neither additional multi-view data collection nor VLA fine-tuning, which can incorporate with exsiting VLAs. It reconstructs a scene point cloud and renders the observation from the VLA's training viewpoint to restore cross-view coordination. In CoRe-VLA, Render-to-Camera (R2C) Restoration reduces rendering-induced visual degradation, while Execution-Trajectory-Conditioned Alignment (ETCA) reduces robot idle time and mitigates motion conflicts during asynchronous execution. Experiments on 5 real-robot tasks, LIBERO-100 and LIBERO-Plus demonstrate CoRe-VLA substantially improves task success across mainstream VLAs under camera shifts. For example, CoRe-VLA raises PI0.5's success rate from 13.3% to 83.3% at a 1.6m camera shift in real-robot environment.

Figures & tables

Figure 1: (a) The external camera shifts while the wrist-camera mounting remains unchanged. The accompanying curves compare task success rates on 5 real-robot tasks with and without CoRe-VLA. (b) Under the shifted view, the robot grasps the tape measure visible in the wrist view while the basket lid remains closed, violating the required subtask order. (c) The core idea of CoRe-VLA: using point-cloud reconstruction and re-rendering to restore the external observation and recover cross-view coordination.

ID	Category	Task instruction
T1	Basic	Pick up the green glue stick and place it in the pink cup.
T2	Long-horizon	Place the yellow tape measure and the pink cup into the basket in sequence.
T3	Long-horizon	Open the lid, then pick up the yellow tape measure and place it in the basket.
T4	Spatial	Pick up the green glue stick and place it on the side of the pink cup farther from the basket.
T5	Spatial	Pick up the object closest to the basket and place it inside.

Table 1: Real-world task suite used to diagnose VLA failures under external-camera shifts. T1 evaluates basic manipulation, T2–T3 long-horizon execution, and T4–T5 spatial reasoning.

Figure 2: Outcome distributions on five real-robot tasks under external-camera shifts for (a)

\pi_{0.5}

, (b) OpenVLA-OFT, and (c) GR00T. Failure Type 1 denotes failure to complete any meaningful subtask, while Failure Type 2 denotes local manipulation inconsistent with the global task state.

Figure 3: Overview of CoRe-VLA, comprising three components: geometric viewpoint restoration, R2C for image restoration, and ETCA for asynchronous control.

Figure 4: ETCA matches recorded robot motion to a new action chunk and selects where execution resumes.

Shift (m)	VLA	Metric	T1		T2		T3		T4		T5		Overall
Shift (m)	VLA	Metric	Raw	CoRe	Raw	CoRe	Raw	CoRe	Raw	CoRe	Raw	CoRe	Raw	CoRe
0.0	$\pi_{0.5}$	SR	100.0	83.3	83.3	83.3	100.0	100.0	100.0	100.0	100.0	100.0	96.7	93.3
	$\pi_{0.5}$	F2	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0	0.0
	OpenVLA-OFT	SR	83.3	83.3	83.3	100.0	100.0	83.3	100.0	100.0	83.3	83.3	90.0	90.0
	OpenVLA-OFT	F2	0.0	0.0	16.7	0.0	0.0	0.0	0.0	0.0	16.7	16.7	6.7	3.3
	GR00T	SR	83.3	66.7	66.7	66.7	66.7	66.7	100.0	83.3	100.0	100.0	83.3	76.7

Table 2: Real-robot outcome rates (%) across external-camera shifts. Raw and CoRe denote the shifted view and CoRe-VLA-restored view, respectively. SR denotes success rate, and F2 denotes the proportion of all trials classified as Failure Mode 2 (task-state inconsistency).

Figure 5: Real-robot performance across 5 manipulation tasks under external-camera shifts. (a) Task success rates. (b–d) Outcome proportions for

\pi_{0.5}

, OpenVLA-OFT, and GR00T, respectively. Failure Types 1 and 2 denote local execution failure and task-state inconsistency, respectively.

G_{0}

and

G_{s}

denote raw external-camera observations captured at the nominal and shifted positions, respectively, while

G_{r}

denotes observations transformed by CoRe-VLA.

Figure 6: Task success rates on LIBERO-100 under external-camera shifts, without CoRe-VLA (solid) and with CoRe-VLA (dashed).

Method	Backbone	SR (%) $\uparrow$
VLA based methods
Cross-View AC ( Huang et al., 2026b )	$\pi_{0.5}$	87.2
GAM ( Han et al., 2026 )	DA3-Giant	83.1
Anchor-Align ( Dalal et al., 2026 )	Prismatic- Qwen2.5-0.5B	96.3
AVA-VLA ( Xiao et al., 2026 )	OpenVLA-OFT	69.4
SRPO ( Fei et al., 2026a )	OpenVLA	83.4

Table 3: Success rates (%) on LIBERO-Plus Camera.

Figure 7: Effect of R2C Restoration on

\pi_{0.5}

task success under external-camera shifts.

Method	Latency (ms)	Success (%)	Completion (s)	Idle (s)
$\pi_{0.5}$	244	96.7	47.2	22.5
$\pi_{0.5}$ + ETCA	244	96.7	34.4	9.0
$\pi_{0.5}$ + CoRe-VLA w/o ETCA	478	93.3	64.5	39.1
$\pi_{0.5}$ + CoRe-VLA	478	93.3	43.7	13.8

Table 4: Effect of ETCA on execution efficiency.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Task	Success criterion
T1	The robot grasps the green glue stick and releases it inside the pink cup.
T2	The robot places and releases the yellow tape measure inside the basket, then places and releases the pink cup inside the basket.
T3	The robot first opens the lid so that it no longer covers the basket, then places and releases the yellow tape measure inside the basket.
T4	The robot grasps the green glue stick and releases it on the side of the pink cup farther from the basket, within a $\pm 45^{\circ}$ sector about the basket-to-cup direction.
T5	The robot grasps the object initially closest to the basket and releases it inside the basket.

Appendix

Table 5: Success criteria for the five real-robot tasks.

Figure 8: Representative real-robot execution sequences under nominal (Raw), shifted, and CoRe-VLA-restored external views. Arrows indicate action order; check marks and crosses denote success and failure, respectively.

Figure 9: Qualitative ablation of the R2C training losses. From left to right: original image, degraded input, and restorations using color loss, color and pixel losses, and the full objective including perceptual loss.

Explore similar work

CoRE-VLA: Towards Scalable and Robust Vision-Language-Action Modeling via Conditional Routing of Experts
Jul 4, 2026Haozhe Zhang, Sixian Li, Yifei Zhang +5Diffusion-Based Vision-Language-Actions Robotic Manipulation
Zhejiang University · Shanghai Innovation Institute · Fudan University +2
Does Adversarial Training Improve Generalization in Multi-View VLAs? Revealing and Mitigating View Collapse
Sep 27, 2026Futa Waseda, Shuhei Kurita, Isao EchizenDiffusion-Based Vision-Language-Actions Vision-Language Model Adaptation
National Institute of Informatics Tokyo, Japan
Cross-View Action Consistency for Camera-Robust Vision-Language-Action Policies
Aug 7, 2026Bingqi Huang, Bingchuan Wei, Xuan Wang +2Flow-Matching Vision-Language-Action Photometric Supervision
Tsinghua University · Informatics Institute, University of Amsterdam & Faculty of Science, Vrije Universiteit Amsterdam

Jul 4, 2026, cs.ROJ/K move · Enter open · S save

CoRE-VLA: Towards Scalable and Robust Vision-Language-Action Modeling via Conditional Routing of Experts

Haozhe Zhang, Sixian Li, Yifei Zhang, Zezheng Huai+4

Zhejiang University · Shanghai Innovation Institute · Fudan University+2

Diffusion-Based Vision-Language-Actions Robotic Manipulation

Open paper AskPDF arXiv

Vision-language-action (VLA) models have advanced generalist robotic manipulation, yet real-world deployment reveals a fundamental challenge: robots are equipped with diverse and heterogeneous sensor configurations, auxiliary sensors can fail unexpectedly during operation, and different robot embodiments often lack certain sensors by design. A unified policy that can exploit auxiliary perceptual inputs when available while remaining reliable under sensor absence, whether incidental or by design, is therefore essential for practical deployment. However, existing VLA policies couple action generation to a fixed sensor set through shared dense computation, making them brittle when sensors are missing and limiting their ability to specialize across diverse tasks and long-horizon behaviors. We propose CoRE-VLA, a scalable and robust VLA framework that formulates action generation as context-conditioned sparse computation. Sensor availability gates modality-specialized experts, enabling graceful degradation under missing sensors without retraining. Task intent further routes action-side representations to task-relevant experts, improving specialization across diverse tasks and long-horizon subgoals. While the framework is designed to accommodate different auxiliary sensors, we focus on depth as a representative and practically important auxiliary modality in our experiments. Experiments on LIBERO, RoboCasa GR1 Tabletop, and real-world dual-arm manipulation show that CoRE-VLA achieves strong results on long-horizon and multi-task benchmarks, and outperforms both a dense-action-generator ablation and a strong pretrained VLA baseline, including in zero-shot generalization to unseen scenarios. Modality analysis shows that CoRE-VLA can exploit auxiliary depth when available while remaining robust when depth is unavailable during deployment.

Abstract

Explore similar work

CoRE-VLA: Towards Scalable and Robust Vision-Language-Action Modeling via Conditional Routing of Experts

Does Adversarial Training Improve Generalization in Multi-View VLAs? Revealing and Mitigating View Collapse

Cross-View Action Consistency for Camera-Robust Vision-Language-Action Policies