Multi-View Unified Camera Fields: Geometry-Shaped Action-Facing Representations for RGB-Only Multi-Camera VLA Policies

2026-08-03

AI Summary

This work addresses the challenge that multi-camera vision-language-action (VLA) models often produce metrically shallow and cross-view inconsistent action representations in complex contact-rich tasks due to naive view concatenation. The authors propose a multi-view unified pinhole camera framework, employed only during training, which constructs a shared, action-oriented geometric representation in an implicit field by aligning cross-view features through coordinate-based depth queries. This approach achieves cross-view geometric alignment and recoverable metric depth in RGB-only multi-camera VLA policies without incurring additional inference overhead—the first method to do so. Experiments demonstrate state-of-the-art performance, with 98.9% success on LIBERO, a 22.4-point improvement on LIBERO-Plus, and a 23.3-point average gain across six RoboTwin tasks, while also validating effectiveness on a real humanoid robot.

Abstract

Vision-Language-Action (VLA) models have shown strong generalization in robotic manipulation, yet complex contact-rich tasks often benefit from multi-camera observations that jointly capture the end effector, objects, and targets under occlusion. Existing multi-camera VLAs usually concatenate view tokens, leaving action representations weak in metric depth and inconsistent across cameras. We introduce Multi-View Unified Camera Fields (MVUCF), a training-only framework that forms a shared action-facing latent field across views. A coordinate-query depth objective makes metric depth recoverable, while a preprocessing-aware correspondence objective aligns tokens observing the same physical point from different cameras. Both directly shape the hidden states consumed by the action module. After geometry injection, depth, camera calibration, and auxiliary heads are removed, so deployment uses the original RGB-only graph with no extra inference FLOPs. Held-out probes confirm stronger depth recovery and cross-view matching. Under matched GR00T-N1.6 settings, MVUCF reaches 98.9% on LIBERO, improves LIBERO-Plus by 22.4 points, and raises success by 23.3 points across six RoboTwin tasks spanning three action families: touch, move-and-place, and contact interaction. Real-world humanoid experiments further provide evidence of its practical effectiveness under RGB-only deployment.

Problem

• multi-camera VLA
• action representation
• metric depth
• cross-view consistency
• RGB-only policy

Innovation

• Multi-View Unified Camera Fields
• geometry-shaped representation
• metric depth recovery
• cross-view correspondence
• RGB-only VLA