VersaCamVLA: Camera-Configurable VLA Policies for Robotic Manipulation

Forum topic · 小凯 · 2026-10-10

Summary

VersaCamVLA is a camera-configurable framework for Vision-Language-Action (VLA) models in robotic manipulation, addressing the brittleness that fixed camera setups introduce at deployment. The method decouples camera-set representation from action learning by learning a unified scene-token interface that maps an arbitrary, variable number of posed RGB views into fixed-size latent scene tokens. Training combines multi-signal target-view prediction with Wrist-Augmented Pose Sampling (WAPS), which exploits natural wrist-camera motion to obtain pose diversity without extra cost. At deployment, a lightweight spatial encoder injects these compact scene tokens into a pretrained foundation VLA model as supplementary visual conditioning, requiring neither explicit 3D perception nor novel-view rendering. Experiments on RoboTwin, LIBERO, and real robot platforms show VersaCamVLA consistently outperforms prior VLA methods and direct multi-view baselines, maintaining robust performance across varying camera counts and unseen camera poses.

Paper Overview

Field: Computer Vision / Robotics Authors: Boyao Han, Chen Shi, Jingjing Qian, ZhuoTan Tian, Li Jiang Date: 2026-10-08 arXiv: 2610.12451

Abstract (translated)

Vision-Language-Action (VLA) models have emerged as powerful foundations for robotic manipulation, but their reliance on fixed camera configurations during training makes them brittle to changes in camera count or pose during deployment. To overcome these limitations, the authors propose VersaCamVLA, a camera-configurable framework that decouples camera-set representation from action learning.

VersaCamVLA learns a unified scene-token interface that maps an arbitrary, variable set of posed RGB views into fixed-size latent scene tokens. This is achieved via multi-signal target-view prediction and Wrist-Augmented Pose Sampling (WAPS), which leverages natural wrist-camera motion for free pose diversity. At deployment, a lightweight spatial encoder injects these compact scene tokens into a pretrained foundation VLA as supplementary visual conditioning, without requiring explicit 3D perception or novel-view rendering.

Key Contributions

Camera-configurable policies: Robust to changes in both camera count and camera pose between training and deployment.
Unified scene-token interface: Variable posed RGB views are compressed into fixed-size latent scene tokens.
Wrist-Augmented Pose Sampling (WAPS): Gains camera-pose diversity for free from natural wrist-camera motion.
Plug-in design: A lightweight spatial encoder adds scene tokens to pretrained foundation VLAs without explicit 3D perception or novel-view rendering.

Results

Experiments on RoboTwin, LIBERO, and a real robot platform demonstrate that VersaCamVLA consistently outperforms prior VLA methods and direct multi-view baselines, maintaining robust performance across different camera counts and unseen camera poses.

Paper link: https://arxiv.org/abs/2610.12451

--- *Auto-collected on 2026-10-10.*