SpaceVLA: Spatially Grounded VLA for Robotic Manipulation with User-Authored Grasp and Place Anchors

2026-08-06

AI Summary

This work addresses the limited ability of existing vision–language–action (VLA) models to explicitly model spatial manipulation intent, which hinders precise control over grasping and placing locations. To overcome this, the authors propose a visual intent anchoring mechanism that renders user-specified grasp and place regions—designated via an XR interface—as image-space overlay masks and integrates them into the VLA framework, enabling joint modeling of linguistic instructions and pixel-level spatial intent. Building upon the OpenVLA-7B architecture and employing LoRA-based fine-tuning, the method is trained on 200 annotated demonstration trajectories in a Unity simulation environment to predict 7-DoF incremental actions from RGB images and language commands. Closed-loop experiments demonstrate a grasp success rate of 91.25%, with average grasp and place errors of 0.5 cm and 0.7 cm, respectively, substantially improving manipulation accuracy and controllability.

Abstract

Vision-language-action (VLA) models follow language commands but often lack explicit spatial intent for manipulation. We present Visual Intent Anchors, an XR pipeline that lets users specify grasp and placement regions and renders them as image-space overlays for VLA control. We collect 200 Unity pick-and-place demonstrations and fine-tune OpenVLA-7B with LoRA on temporally subsampled annotated observations. The policy predicts tokenized 7-DoF incremental actions from marked RGB observations and language. We evaluate the policy in closed-loop Unity trials, achieving a grasp success rate of 91.25% and mean grasp and placement errors of 0.5 cm and 0.7 cm, respectively.

Problem

• Vision-language-action
• spatial intent
• robotic manipulation
• grasp and place

Innovation

• Visual Intent Anchors
• Vision-Language-Action (VLA)
• Spatial Grounding
• Robotic Manipulation
• LoRA Fine-tuning