3DGS Generates Training Data for Wrist-Camera VLA Policies, Validated on Reaching in Motherboard Disassembly

What happened

This article covers the paper "Robot-Aligned 3D Gaussian Splatting for Precise Reaching with Wrist-Camera VLA Policies," posted to arXiv and dated October 26, 2026. The authors are Zuoxu Wang and Xiao Liang. A VLA is a robot policy that outputs actions from images and language instructions.

The paper points out that data collection is a major barrier to VLA training. It says this is especially serious for tasks that require precise positioning of the end effector (the tip of the robot) and alignment between camera images and actions.

The target application is electronic waste disassembly. In this environment, visually similar parts are packed densely, and many parts are thin, reflective or structurally complex. The paper says that this makes it hard to build accurate per-object meshes and contact models, so fully digital-twin simulation is difficult.

The authors therefore focused on the "reaching motion" before touching an object. They reconstruct the real workspace with 3DGS and render a variety of wrist-camera observations from robot-aligned poses. These are combined with smooth, structured reaching trajectories to form visual and action supervision data for VLA fine-tuning.

They also add an instruction-conditioned "reach verifier" to the VLA. It estimates proximity to the target and is intended to reliably stop the motion before physical contact occurs.

The experiments were conducted on motherboard disassembly. The authors report that precise reaching was achieved using only wrist-camera observations, without an external side camera.

How it works

The pipeline is as follows. (1) Reconstruct the real workspace with 3DGS. (2) Render many wrist-camera images from robot-aligned poses. (3) Combine them with smooth reaching trajectories to create image-action pairs. (4) Fine-tune the VLA on this data. (5) The reach verifier estimates proximity to the target and terminates the motion before contact.

A distinguishing feature is that it increases training data through visual reproduction alone, without building meshes or contact models. Contact-rich manipulation is out of scope; the method is limited to reaching.

By the numbers

Why it matters

This is an example of using 3DGS for Real2Sim data generation centered on visual reproduction. It suggests that training data could be created from short captures even in settings with reflective or thin objects that are hard to mesh.

For those considering robot deployment in settings with densely packed parts, such as disassembly, sorting and inspection, it is a useful reference as an option for lowering data collection costs. However, the results are currently at the paper stage.

How to try it

1. Check the paper page and project site for the method's figures and videos, and whether code has been released.
2. To try it in your own environment, first reconstruct a small workbench with 3DGS. Next, render viewpoints equivalent to a wrist camera and check whether they can be used as input for your existing VLA fine-tuning.

Caveats and open questions

• The sources are limited to the arXiv abstract and submission information; details of the paper's full text have not been confirmed. The reconstruction procedure, number of captured images, number of evaluation trials and similar details are not confirmed in the sources.
• The success rates are values reported by the authors; third-party verification is not confirmed in the sources.
• Validation is limited to reaching motions in motherboard disassembly. The success rates for contact-involving grasping or for disassembly itself are not stated in the sources.
• "Less than 10 minutes" refers to active human collection time. The computation time required for reconstruction and generation is not confirmed in the sources.
• The release status of code and data is unconfirmed. The abstract gives a project site URL, but its contents have not been checked.

Glossary

VLA

Vision-Language-Action. A model that takes images and language instructions as input and outputs robot actions.

3DGS

3D Gaussian Splatting. A 3D reconstruction method that represents a scene with many Gaussian distributions and renders it quickly.

Wrist camera

A camera mounted on the wrist of a robot arm. It provides a view from the tip of the arm.

π0.5

The VLA model used in the paper. Details are not confirmed in the sources for this article.

Reach verifier

An additional module that estimates how close the robot is to the instructed target and decides when to terminate before contact.

Sources

Primary sources (publisher, paper, repository)

• Robot-Aligned 3D Gaussian Splatting for Precise Reaching with Wrist-Camera VLA PoliciesarXiv（arxiv.org）・en

The X post that surfaced this story

• View @OWW's post on Xx.com

This article was generated in Japanese by AI from the sources listed (primarily primary sources) and published automatically, without prior human review, after automated checks such as source matching (how we use AI). Figures and specifications are as stated by the publishers and have not been independently verified by us. Please consult the original sources. To report an error, use our contact form; we will review it promptly and correct it. This page is an AI translation of the Japanese original. If anything differs, the Japanese version prevails. Japanese original