New Tool Automates Tedious Collision Mesh Editing in 3D Design Software

Collision meshes—the geometric boundaries that physics engines use to detect object interactions—are essential in modern 3D software, yet editing them remains a labor-intensive chore. A team of researchers has introduced a neuro-symbolic approach that learns editing patterns from user examples and applies them across multiple assets at once.

The Problem with Manual Mesh Editing#

Automated tools can generate collision meshes from visual geometry, but they frequently miss the specific interaction behaviors designers intend. This forces developers to manually refine each mesh individually—a time-consuming task that becomes a major bottleneck as projects grow. The heterogeneity of meshes and the variety of required edits make scaling this work particularly difficult.

How MeshForge Learns from Examples#

The solution, implemented in a tool called MeshForge, frames mesh editing as a programming-by-example problem. Users provide a few demonstrations showing how to transform collision meshes with a similar editing goal. The system then synthesizes a reusable program that captures that intent and applies it automatically to all other meshes in the batch.

Measured Success#

When tested across 24 different editing tasks involving 600 collision meshes, MeshForge successfully synthesized solutions for 23 tasks. On average, it required only 2.2 user demonstrations and took 3.5 seconds to complete synthesis—making it practical for real-world design workflows.

The work demonstrates how program synthesis can ease the cognitive load on developers working with 3D assets, extending the reach of automation beyond traditional code and interfaces into the spatial domain.