AI-Driven Repair of Missing Automotive Components

Kazim Abbas, Zheng LiuarXiv 2026 · arXiv

TL;DR

This paper introduces a CAD-guided pipeline designed to reconstruct missing sections of damaged automotive components as 3D-printable "repair patches." By leveraging multi-view voxel reconstruction and depth-map comparison, the researchers achieved a median reduction of over 93% in material usage and 89% in printing time compared to replacing entire parts.

What Problem Does This Paper Address?

Replacing damaged automotive components is often a binary choice: find a replacement part or 3D print a new one from scratch. Both options are resource-intensive, especially for discontinued parts or during supply chain disruptions. However, in many cases, the majority of a component remains structurally sound, with damage limited to a specific section.

The challenge lies in partial reconstruction: how do you accurately isolate and reconstruct only the missing geometry? This is technically difficult because "seeing" the missing volume requires perfect alignment between the damaged part and its original CAD model, as well as overcoming "visual hull" limitations—where internal cavities or hidden features are obscured from the camera's view.

Methods

The authors propose a multi-stage pipeline to generate a repair STL file:

  1. Damage Localization: The system renders the original CAD model and the damaged part from identical viewpoints. By comparing depth maps, it identifies "missing" pixels where the damaged surface is significantly deeper than the original or entirely absent.
  2. Multi-View Voxel Reconstruction: Information from multiple camera angles is combined into a 3D voxel grid. The system uses "depth carving" to eliminate empty space and identifies voxels that exist in the intact model but not in the damaged one.
  3. DINOv2 Integration (Experimental): The researchers tested a variant using DINOv2, a pretrained vision transformer, to validate missing regions by comparing visual features between the intact and damaged RGB images.
  4. Mesh Generation: The final voxel set is cleaned (removing noise/small fragments) and converted into a watertight STL mesh using the Marching Cubes algorithm.
Figure 1: Overview of the CAD-guided reconstruction pipeline for generating the repair patch

Datasets

The study utilized a custom dataset of 20 unique automotive CAD models, including clips, brackets, cover caps, and flanges.

  • Synthetic Damage: For each model, a section representing 3.4% to 5.4% of the total volume was removed using a "box cutter" tool with a jagged surface to simulate a realistic break.
  • Trials: The pipeline was tested across 160 trials, varying camera coverage (8 views vs. 4 views) and lighting conditions.
The 20 automotive CAD components used in the evaluation. Green identifies the synthetically removed region.

Key Insights and Ideas

  • The "Extra Material" Pitfall: The most common failure mode was the incorrect addition of material in areas that should be empty. This typically happened in components with hollow openings or internal cavities. If a camera cannot "see" through a hole to verify it is empty, the voxel reconstruction may assume that space is part of the missing solid.
  • AI Isn't Always the Answer: Surprisingly, adding DINOv2 to the pipeline did not improve accuracy and significantly increased processing time (from 5.4s to 63.4s). The authors suggest the filtering stage might have been too restrictive, rejecting valid geometric candidates that the depth-based method had correctly identified.
  • Coverage is Crucial: Moving from 8 distributed views to 4 adjacent views (covering only a 90° range) caused a sharp drop in success rates (from 16/20 to 10/20 passing components). This highlights the necessity of "seeing" the damage from multiple sides to resolve geometric ambiguities.

Key Results

  • Success Rate: 109 out of 160 experimental trials passed the strict accuracy thresholds (Voxel Intersection-over-Union $\ge 0.85$ and Absolute Relative Volume Error $\le 15\%$).
  • Reconstruction Quality: The median IoU for passing trials was 0.902, indicating high fidelity to the original missing geometry.
  • Efficiency Gains: For the 13 verified tests, printing only the repair patch resulted in a median 93.08% reduction in filament use and an 89.39% reduction in printing time.
Comparison of exhaust flange reconstruction under different camera conditions. Red indicates extra material; blue indicates missed material.

Limitations and Future Work

The study was conducted in a controlled synthetic environment with noise-free images and perfect camera positioning. Real-world applications would face challenges like sensor noise, lighting variations, and the difficulty of perfectly aligning a physical 3D scan of a broken part with a digital CAD file.

Future research directions suggested by the authors include:

  • Testing with real-world 3D scans and physical damaged components.
  • Developing methods to better identify and "see" into hidden internal spaces.
  • Exploring different AI architectures that might complement geometric reconstruction more effectively than the DINOv2 implementation used here.

Summary updated

Abstract

Due to supply constraints, replacing damaged automotive components can be expensive and time-consuming, while replacing the entire component can be wasteful when much of it remains usable. This paper presents a pipeline to investigate whether missing geometry can be reconstructed as a repair patch using paired observations of the intact models and synthetically damaged components. Noise-free color, depth, and silhouette images were produced using the camera viewpoints. Depth differences were then used to identify the potential missing regions, while multi-view voxel reconstruction was used to estimate the missing portion. A pretrained DINOv2 model was also added to validate missing candidates and determine whether pretrained visual features could improve reconstruction. We tested the pipeline on 20 automotive CAD models, each with a synthetically removed portion. Each model was tested under four different conditions using two reconstruction methods for a total of 160 trials. The predictions were compared with the missing geometry of the unchanged ground-truth using predefined thresholds of the voxel intersection-over-union over 0.85 and the absolute relative volume error less than 15%. Of the 160 experimental trials, 109 of them satisfied both thresholds. Geometry-only reconstruction passed 16 of 20 components with eight views but only 10 of 20 with four adjacent views. Further analysis showed that 24 out of the 37 failed reconstructions could be categorized as unsuccessful because extra material was incorrectly added to areas that should have been empty. The implementation of DINOv2 into the pipeline did not improve the overall reconstruction accuracy and made prediction generation slower. For 13 verified tests, printing only the repaired sections instead of the entire parts reduced filament use by 93.08% and printing time by 89.39% based on the median results.
BibTeX
@misc{abbas2026ai,
  author = {Kazim Abbas and Zheng Liu},
  title = {{AI-Driven Repair of Missing Automotive Components}},
  year = {2026},
  eprint = {2611.10318},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2611.10318},
}
Figure