Back to Feed
Computer Vision / Reinforcement Learning

Building 3D Designs One Step at a Time

Original: CADENA: Stepwise CAD Reverse Engineering
By Soslan Kabisov, Gennadiy Savrasov, Maksim Elistratov, Antonio Rodriguez, Daniil Ignatiev, Nikita Gavrilov, Rustam Uzdenov, Alexey I. Boyko, Igor Pasechnik, Anton Konushin, Andrey Kuznetsov, Dmitrii Zhemchuzhnikov Published 2026-08-01

Listen to the summary

Uses a voice available on your device

Audio options
On this page

Key Takeaways

  • CADENA treats design reconstruction as a series of individual decisions rather than generating the entire program at once.
  • The system uses a vision model to look at both the original goal and the current partial design, allowing it to spot and fix differences in real time.
  • CADENA achieved record breaking performance across five industry standard design benchmarks.
  • The approach uses hint points to help the model decide what to do next when multiple parts of a design need attention.

Summary & Methodology Analysis

Traditional artificial intelligence tools for design often fail because they try to create a whole three dimensional object in one go, which leads to mistakes that add up over time. CADENA fixes this by breaking the process into smaller steps. The model treats design as a sequential decision problem, where it adds one operation at a time. To ensure accuracy, it uses an observation function that renders the original design and the current progress into a single image. This visual feedback allows the system to see exactly where the current build deviates from the intended target.

Interactive System Flowchart

Click diagram to expand and zoom

Cross-Examination & FAQs

A deeper dive clarifying mechanics, constraints, and baseline evaluations.

Q1. What is the main goal of this research?

The researchers aimed to improve how artificial intelligence turns images into three dimensional computer aided design files by correcting errors as they happen.

Q2. How does CADENA differ from previous methods?

Older systems try to build the design in a single pass, whereas CADENA builds it piece by piece while checking its work against the target at every stage.

Q3. What kind of models were used to build this system?

The system relies on a vision language model called Qwen2-VL as its policy to interpret visual data and guide the construction process.

Q4. What role does reinforcement learning play in this system?

The authors used reinforcement learning with a technique called Group Relative Policy Optimization normalization to update and improve the policy during training.

Q5. How does the system know which part of the design to edit next?

When multiple areas of the design differ from the goal, the model uses a text based hint point to focus on a specific region for the next step.

Q6. What datasets were used to validate the performance of CADENA?

The system was tested against five datasets including DeepCAD, Fusion 360, MCB, CADENA-Bench, and BenchCAD.

Q7. Are there any limitations to how the system observes the design?

Yes, parts of the geometry that are hidden or occluded from the view are not visible to the system, and it may struggle with countable features.

Q8. Can the system recover from mistakes made early in the process?

No, the paper states that errors occurring in early frames cannot be recovered by the current system.

Q9. How does the system handle features it does not recognize?

The system attempts to approximate out of vocabulary features using the available operations in its domain specific language.