3D modeling in robotics with LLMs: now feasible
Anthropic
Google/DeepMind
OpenAI
A year ago, large language models failed to generate a correct URDF model for a robotic arm from STL files. Today, newer models like Opus 4.8 and Fable 5 can produce a nearly functional model in a couple of hours, though some manual tweaking is still required.
An engineer tested whether modern LLMs could handle the tedious task of creating a URDF model for a DIY robotic arm (PhantomX Pincher clone) from STL files. A year ago, models like Sonnet 4, Opus 4, Gemini 2.5, and GPT-4.1 produced random, misaligned parts. Recently, the same prompt was given to Opus 4.8 and Fable 5. Opus 4.8 correctly scaled the STL parts and identified their roles, but final joint orientations were still off. Fable 5, after receiving a screenshot of the flawed assembly, correctly diagnosed and fixed the issues, producing a nearly correct URDF. The engineer estimates the task now takes a couple of hours instead of several days, with minor manual adjustments still needed.
- Сокращения
- LLM = Large Language Model — большая языковая модель
- URDF = Unified Robot Description Format — унифицированный формат описания робота
- ROS = Robot Operating System — операционная система для роботов
- STL = Standard Triangle Language (or Stereolithography) — стандартный формат 3D-моделей
- DIY = Do It Yourself — сделай сам
- UI = User Interface — пользовательский интерфейс
Source: Habr — хаб ИИ —
original