HN Showcase
Back to browse
Letting an LLM write robot programs

Letting an LLM write robot programs

by encrux·Apr 7, 2026·3 points·0 comments
Visit ProjectView on HN

AI Analysis

●●●BangerBig BrainNiche Gem

Using symbolic object names prevents dangerous coordinate hallucinations during execution.

Strengths
  • •Using symbolic object names prevents dangerous coordinate hallucinations during execution.
  • •MuJoCo physics simulation runs entirely in the browser via WASM.
Weaknesses
  • •Public demo relies on external free-tier API with rate limiting.
  • •Real industrial application requires more than just trajectory planning logic.
Category
AI/ML
Target Audience

Robotics engineers, AI researchers

Similar To

Code as Policies · Google RT-2

Post Description

Try the demo here: https://llm-trajectory.boesch.dev/

This is a write-up of a system I designed at a previous job. It's a natural language interface for industrial robots. The demo runs MuJoCo in the browser via WASM.

The demo uses a free-tier LLM API with rate limiting, so it might be slow or unavailable under load, but feel free to check out the code and hook up your own inference service: https://github.com/Encrux/llm_trajectory

Similar Projects