---
license: apache-2.0
library_name: jax
tags:
- robotics
- vision-language-action
- vla
- flow-matching
- manipulation
- garment-folding
- lehome-challenge
- sim-to-real
pipeline_tag: robotics
datasets:
- lehome/dataset_challenge_real
---

# LeHome Challenge 2026 — Real-World Policy

This is my policy for the **real-world final** of the [LeHome Challenge 2026](https://github.com/lehome-official/lehome-challenge) (ICRA 2026), a competition on bimanual garment folding with a physical LeRobot SO-ARM101 setup. It placed **2nd**.

It is a π0.5-based vision-language-action policy, trained as behavior cloning on a mix of organizer teleop data, my own teleop / DAgger corrections, and heavily-augmented simulation replays. It builds on the [simulation-round policy](https://huggingface.co/IliaLarchenko/lehome_sim).

For how it works and how to run it, please refer to the code repository and the tech report.

- 💻 **Code:** https://github.com/IliaLarchenko/lehome_solution
- 📄 **Tech report:** https://arxiv.org/abs/2606.27163
- ✍️ **Blog post:** https://ilialarchenko.com/projects/lehome2026
- 🤖 **Simulation-round policy:** https://huggingface.co/IliaLarchenko/lehome_sim