xle-hack

Hackathon weekend stack for Physical AI Hack World Tour — SF (May 9-10, 2026, Founders Inc). MakerMods + XLeRobot.

What lives here

• Modal trainer — fine-tune lerobot policies (SmolVLA primary, X-VLA secondary) on rented compute. Image, volume, train function.
• Dataset prep — pulls + filters the AllenAI MolmoAct2-SO100/101 corpus to a bimanual-SO101 subset for pre-fine-tuning.
• Perception layer — Molmo / Gemini ER as a "find the thing" wrapper, feeding pixel coords or pose targets into the policy prompt.
• Demo CLI — prompt → action wrapper. The thing we present.

Companion repos

• ScavieFae/xlerobot-onboard — public install script + Claude skill we ship to fellow hackers.
• new-theory/so101-dev (private) — broader SO-101 work; this repo is hackathon-scoped.

Quick start

uv venv --python 3.12 .venv && source .venv/bin/activate
uv pip install -e .

Then see modal/ for training and data/ for dataset prep.