A portable knowledge vault for working with an SO-ARM100 leader/follower pair using LeRobot. It captures hardware facts, calibration files, copy-paste workflows, experiment post-mortems, and one-lesson-per-note troubleshooting, so the work can move between machines:
| Machine | Role |
|---|---|
| Ubuntu desktop (RTX 5060 Ti 16 GB) | record + train locally (ACT runs of June/July 2026 live here) |
| MacBook Pro M3 | record datasets anywhere (talks, demos), push to the Hub |
| Hugging Face Jobs GPU | train VLAs (SmolVLA, pi0) when the desktop is not available |
It is an Obsidian vault (open the folder as a vault for wiki-links and the graph)
and also a plain Markdown repo readable on GitHub. CLAUDE.md is the entry point for AI coding agents.
- 00-Start-Here — map of content
- timeline — what happened, when (May → Sep 2026)
- troubleshooting — symptom → fix
CLAUDE.md agent entry point: hard facts, read order, conventions
hardware/ arms, calibration, cameras, Feetech firmware, breakage history
machines/ per-machine profiles + .env files (ports, cameras, device, runner)
workflows/ 01..12: environment → ports → calibrate → cameras → teleop → record → train → eval → hub sync
experiments/ dated post-mortems (Lego skeleton ACT, Fool's mate ACT) + next experiment (medicaments VLA)
lessons/ one insight per note
tools/ env.sh + thin shell wrappers over lerobot-* commands, bench.py
assets/ calibration JSONs, camera probe images, exact train configs
archive/ original guides (verbatim), VLA roadmap, reading list
presentation/ demo plan for the "From Token to Torque" VLA talk
git clone <this repo> LeRobot-Second-Brain
cd LeRobot-Second-Brain
source tools/env.sh desktop # or: macbook
tools/find.sh # lerobot-find-port + lerobot-find-cameras
tools/teleop.sh # sanity check
tools/record.sh my_task "Pick the medicament box and place it in the tray" 50Nothing in this repo uploads anything. Hub pushes are opt-in flags documented in workflows/11-hub-sync.
Bóveda de conocimiento del brazo SO-100 con LeRobot: hardware, calibración, comandos listos para copiar, experimentos (qué funcionó y qué no) y lecciones. Sirve para trabajar igual en el escritorio con GPU, en la MacBook, o entrenando en GPUs de Hugging Face. Abrir como vault de Obsidian.