Teach a Robot to Make Pizza in 48 Hours
October 30 – November 1, 2026 · [Venue / City] · 8–12 teams · 2–4 participants per team
Each team has two kits: a LeRobot SO-101 pair and an ElRobot pair. Pick the arm directory for the stack you are on. They do not share a Python environment.
| Arm | Directory | Stack |
|---|---|---|
| LeRobot SO-101 | so101 | uv + LeRobot 0.6.1 |
| ElRobot | elrobot | NormaCore Station binary |
SO-101 train → deploy: DevLab · Train and deploy · ElRobot train → deploy: Train and deploy
Status: Draft v0.1.
From so101. Full steps: so101/README.md.
1. Install uv once if uv is not a command (DevLabs Jupyter, a fresh machine). The installer writes ~/.local/bin/uv and updates the shell profile. The terminal you installed in still needs source.
curl -LsSf https://astral.sh/uv/install.sh | sh
source "$HOME/.local/bin/env"2. Create the environment. DevLabs’ (base) Python is /usr/bin/python3.12 and has no Python.h. On Linux, LeRobot pulls pynput, which pulls evdev, and that package compiles from source. uv sync has to use uv’s own CPython 3.13 (so101/.python-version). hf is installed inside the venv.
cd so101
unset UV_PYTHON
rm -rf .venv
uv python install 3.13
uv sync --managed-python --python 3.13
uv run hf auth login
cp station.env.example station.env
cp ../.env.example ../.envThe sync log must name an interpreter under ~/.local/share/uv/python. A line that says /usr/bin/python3.12 means this environment will fail the same way again.
That install includes LeRobot 0.6.1 with robot CLIs (core_scripts + feetech) and policy extras ACT (core), SmolVLA, Diffusion, π0 / π0.5, and GR00T N1.7.
| Step | Command (from so101) |
|---|---|
| Find USB ports | bash scripts/list_ports.sh |
| Calibrate | bash scripts/calibrate.sh follower then leader |
| Teleop | bash scripts/teleoperate.sh |
| Record 50 demos | bash scripts/record.sh |
| Preview in Rerun | bash scripts/viz_dataset.sh |
| CPU ACT smoke (20 steps) | bash training/scripts/train_local.sh act lerobot/pusht 20 |
| DevLab fine-tune (recommended) | DevLab guide |
| Job canary (500 steps) | bash training/scripts/run_serverless.sh --debug act |
| Job ACT | bash training/scripts/run_serverless.sh act |
| Copy checkpoint from FS mount | bash training/scripts/download_checkpoint.sh <job-name> |
| Run policy | bash scripts/rollout.sh <checkpoint> |
Robot work uses the uv venv in so101. train_local.sh is a CPU smoke. GPU fine-tune is a DevLab (recommended, RTX PRO 6000 by default) or a Job. L40S and H100 are the other platforms.
Upstream: LeRobot install · SO-101 · real-robot IL
From elrobot. Full steps: elrobot/README.md.
cd elrobot
bash scripts/install.sh
bash scripts/run.shinstall.sh downloads Station v0.1.0-beta.10 and shallow-clones that tag into norma-core/. Serials and the task sentence live in elrobot.env. The SmolVLA trainer in the tag is 6-DoF and needs CUDA; ElRobot has 8 servos.
| Step | Command (from elrobot) |
|---|---|
| Install | bash scripts/install.sh |
| Config | cp elrobot.env.example elrobot.env |
| Start Station | bash scripts/run.sh |
| Calibrate follower | bash scripts/calibrate.sh |
| Teleop | bash scripts/teleop.sh |
| Export parquet | bash scripts/export_dataset.sh <START> <END> |
| Convert to LeRobot v2.1 | data collection (scripts/export_to_lerobot.py) |
Upstream: norma-core
| Document | Covers |
|---|---|
| SO-101 setup | Install, calibrate, collect, train, deploy |
| ElRobot setup | Install, calibrate, collect, train, deploy |
| SO-101 data collection | lerobot-record, H.264, LeRobot v3 |
| ElRobot data collection | Station tags, parquet export, LeRobot v2.1 conversion |
| SO-101 DevLab | Recommended GPU workspace: data, fine-tune, policy server |
| SO-101 train and deploy | CPU smoke, DevLab, Jobs, rollout |
| ElRobot train and deploy | SmolVLA fine-tune, deploy, evaluation |
The two datasets are separate. Norma's SmolVLA trainer reads Station parquet. elrobot/scripts/export_to_lerobot.py writes a LeRobot v2.1 folder from those parquets. SO-101 training reads LeRobot v3.
Nothing under data/ is committed.
| Stack | Where recordings land |
|---|---|
| ElRobot export | data/elrobot-pizza.parquet and data/elrobot-pizza.mp4 (prefix ../data/elrobot-pizza in elrobot.env) |
| SO-101 | ~/.cache/huggingface/lerobot/${HF_USER}/${DATASET_NAME}_<timestamp>/ |
- Format: Teleoperate → Collect Data → Fine-Tune → Deploy → Compete
- Three challenges: Precision Topping (30 pts) · Recipe Assembly (40 pts) · Mystery Pizza (30 pts) — 100 pts total
- Primary KPI: ≥70% of participating teams successfully deploy one autonomous learned policy to a real robot
- Platforms: each team uses both the SO-101 kit (LeRobot) and the ElRobot kit (NormaCore Station)