Advertisement
zk0.bot is an open-source federated-learning network for robot arms: independently owned SO-101 arms train one shared SmolVLA model on data that never leaves the owner's desk — only model updates move. This is session 3 of 6 in the zk0.bot Workshop Series at Pirin.ai's Build Table, walking the arc from a first assembled node to the M1 experiment: local-only vs federated vs centralized training on a real fleet, held-out sites, pre-registered bars.You leave with your first batch of standard-task episodes — successes and labelled failures — recorded to the federation's quality bar, plus your site's coverage map and episode target.
What you'll leave with
A first episode batch meeting the quality bar and a personal coverage map (what your site adds: environments, objects).
Workshop outline
Data beats parameters — the receipts: community pretraining lifted SmolVLA on SO-100 from 51.7% → 78.3% with no architecture change; generalization scales with environments/objects, while repeat takes saturate.
The standard task: the sorting family, defined precisely — layout, object set, success criteria — fixed across the cohort so episodes are comparable and M1 stays readable.
Episode anatomy: framing, consistent task labels, the full attempt (approach, grasp, place, settle). An episode is a unit of evidence.
Operator protocol consistency: same motions, pace, recovery moves — operator style is a variable, and uncontrolled style (velocity multimodality) can hurt.
Recording at training fidelity: wrist camera on; LeRobot Real-Time Chunking at rollout so execution has no stale chunk seams; record the way the policy will run.
Failures are data — labelled, not deleted: near-miss grabs and dropped places, judge-labelled (frontier-VLM pattern), with sampled human spot-checks keeping the judge honest.
Dataset hygiene before volume: cleanup, judge labels applied, dedup by scene similarity; contribution weighting follows verified quality, not raw counts.
Sprint: record on the standard task live, room reviewing framing and labels; set the target toward your site's M1 share (50–225 episodes).
Bring: Calibrated arm, the standard object set.
Slides: shared with confirmed guests ahead of the session.
The full zk0.bot Workshop Series:
Session 1 — Node Setup (Nov 30): https://luma.com/jcowyitm
Session 2 — Calibration Standard (Dec 7): https://luma.com/zunvvhi1
Session 3 — Data Sprint (Dec 14): https://luma.com/lyf25edz
Session 4 — Training & Rounds (Jan 4): https://luma.com/7ca8lhrn
Session 5 — Reading Results (Jan 18): https://luma.com/v5wpxrvw
Session 6 — The M1 Experiment (Feb 1): https://luma.com/myfflwx5
Hybrid: 16 seats at the East 6th studio (bring a laptop), or join by video. Instant confirmation. 3:00-5:00 PM CT.
Two stipend seats per session for operators who can't swing the $99 — pick the Stipend Seat ticket and tell us what you're building.
Next step: Commit an episode target toward the M1 design. Compute contributors: pair with a data partner for the Session 4 round.
Background: Data Beats Parameters — why zk0 trains a 450M robot model · Community: zk0 Discord · Code: github.com/ivelin/zk0
This session is recorded. Short clips may feature participants who opt in.
Advertisement
Event venue
Austin, United States
Tickets
usd 0
Discover more events by tag
Technology in AustinWorkshops in AustinRobotics in AustinArtificial-intelligence in AustinData-science in AustinMachine-learning in Austin