Frontier models as robot training teachers: zk0's federated learning for SmolVLA
DJiafei · x · 2026-11-21
The zk0 team argues frontier models are becoming great teachers for on-device robotics models: their GPT-6.2 Astra (Muse) manages the training pipeline — dataset cleanup, sim testing, and sim enrichment of novel edge-case tasks, unthinkable a year ago.
zk0 itself is an open-source federated learning platform:
- Trains the 450M-parameter vision-language-action model SmolVLA on SO-100 manipulation tasks via Flower, with privacy-preserving updates and no raw data sharing;
- Benchmarked six physics engines for SO-101 simulation and built a grasp-quality gate that flipped five of its own passing grades to FAIL;
- Backs a data-first bet ("data beats parameters"), publishes a phased roadmap, and recruits data and compute contributors.
More from AGI Musings
- Plinyus criticizes closed frontier models: collective knowledge hoarded behind closed doors — alexcovo_eth · 2026-11-21
- goodside jokes: don't be surprised the model built a Dyson sphere—it's a verifiable task — goodside · 2026-11-21
- Coding Value Trends to Zero, Systems Thinking to Infinity: AI Skills Debate — attentionmech · 2026-11-21
- AI Safety Is So New It Levels the Playing Field for Aspiring Researchers — eigenron · 2026-11-21
- Ex-OpenAI policy VP Miles Brundage jokes about AI containment via Pacific Rim — Miles_Brundage · 2026-11-21
- Tweet compares historical bubbles: 2025 is AI and housing — PAstynome · 2026-11-21