Scale AI's FineART: 1,700-Hour Dual-Arm Robot Dataset Where Human Subtask Labels Enable Cross-Robot Transfer
yacinelearning · x · 2026-11-18
An in-depth interview with Jade Choghari of Scale AI covering the FineART dataset and FineART-VLA policy.
The problem: most robot datasets give one instruction per episode (e.g., "clean the table") for minutes of motion, forcing the model to figure out where each step ends — which is very hard.
FineART's approach: a 1,700-hour two-arm dataset where every single step is labeled in plain words by humans (534k labels, no VLM in the loop). These subtask labels do heavy lifting: without them, even a big vision-language backbone can't reliably tell the left bin from the right; with them, steps learned in words transfer to a completely different robot — completing a long breakfast task never practiced there and even reusing an "adjust" move only seen on the previous robot.
The chat also covers how Scale AI collects this data behind the scenes, defining subtasks, label review passes, why knowledge insulation protects the VLM, why the ABC baseline collapsed while FineART kept climbing, and how to mix sim + teleop + human video data.
More from Embodied
- VITURE launches $599 Phantom Beast XR glasses co-branded with Phantom Blade Zero — Scobleizer · 2026-11-18
- StarVLA open-sources a Lego-like codebase for VLA models, nearing 4,000 GitHub stars — wzenus · 2026-11-18
- Museo, a whisperable AI ring for phone-free walks, ships next month with open dev platform — Scobleizer · 2026-11-18
- Claude Fable 5.2 tops LIBERO-10 LLM robot control benchmark, lifting success to 54% — YuXiang_IRVL · 2026-11-18
- RL-trained 5-tonne skid-steer loader balances on two wheels after sim training — Scobleizer · 2026-11-18
- Valkryn raises $17M led by a16z to build a Battle Lab for robotic warfare — espricewright · 2026-11-18