Scale AI labels fine-grained robot subtasks — one 195s task split into 11 segments
yacinelearning · x · 2026-11-17
In a 1.5-hour interview with Scale AI research engineer Jade Choghari ("FineART: Dense Subtask Make Better Robot Policies"), the core thesis is that better model performance often comes from better data — robotics included. Scale AI segments robot demonstrations into fine-grained subtasks with precise timestamps: e.g. a 194.8s "pack items into zip bag" task is split into 11 labeled subtasks grouped as open box/unzip bag, pack six items, and zip/close/return. Dense subtask labels are then used to train better robot policies.
Related event: Scale AI Boosts Robot Policies with Dense Subtask Annotation(2 posts)→
More from Embodied
- Open-Source reBot Arm B601-DM Robotic Arm on Sale for $1,090 — philfung · 2026-11-17
- NC State builds Howl, a Unitree G1-based humanoid that collects floor eggs in chicken farms — CyberRobooo · 2026-11-17
- UR5 arm threads cable through car door autonomously, end-to-end on Neuracore — stepjamUK · 2026-11-17
- JEPA4R: point-cloud JEPA world model transfers zero-shot to real robot arms — AIGeeksGroup · 2026-11-17
- Generalist AI's Andy Barry to Present Foundation Models in Robotics at MIT on Nov 19 — HanGuo97 · 2026-11-17
- NVIDIA open-sources USDCraft-10K: 10K articulated OpenUSD assets for robot training — chris_j_paxton · 2026-11-17