Samsung Robotics Proposes ε4P: Route Imperfect Robot Data by Noise Level for Precision Manipulation
jiqizhixin · x · 2026-11-08
Samsung Robotics eXperience researchers introduce ε4P (Imperfection for Precision), a preprint method for high-precision manipulation showing that imperfect robot data still provides useful supervision if injected at the right noise level.
- Separate the mismatches: low-precision demos of the target task preserve useful context; precise demos from other tasks preserve useful action detail — conflating them hurts learning.
- Assign training regions: target-task low-precision data contributes at high noise, task-mismatched precise data at low noise, with boundaries set by a flow-time-conditioned source classifier and action-context reliance analysis. Policy architecture and flow-matching objective stay unchanged.
- Results: improved success on real-robot ATX connector insertion (figures truncated in source, roughly 4x).
The accompanying figure details both boundary mechanisms: task-matched/precision-mismatched via source-classifier AUC, precision-matched/task-mismatched via action-context reliance comparison.
More from Embodied
- PyTorch ATX and Austin Robotics host robot learning meetup alongside CoRL — kscottz · 2026-11-08
- Human2Any Transfers Human Video Skills to Robots with Zero Real-World Training Data — danfei_xu · 2026-11-08
- Amazon FAR's 'Pick Anything' grasps 66 object types with a single policy — JitendraMalikCV · 2026-11-08
- A group of humanoid robots walks through Figure AI HQ, Westworld vibes — CyberRobooo · 2026-11-08
- CoRL 2026 sells out: record 2,094 submissions, 687 accepted papers — Majumdar_Ani · 2026-11-08
- Clockwork uConsole kit boots on first assembly, but the screen stays dark — vincent_koc · 2026-11-08