UIUC & CMU's LUCID: Teaching Robots Manipulation Skills from Human Videos
lukas_m_ziegler · x · 2026-08-09
UIUC and CMU introduced LUCID, a system that learns dexterous manipulation skills from unstructured human videos. It is embodiment-agnostic, meaning it can work across different robot bodies without retraining.
Notably, the researchers published the actual failure modes instead of hiding them:
- Stirring: Hand and pot occlude the spoon, causing tracking loss and revealing perception gaps.
- Wiping: The towel drops mid-motion and the hand presses into the table, making it unrecoverable.
- Binning: Grasp slips due to insufficient force control for objects with different properties.
- Cable routing: The dexterous hand gets stuck and fails to grasp the cable, showing some tasks remain highly difficult.
More from Embodied
- Leg-KILO: Open-Sourcing Robust SLAM for Dynamic Legged Robots — rsasaki0109 · 2026-08-09
- ω-0 Model Enables Humanoid Robots to Perform Household Tasks with 81.8% Success — CyberRobooo · 2026-08-09
- Waymo Reportedly Importing Chinese EVs by the Boatload Despite US Bans — alexmacgregor__ · 2026-08-09
- Stanford Team Unveils OpenDerm: An Open-Source Robot for At-Home Skin Cancer Tracking — StanfordAILab · 2026-08-09
- Debate on Robot AI Brains: Is VLA Dead or Converging with World Models? — Scobleizer · 2026-08-09
- Genspark Launches First Hardware Product SecondBrain Note: MagSafe Recorder with AI Summaries — dr_cintas · 2026-08-09