Post by Robotrove

𝗧𝗵𝗲 𝗿𝗼𝗯𝗼𝘁𝗶𝗰𝘀 𝗱𝗮𝘁𝗮 𝗯𝗼𝘁𝘁𝗹𝗲𝗻𝗲𝗰𝗸 𝗷𝘂𝘀𝘁 𝗴𝗼𝘁 𝘂𝗻𝗹𝗼𝗰𝗸𝗲𝗱. 14.6 years of internet video. 147 million action segments. Zero annotation. Rice University's @RobotPI Lab just dropped EgoInfinity — not another static dataset, but a data engine that turns arbitrary internet videos into executable robot training data. Here's what it does: → Takes any static-camera RGB video of human manipulation → Recovers metric 3D hand trajectories + 6-DoF object poses + contact states → Retargets the motion onto any robot morphology No mocap. No wearables. No manual labels. No human in the loop. The key insight: it's not a naive pipeline. Cross-module metric calibration and interaction-aware refinement fix the drift and contact inconsistency you'd get from just chaining vision models together. Already validated on 4 robots: Unitree G1, Robonaut2, Dual-Franka, XLeRobot — with real-robot demos on Franka FR3 and LEAP dexterous hand (grasping, cutting, wiping, pouring). The addressable corpus is bounded only by Action100M. The engine itself is corpus-agnostic. Scale is literally infinite. This is how you go from hundreds of lab demos to millions of real-world manipulation trajectories. Huge thanks to the team behind it Kejia Ren, @Yiting Chen, Howard Qian, Podshara Chanrungmaneekul, and Kaiyu Hang, with @Andrew Morgan from RAI Institute.  Project: https://lnkd.in/eDCtAtZU  Paper: https://lnkd.in/eRVG3xiU  Code: https://lnkd.in/eEwsatwx Follow the Robotrove page for more drops that push physical AI forward.

Video Content