Modelpublic

ASethi04/pi05-BimanualYAM-freshbase-umi100-coherence-balanced-ee20

sourceHugging Faceupdated 1mo agoView on Hugging Face

0likes12downloads

Model Card

pi05-BimanualYAM-freshbase-umi100-coherence-balanced-ee20

Fresh-base Pi0.5 trained for 12,000 optimizer steps on pure canonical UMI data with trajectory-coherence-balanced sampling and zero teleop frames.

—base: lerobot/pi05_base@b211f3d44c36b6acfcf7ae94a64e8e96f75a64ba
—UMI dataset: brandonyang/dual-lidar-umi-independent@a95b079b2b3dc73a912ecd12967f22f825d04fa8
—action: H24 EE20 current-relative SE(3), R6D rows, absolute future gripper
—task: pick up oranges and place them in the bowl
—rotation contraction: none
—state history: none
—variant: coherence

The bundled evaluation is a matched training-set observation replay, not a held-out or hardware-success claim. Hardware use requires the standard EE-to-IK safety path and operator supervision. Pin the immutable Hub revision rather than main.