Skip to content
humanoidrobots.training

UMI Gripper (Universal Manipulation Interface) · published weights

In-the-wild cup arrangement (rotate espresso cup, place on saucer)

Record

Framework
Diffusion Policy (UMI), fine-tuned CLIP ViT-L backbone · other
Policy license
not stated by the package · robot software repos: MIT
File
https://real.stanford.edu/umi/data/pretrained_models/cup_wild_vit_l_1img.ckpt
Note
Trained on handheld UMI data from 30 locations; deployed on UR5e + WSG-50 with a GoPro. Real-world checkpoint, about 3.05 GB. README warns it does not work well in direct sunlight.

Not in the tab yet

To replay this policy in the browser we need an ONNX export of the named checkpoint, the observation and action contract copied from the authors' own sim2sim or deploy script, and a MuJoCo model of the robot that the contract matches. Then it must walk 60 s in our runtime without a fall. Until then the source link is the way in.