A quadruped solved a Rubik's cube with its feet
ETH Zurich and NVIDIA got a four-legged robot to pin and turn a giant Rubik's cube with the same two feet holding it up, three full autonomous solves on real hardware. Plus an open-source bimanual data rig and a gripper that feels force by looking at itself.
Shown this week
UMI-Live open-sources a two-handed data rig and 40 hours of data
Amish Sethi released UMI-Live, a fully open-source bimanual handheld rig that goes from pressing record to a trained robot policy. Each gripper carries its own camera, LiDAR, compute and battery, so there is nothing to calibrate in the scene, and finished episodes become LeRobot datasets uploaded to Hugging Face automatically.
You press one button and both hands start recording.
Two papers: humanoids manipulating with their chests, grippers feeling force by sight
Conor Mc Gartoll introduced Whole-Body Dexterity, which treats the entire humanoid body as one dexterous manipulator: bracing boxes against the chest, tucking objects under arms, rolling balls along forearms.
Can a robot see how hard it is squeezing without force sensors?
ETH Zurich quadruped solves a giant Rubik's cube using only its feet
Lukas Ziegler flagged new work from ETH Zurich and NVIDIA Robotics in which a four-legged robot autonomously solves a giant Rubik's cube with the same two feet it stands on, which the authors call pedipulation.
Pin and push at once, on legs that are also holding the robot up.