Humanoid G1+Dex3 Bimanual Grasping Dataset | Envelop Oppose-Pinch Container | Dual-Hand Sync | RGB-D IMU Contact | Offline RL

The provider Humanoid G1+Dex3 Bimanual Grasping Dataset This is a training‐ready robotics data product. Expert‐level two‐hand manipulation trajectories are generated on our in‐house simulation platform built upon NVIDIA Isaac Sim and Isaac Lab. The hardware setup employs the Unitree G1 humanoid robot with Dex3 three‐fingered dexterous hands mounted on both wrists. It executes complete approach, grasp, transport and placement workflows via bimanual coordination strategies. Each data episode is dedicated to training one single bimanual action. This dataset is collected after the single‐hand dataset meets a high success‐rate gate; bimanual timing logic is built upon mature unimanual prior capabilities. Three bimanual grasp actions (props are not hard‐coded; users train over coordinated motion patterns): - Bimanual envelop grasp: Both hands approach laterally at approximately 45° and jointly wrap around the object rim. During placement, the held object is rotated back to its canonical pose. After stable settling, both hands open synchronously. Throwing and single‐hand release are prohibited. - Bimanual oppose pinch: Two hands squeeze opposing vertical faces of the object in a horizontal closing motion. For placement, lift the object above the bin, then open both hands simultaneously to drop the object inside. - Bimanual container manipulation: The primary grasp hand envelopes the container side‐wall to bear the main load; the assist hand rests lightly against the upper rim for stabil