Innovative Handheld Interface Enhances Multi-Task Robot Training
A new handheld manipulation interface improves robot training efficiency.
Researchers have developed a battery-powered handheld Universal Manipulation Interface (UMI) designed to enhance the training of robot policies for multi-task pick and place operations. This new interface features soft, angled fingers that adapt to objects and provide visual feedback on contact forces, addressing limitations of existing UMIs.
The team collected a comprehensive dataset over 99.7 hours involving 53 objects, successfully training robot policies on 49 of them. Notably, the inclusion of a top camera improved success rates significantly, and mid-training adjustments led to enhanced performance on previously unseen objects.