Manipulation with Bimanual Dexterous Hands
• Completed
Manipulation Imitation Learning
This is an project about bimanual dexterous manipulation on Fourier GR1/2 humanoid robots.
Project Overview
- Enhanced the generalization of object positions in dual-arm dexterous humanoid robots Fourier gr1/2.
- Identified shortcut learning as a key generalization bottleneck for algorithms like ACT and Diffusion Policy. This was accomplished by applying visualization techniques such as Grad-CAM for attribution analysis and conducting comparative experiments across different observation and action spaces.
- Fundamentally improved training data quality and generalization potential by optimizing the data collection pipeline. A standardized data collection process was established, incorporating diverse action spaces, quantitative metrics, and realistic human-induced perturbations. By further decoupling hand and arm motions, this ultimately led to a significant improvement in visual understanding and positional generalization for manipulation tasks.