YOTO++: Learning Long-Horizon Closed-Loop Bimanual Manipulation from One-Shot Human Video Demonstrations.

Area of Science:

• Robotics
• Artificial Intelligence
• Computer Vision

Background:

• Bimanual robotic manipulation is complex due to dual-arm coordination challenges.
• High-dimensional action spaces hinder efficient skill learning.
• Current methods often require extensive data and programming.

Purpose of the Study:

• To present YOTO++ (You Only Teach Once), a unified one-shot learning framework for bimanual skills.
• To enable robots to learn complex manipulation from third-person human video demonstrations.
• To improve the generalization and adaptability of robotic systems in bimanual tasks.

Main Methods:

• Extracting 3D hand motions from binocular vision.
• Distilling motions into keyframe-based trajectories for dual-arm execution.
• Utilizing a demonstration proliferation strategy for synthetic data augmentation.
• Developing a customized bimanual diffusion policy.
• Implementing a visual alignment mechanism for closed-loop control.

Main Results:

• YOTO++ demonstrates strong generalization across diverse bimanual tasks (asynchronous, synchronous, contact-rich, non-prehensile).
• The framework shows effective learning of novel skills and adaptation to new objects.
• Successful cross-embodiment transfer to an unseen dual-arm robotic platform without retraining.
• Achieved high accuracy, robustness, and scalability in evaluations.

Conclusions:

• YOTO++ advances one-shot learning for complex bimanual robotic manipulation.
• The framework offers a practical solution for general-purpose bimanual systems.
• Enables efficient skill acquisition from limited demonstrations, reducing programming effort.