UT Austin朱玉可：人形机器人的数据困局怎么破？答案藏在「数据海绵」里 ｜ICRA 2026

His work on models like SONIC and EgoScale demonstrates that with less than 1% real robot data, complex tasks can be achieved, leveraging vast human video data and synthetic data for training.

Key Points

• Yuke Zhu introduced a 'data pyramid' framework for humanoid robots at ICRA 2026.
• SONIC model uses human motion capture data for training, simplifying reinforcement learning.
• EgoScale requires less than 1% real robot data by leveraging first-person videos.
• NVIDIA's GR00T model enables end-to-end training for complex robotic tasks.
• Open-source initiatives are crucial for advancing humanoid robotics technology.