独家｜北大董豪：「仅停留在数据层面的Scaling Law，教不出通用机器人」

This approach aims to reduce data usage while increasing task success rates, crucial for the widespread deployment of general-purpose humanoid robots.

Key Points

• Current imitation and reinforcement learning methods have significant limitations in robot training.
• Dong's lab achieved fully autonomous laundry processes using a combination of imitation and reinforcement learning.
• The proposed two-dimensional Scaling Law integrates task quantity with data volume for better growth.
• Generative AI can create 50 realistic training samples from a single real-world trajectory.
• The focus on a unified growth curve aims to facilitate the commercialization of general-purpose robots.