Sign in Subscribe
Machine learning

Robots Master Jenga Whipping with AI-Powered Training

Strategic Insights

— 1 min read

Researchers at UC Berkeley have developed an AI-powered training method that enables robots to master complicated tasks like Jenga whipping with a 100% success rate. The system, called Human-in-the-Loop Sample Efficient Robotic Reinforcement Learning (HiL-SERL), teaches robots how to perform tasks by studying demonstrations and learning from both human feedback and its own real-world attempts. This training protocol allows robots to learn complex tasks such as assembling a computer motherboard, building a shelf, and more, within one to two hours.

Key Takeaways:

  • The new system, HiL-SERL, enables robots to master complicated tasks like Jenga whipping with a 100% success rate.
  • The robots are taught at an impressive speed, allowing them to learn within one to two hours how to perfectly assemble a computer motherboard, build a shelf, and more.
  • The system combines reinforcement learning with human intervention, allowing humans to correct the robot's course and incorporate those corrections into the robot's memory bank.
  • The researchers put the robotic system through a gauntlet of complicated tasks, including flipping an egg in a pan, passing an object from one arm to another, and assembling a motherboard, car dashboard, and timing belt, all with 100% accuracy.
  • The system was compared to a common "copy my behavior" method known as behavioral cloning, and the results showed that the new system made the robots faster and more accurate.
  • The researchers emphasized that the bar for robot competency is very high, and that made-to-order manufacturing processes could benefit from robots that can reliably and adaptably learn a range of tasks.

Statistics:

  • 100% success rate in mastering complicated tasks like Jenga whipping
  • 1-2 hours to learn complex tasks such as assembling a computer motherboard and building a shelf
  • 30% human intervention in the initial stages of training gradually decreasing to less attention
  • 100% accuracy in tasks such as flipping an egg in a pan, passing an object from one arm to another, and assembling a motherboard, car dashboard, and timing belt

Sources:

  • "Human-in-the-Loop Sample Efficient Robotic Reinforcement Learning" study appearing August 20 in the journal Science Robotics.
  • Sergey Levine's Robotic AI and Learning Lab at UC Berkeley.
  • Jianlan Luo, postdoctoral researcher at UC Berkeley.
  • Charles Xu and Jeffrey Wu, researchers at UC Berkeley.
  • Science Robotics journal.

Read more