Robot Learns to Ride a Scooter with 52% Improvement through Machine Learning

Strategic Insights -

Researchers at National Taiwan Normal University have successfully taught a humanoid robot to ride a scooter using machine learning and AI technology. The team, led by Professor Jacky Baltes, designed two control methods to operate the robot and tested them at different speeds and rotation rates. The robot trained with machine learning showed an average performance improvement of about 52%, indicating a significant advancement in the robot's ability to balance and steer the scooter.

Key Takeaways:

  • The team created 3D digital models of a humanoid robot and a scooter, then used machine learning and AI technology to teach the robot to ride the vehicle.
  • Two-wheeled vehicles are harder to balance than four-wheeled ones when stationary or moving at low speeds, according to Professor Jacky Baltes.
  • The team designed two control methods to operate the robot and tested them at different speeds and rotation rates, including maintaining balance while moving, recovering quickly from disturbances, and navigating winding paths.
  • The robot trained with machine learning showed an average performance improvement of about 52%.
  • The team trained the robot using Nvidia Isaac Gym, a GPU-based simulation platform, which allowed for efficient motion training on a single GPU.
  • The team aims to enable the robot to demonstrate balance and steering control without modifying the specifications of the electric scooter.
  • The robot's scooter-riding capability has been verified in the laboratory, and the team plans to adapt the technology for testing in larger, real-world environments.
  • Teaching robots various skills is crucial for developing general-purpose robots, which can assist in disaster relief, senior care, and daily life.
  • The team's research can potentially lead to advancements in AI and machine learning for humanoid robots.

Statistics:

  • 52%: Average performance improvement of the robot trained with machine learning.
  • Nvidia Isaac Gym: A GPU-based simulation platform used to train the robot.
  • Thousands of CPU cores: The number of CPU cores required before the advent of GPU computing for tasks like motion training.
  • 1: The number of GPU required to perform tasks that once required thousands of CPU cores.

Sources:

  • ETtoday, the name of the publication that reported on the research as mentioned in the original text.