The Importance of Neural Probabilistic Motor Primitives in AI Research
A recent article by Siqi Liu, Leonard Hasenclever, Steven Bohez, Guy Lever, Zhe Wang, and S. M. Ali Eslami, published in Science Robotics, discusses the significance of using human and animal motions to teach robots to complete complex tasks. Five years ago, the team demonstrated the potential of reinforcement learning (RL) by teaching a humanoid character to navigate obstacle courses. This exploration revealed the challenges of reusing previously learned behaviors and generating idiosyncratic movements.
To address these challenges, the team developed neural probabilistic motor primitives (NPMP) as a solution. NPMP involves guided learning with movement patterns derived from humans and animals. This approach was used to teach humanoid characters to carry objects and play football. The NPMP distills data into controllable motor primitives, enabling efficient exploration and the development of natural and practical movement skills.
In their latest work, the team applied NPMP to teach humanoid characters to play football competitively using multi-agent RL. This enabled the characters to exhibit agile locomotion, passing, and coordinated team play.
Additionally, NPMP was used to teach robotic control in the real world, such as dribbling a ball or carrying objects using vision. The NPMP approach also allowed for safe and efficient control of real robots, by learning skills and controllers in simulation that could be deployed on real humanoid and quadruped robots.
In summary, the NPMP model has proven to be an invaluable tool for AI research, enabling embodied agents to learn more quickly, naturalistic behaviors and safe, efficient, and stable behaviors suitable for real-world robotics.
For more information on this groundbreaking research, visitors can refer to the original article published in Science Robotics.