Exploring Neural Probabilistic Motor Primitives for Complex Tasks
Research
Teaching robots complex tasks like dribbling a ball or carrying boxes using human and animal motions.
Humanoid character learning to traverse an obstacle course through trial-and-error, demonstrating the emergence of locomotion behaviors.
Five years ago, a challenge was taken to teach humanoid characters complex tasks. This led to the development of neural probabilistic motor primitives (NPMP) to overcome challenges in embodied intelligence.
An agent learning to imitate a motion capture trajectory.
The NPMP model consists of an encoder and a low-level controller to distil and control motor intentions.
NPMP model distilling reference data into a low-level controller for new tasks.
Reusing the low-level controller enables efficient exploration and coordinated team play in humanoid football.
Learning football skills using the NPMP prior.
NPMP enables agile locomotion and teamwork in humanoid football.
Agents learning complex tasks like box carrying and ball dribbling.
NPMP approach solving tasks involving manipulation, perception, and memory.
Locomotion skills learned from biological motion aiding real-world robotic control.
Benefits of using NPMP for learning complex and naturalistic behaviors in robotic systems.